Add Batch 8f7bedcb-c756-40ab-aaa9-c1b52ccb082b data

.gitattributes

@@ -5130,3 +5130,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/Variational[[:space:]]Task[[:space:]]Vector[[:space:]]Composition/19fb88ee-2360-41e0-815e-64eafffa5c5e_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/Variational[[:space:]]Uncertainty[[:space:]]Decomposition[[:space:]]for[[:space:]]In-Context[[:space:]]Learning/046e33aa-5e08-446b-a7f3-9790f2549123_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/Vector[[:space:]]Database[[:space:]]Watermarking/82ad6074-a8de-4c07-9535-55b87ca87075_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vector[[:space:]]Quantization[[:space:]]in[[:space:]]the[[:space:]]Brain_[[:space:]]Grid-like[[:space:]]Codes[[:space:]]in[[:space:]]World[[:space:]]Models/d234e4d8-46ca-4953-9b53-72bd3966fe3a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Venus-MAXWELL_[[:space:]]Efficient[[:space:]]Learning[[:space:]]of[[:space:]]Protein-Mutation[[:space:]]Stability[[:space:]]Landscapes[[:space:]]using[[:space:]]Protein[[:space:]]Language[[:space:]]Models/d6e2fc33-adea-4d36-b95e-087103bd5d6d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VeriLoC_[[:space:]]Line-of-Code[[:space:]]Level[[:space:]]Prediction[[:space:]]of[[:space:]]Hardware[[:space:]]Design[[:space:]]Quality[[:space:]]from[[:space:]]Verilog[[:space:]]Code/f0ee19e6-974b-4839-8629-990369c86a68_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VeriThinker_[[:space:]]Learning[[:space:]]to[[:space:]]Verify[[:space:]]Makes[[:space:]]Reasoning[[:space:]]Model[[:space:]]Efficient/a0289d16-03a4-4105-9216-1417938420cf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Versatile[[:space:]]Transferable[[:space:]]Unlearnable[[:space:]]Example[[:space:]]Generator/0c230d75-f6ec-478b-87fa-3d67ddcc818b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vertical[[:space:]]Federated[[:space:]]Feature[[:space:]]Screening/6990da2f-c640-414b-8dc1-c12a32a82387_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vgent_[[:space:]]Graph-based[[:space:]]Retrieval-Reasoning-Augmented[[:space:]]Generation[[:space:]]For[[:space:]]Long[[:space:]]Video[[:space:]]Understanding/c5a4b368-9f46-42fb-9ad9-4fc0cdb91820_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViCrit_[[:space:]]A[[:space:]]Verifiable[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]Proxy[[:space:]]Task[[:space:]]for[[:space:]]Visual[[:space:]]Perception[[:space:]]in[[:space:]]VLMs/9f5de5f5-6130-4a1f-933d-16e900744798_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViDAR_[[:space:]]Video[[:space:]]Diffusion-Aware[[:space:]]4D[[:space:]]Reconstruction[[:space:]]From[[:space:]]Monocular[[:space:]]Inputs/20739a5b-dfd4-4e04-811c-f58f3ce782f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViSPLA_[[:space:]]Visual[[:space:]]Iterative[[:space:]]Self-Prompting[[:space:]]for[[:space:]]Language-Guided[[:space:]]3D[[:space:]]Affordance[[:space:]]Learning/e25a58db-a09d-4de2-9ae0-0e091f0d7926_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViSpec_[[:space:]]Accelerating[[:space:]]Vision-Language[[:space:]]Models[[:space:]]with[[:space:]]Vision-Aware[[:space:]]Speculative[[:space:]]Decoding/3e97780a-e086-44a6-a37c-203633bc0ec3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vicinal[[:space:]]Label[[:space:]]Supervision[[:space:]]for[[:space:]]Reliable[[:space:]]Aleatoric[[:space:]]and[[:space:]]Epistemic[[:space:]]Uncertainty[[:space:]]Estimation/35d43253-98be-4890-97d9-0f57d6e1c622_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vicinity-Guided[[:space:]]Discriminative[[:space:]]Latent[[:space:]]Diffusion[[:space:]]for[[:space:]]Privacy-Preserving[[:space:]]Domain[[:space:]]Adaptation/9351bd54-796d-48be-8005-832ff254c01f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vid-SME_[[:space:]]Membership[[:space:]]Inference[[:space:]]Attacks[[:space:]]against[[:space:]]Large[[:space:]]Video[[:space:]]Understanding[[:space:]]Models/254f2438-216a-470d-bcc4-33e679aa2903_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VidEmo_[[:space:]]Affective-Tree[[:space:]]Reasoning[[:space:]]for[[:space:]]Emotion-Centric[[:space:]]Video[[:space:]]Foundation[[:space:]]Models/64aff70f-f323-4799-a76d-4eba8d6bef19_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]Diffusion[[:space:]]Models[[:space:]]Excel[[:space:]]at[[:space:]]Tracking[[:space:]]Similar-Looking[[:space:]]Objects[[:space:]]Without[[:space:]]Supervision/fe623d1c-16dd-480c-8b28-8cc41b4b00e0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]Perception[[:space:]]Models[[:space:]]for[[:space:]]3D[[:space:]]Scene[[:space:]]Synthesis/02f59b70-9b2d-402a-91cf-c07464f5c34a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]World[[:space:]]Models[[:space:]]with[[:space:]]Long-term[[:space:]]Spatial[[:space:]]Memory/5ab95b39-aac9-4821-bd8e-a19342c636c0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video-R1_[[:space:]]Reinforcing[[:space:]]Video[[:space:]]Reasoning[[:space:]]in[[:space:]]MLLMs/62553eac-03c3-47e2-9335-9b237512f12e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video-RAG_[[:space:]]Visually-aligned[[:space:]]Retrieval-Augmented[[:space:]]Long[[:space:]]Video[[:space:]]Comprehension/a12dbbae-34bf-4f9d-8c9c-1113249082f3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoChat-R1.5_[[:space:]]Visual[[:space:]]Test-Time[[:space:]]Scaling[[:space:]]to[[:space:]]Reinforce[[:space:]]Multimodal[[:space:]]Reasoning[[:space:]]by[[:space:]]Iterative[[:space:]]Perception/57bd5930-864c-418d-b78e-a4c84cf6216d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoHallu_[[:space:]]Evaluating[[:space:]]and[[:space:]]Mitigating[[:space:]]Multi-modal[[:space:]]Hallucinations[[:space:]]on[[:space:]]Synthetic[[:space:]]Video[[:space:]]Understanding/60fdc8b5-cc06-4041-acd9-d25ae76f1285_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoLucy_[[:space:]]Deep[[:space:]]Memory[[:space:]]Backtracking[[:space:]]for[[:space:]]Long[[:space:]]Video[[:space:]]Understanding/9a81f3f4-d243-4259-8d0e-97b0450a6ae8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoMAR_[[:space:]]Autoregressive[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Continuous[[:space:]]Tokens/a83a5bf6-4e8f-421d-9d3a-abec78c8c2ac_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoREPA_[[:space:]]Learning[[:space:]]Physics[[:space:]]for[[:space:]]Video[[:space:]]Generation[[:space:]]through[[:space:]]Relational[[:space:]]Alignment[[:space:]]with[[:space:]]Foundation[[:space:]]Models/5f99cc7d-c897-42a9-b060-ba4577ff9b17_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoRFT_[[:space:]]Incentivizing[[:space:]]Video[[:space:]]Reasoning[[:space:]]Capability[[:space:]]in[[:space:]]MLLMs[[:space:]]via[[:space:]]Reinforced[[:space:]]Fine-Tuning/e9f04270-6fb7-48ce-b73b-24a91f8493f3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoTitans_[[:space:]]Scalable[[:space:]]Video[[:space:]]Prediction[[:space:]]with[[:space:]]Integrated[[:space:]]Short-[[:space:]]and[[:space:]]Long-term[[:space:]]Memory/bf02a1a7-8d3a-4faa-ab03-3af1bc1c2365_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoVLA_[[:space:]]Video[[:space:]]Generators[[:space:]]Can[[:space:]]Be[[:space:]]Generalizable[[:space:]]Robot[[:space:]]Manipulators/3bafe277-2467-44a9-bbe1-e57c23eb57ca_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Videos[[:space:]]are[[:space:]]Sample-Efficient[[:space:]]Supervisions_[[:space:]]Behavior[[:space:]]Cloning[[:space:]]from[[:space:]]Videos[[:space:]]via[[:space:]]Latent[[:space:]]Representations/83d2cbd2-c04e-4aaa-a6f3-ba5a06ae7420_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViewCraft3D_[[:space:]]High-fidelity[[:space:]]and[[:space:]]View-Consistent[[:space:]]3D[[:space:]]Vector[[:space:]]Graphics[[:space:]]Synthesis/c4057c81-cdfc-4b23-a9a7-41186435326c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViewPoint_[[:space:]]Panoramic[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Pretrained[[:space:]]Diffusion[[:space:]]Models/a893ce98-c8d1-42ad-8e94-5f5b3d822895_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VimoRAG_[[:space:]]Video-based[[:space:]]Retrieval-augmented[[:space:]]3D[[:space:]]Motion[[:space:]]Generation[[:space:]]for[[:space:]]Motion[[:space:]]Language[[:space:]]Models/cd8b5ea3-6dbb-46e9-9e5e-f09691051133_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vinci_[[:space:]]Deep[[:space:]]Thinking[[:space:]]in[[:space:]]Text-to-Image[[:space:]]Generation[[:space:]]using[[:space:]]Unified[[:space:]]Model[[:space:]]with[[:space:]]Reinforcement[[:space:]]Learning/382d1bb9-7b7e-4ff4-9015-79a6afee3e6e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Virtual[[:space:]]Fitting[[:space:]]Room_[[:space:]]Generating[[:space:]]Arbitrarily[[:space:]]Long[[:space:]]Videos[[:space:]]of[[:space:]]Virtual[[:space:]]Try-On[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image/97a90bc2-83bd-4004-b031-9e7e9663a9b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Virus[[:space:]]Infection[[:space:]]Attack[[:space:]]on[[:space:]]LLMs_[[:space:]]Your[[:space:]]Poisoning[[:space:]]Can[[:space:]]Spread[[:space:]]_VIA_[[:space:]]Synthetic[[:space:]]Data/1a34b33b-22f4-4c4b-881b-b62acc6db887_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VisDiff_[[:space:]]SDF-Guided[[:space:]]Polygon[[:space:]]Generation[[:space:]]for[[:space:]]Visibility[[:space:]]Reconstruction,[[:space:]]Characterization[[:space:]]and[[:space:]]Recognition/491d44f8-2e2a-4684-8978-71299ac454d0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision[[:space:]]Foundation[[:space:]]Models[[:space:]]as[[:space:]]Effective[[:space:]]Visual[[:space:]]Tokenizers[[:space:]]for[[:space:]]Autoregressive[[:space:]]Generation/486987a7-505b-46ac-be4d-b41c0a57a9e7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision[[:space:]]Function[[:space:]]Layer[[:space:]]in[[:space:]]Multimodal[[:space:]]LLMs/badc8cb1-7218-4829-b7a4-ebd4ab0542d8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision[[:space:]]Transformers[[:space:]]Don't[[:space:]]Need[[:space:]]Trained[[:space:]]Registers/3ca2e7cd-cb80-49a4-8de5-c510c567454c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision[[:space:]]Transformers[[:space:]]with[[:space:]]Self-Distilled[[:space:]]Registers/27e94445-2592-473c-91fd-4a2cfc43a773_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision[[:space:]]as[[:space:]]a[[:space:]]Dialect_[[:space:]]Unifying[[:space:]]Visual[[:space:]]Understanding[[:space:]]and[[:space:]]Generation[[:space:]]via[[:space:]]Text-Aligned[[:space:]]Representations/2132d4a3-1e58-462e-ab69-d5e39cec5b59_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision-and-Language[[:space:]]Training[[:space:]]Helps[[:space:]]Deploy[[:space:]]Taxonomic[[:space:]]Knowledge[[:space:]]but[[:space:]]Does[[:space:]]Not[[:space:]]Fundamentally[[:space:]]Alter[[:space:]]It/60862c61-ec14-478f-ab50-aae62dd534a7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision-centric[[:space:]]Token[[:space:]]Compression[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Model/6ffbd3f9-f576-4ade-8919-53059166ad81_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VisionThink_[[:space:]]Smart[[:space:]]and[[:space:]]Efficient[[:space:]]Vision[[:space:]]Language[[:space:]]Model[[:space:]]via[[:space:]]Reinforcement[[:space:]]Learning/19ce941d-c762-4ec8-9cc5-610cfd44e2cd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision‑Language‑Vision[[:space:]]Auto‑Encoder_[[:space:]]Scalable[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]from[[:space:]]Diffusion[[:space:]]Models/b9e9188c-8efe-45ed-a4ea-5ec8e3280f17_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Anagrams[[:space:]]Reveal[[:space:]]Hidden[[:space:]]Differences[[:space:]]in[[:space:]]Holistic[[:space:]]Shape[[:space:]]Processing[[:space:]]Across[[:space:]]Vision[[:space:]]Models/553fdfb5-108b-4bbf-8356-91ee359f4257_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Diversity[[:space:]]and[[:space:]]Region-aware[[:space:]]Prompt[[:space:]]Learning[[:space:]]for[[:space:]]Zero-shot[[:space:]]HOI[[:space:]]Detection/34f360a3-972d-486c-ae62-a34131fd6e5f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Instruction[[:space:]]Bottleneck[[:space:]]Tuning/ee2cba6a-b9f9-44e5-917b-8fc2cb0715b0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Jenga_[[:space:]]Discovering[[:space:]]Object[[:space:]]Dependencies[[:space:]]via[[:space:]]Counterfactual[[:space:]]Inpainting/c0b01e56-c3b1-4896-b1e8-751bceac0e77_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Structures[[:space:]]Help[[:space:]]Visual[[:space:]]Reasoning_[[:space:]]Addressing[[:space:]]the[[:space:]]Binding[[:space:]]Problem[[:space:]]in[[:space:]]LVLMs/5a960cb9-1923-4c32-b69a-5322d4450b9f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Sync_[[:space:]]Multi‑Camera[[:space:]]Synchronization[[:space:]]via[[:space:]]Cross‑View[[:space:]]Object[[:space:]]Motion/6f8aa5b9-3b00-48de-a5a9-919924beb945_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Thoughts_[[:space:]]A[[:space:]]Unified[[:space:]]Perspective[[:space:]]of[[:space:]]Understanding[[:space:]]Multimodal[[:space:]]Chain-of-Thought/b5ea7783-71de-4955-9a00-2375be0ae593_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VisualLens_[[:space:]]Personalization[[:space:]]through[[:space:]]Task-Agnostic[[:space:]]Visual[[:space:]]History/87b75197-e240-4066-bfd0-73c478c96857_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VisualQuality-R1_[[:space:]]Reasoning-Induced[[:space:]]Image[[:space:]]Quality[[:space:]]Assessment[[:space:]]via[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]to[[:space:]]Rank/a2c6589d-f15d-4021-bf07-9f342d6ad049_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VividFace_[[:space:]]A[[:space:]]Robost[[:space:]]and[[:space:]]High-Fidelity[[:space:]]Video[[:space:]]Face[[:space:]]Swapping[[:space:]]Framework/9c481239-7fb6-4799-81e8-4d3810240a1c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vocabulary[[:space:]]In-Context[[:space:]]Learning[[:space:]]in[[:space:]]Transformers_[[:space:]]Benefits[[:space:]]of[[:space:]]Positional[[:space:]]Encoding/dcfb3c9f-2b6a-4696-bfa6-e64253450a39_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vocabulary-Guided[[:space:]]Gait[[:space:]]Recognition/0b7db8cd-79e3-45a7-be87-ae67f3260dde_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Volume[[:space:]]Transmission[[:space:]]Implements[[:space:]]Context[[:space:]]Factorization[[:space:]]to[[:space:]]Target[[:space:]]Online[[:space:]]Credit[[:space:]]Assignment[[:space:]]and[[:space:]]Enable[[:space:]]Compositional[[:space:]]Generalization/3de3866c-af82-4e5d-9bd4-e356f541e719_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VoxDet_[[:space:]]Rethinking[[:space:]]3D[[:space:]]Semantic[[:space:]]Scene[[:space:]]Completion[[:space:]]as[[:space:]]Dense[[:space:]]Object[[:space:]]Detection/3c33fb97-32ef-4a31-b78a-0774535155c3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vulnerable[[:space:]]Data-Aware[[:space:]]Adversarial[[:space:]]Training/d2d62746-afae-4696-a00a-cf0f61d3cff6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/WALL-E_[[:space:]]World[[:space:]]Alignment[[:space:]]by[[:space:]]NeuroSymbolic[[:space:]]Learning[[:space:]]improves[[:space:]]World[[:space:]]Model-based[[:space:]]LLM[[:space:]]Agents/1aa331fb-c89e-4468-ae8b-f072dce6dcf7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/WHAT[[:space:]]MAKES[[:space:]]MATH[[:space:]]PROBLEMS[[:space:]]HARD[[:space:]]FOR[[:space:]]REINFORCEMENT[[:space:]]LEARNING_[[:space:]]A[[:space:]]CASE[[:space:]]STUDY/edfc9efa-573a-4a40-873f-70f74a93bce5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/WISA_[[:space:]]World[[:space:]]simulator[[:space:]]assistant[[:space:]]for[[:space:]]physics-aware[[:space:]]text-to-video[[:space:]]generation/a70a85ab-d10b-4cb1-bb9e-c69ff2a2069f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/WKV-sharing[[:space:]]embraced[[:space:]]random[[:space:]]shuffle[[:space:]]RWKV[[:space:]]high-order[[:space:]]modeling[[:space:]]for[[:space:]]pan-sharpening/7822acbd-5505-437b-91bb-46d8d695ed6a_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/Vector Quantization in the Brain_ Grid-like Codes in World Models/d234e4d8-46ca-4953-9b53-72bd3966fe3a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c60c3110136752f18ba28f3365943ba772b02092b6a2d54cff3b70bc48b98b9
+size 9796349

2025/Vector Quantization in the Brain_ Grid-like Codes in World Models/full.md

@@ -0,0 +1,734 @@
+# Vector Quantization in the Brain: Grid-like Codes in World Models
+
+Xiangyuan Peng*
+
+1900012762peng@pku.edu.cn
+
+Xingsi Dong*†
+
+dxs19980605@pku.edu.cn
+
+Si  $\mathbf{W}\mathbf{u}^{\dagger}$
+
+siwu@pku.edu.cn
+
+PKU-Tsinghua Center for Life Sciences, Academy for Advanced Interdisciplinary Studies.  
+IDG/McGovern Institute for Brain Research, Center of Quantitative Biology, Peking University.  
+School of Psychological and Cognitive Sciences,  
+Key Laboratory of Machine Perception (Ministry of Education).  
+*: Equal contribution.  
+†: Corresponding authors.
+
+# Abstract
+
+We propose Grid-like Code Quantization (GCQ), a brain-inspired method for compressing observation-action sequences into discrete representations using grid-like patterns in attractor dynamics. Unlike conventional vector quantization approaches that operate on static inputs, GCQ performs spatiotemporal compression through an action-conditioned codebook, where codewords are derived from continuous attractor neural networks and dynamically selected based on actions. This enables GCQ to jointly compress space and time, serving as a unified world model. The resulting representation supports long-horizon prediction, goal-directed planning, and inverse modeling. Experiments across diverse tasks demonstrate GCQ's effectiveness in compact encoding and downstream performance. Our work offers both a computational tool for efficient sequence modeling and a theoretical perspective on the formation of grid-like codes in neural systems.
+
+# 1 Introduction
+
+VQ-VAE [1] introduces discrete latent variables into the autoencoding framework through vector quantization (VQ) [2], allowing the model to compress high-dimensional continuous inputs into discrete, tokenized representations. This capability has led to its widespread application across various domains, including images [3, 4], video [5], speech [6], actions [7], and multimodal data [8], demonstrating its versatility in handling complex, diverse inputs. The success of VQ-VAE underscores the utility of compressing inputs into reusable codes as a general computational strategy for preprocessing and organizing data across a wide range of tasks.
+
+Biological systems face the similar challenge: how to process and represent high-dimensional, continuous inputs arising from multiple sensory and motor modalities. In parallel, the brain exhibits grid-like codes (GCs), which serve as general-purpose neural patterns for encoding information. GCs are extensively observed across various brain regions. Initially identified in the medial entorhinal cortex for spatial navigation [9], GCs have since been observed in the neocortex [10, 11, 12] and associated with representing abstract concepts beyond space, such as time and relational knowledge [10, 11, 13, 14]. This widespread neural activity is characterized by bump-like patterns, periodicity, and typically disentangled representations.
+
+Building on this insight, we propose a brain-inspired VQ method, Grid-like Code Quantization (GCQ), which uses the principles of GCs to structure the codebook. Specifically, we use continuous attractor neural networks (CANNs) [15, 16, 17] to generate grid-like activity patterns, where each stable
+
+state—bump—acts as a codeword. Due to the finite number of neurons, these bumps naturally form a discretized representation [18]. Unlike traditional VQ methods that use a static codebook, GCQ introduces an action-conditioned codebook: a dynamic set of codewords formed by CANN-generated bumps whose transitions are modulated by actions. This enables GCQ to perform quantization not on isolated observations, but on observation-action sequences, allowing the representations to capture temporal dependencies and behavioral context. Moreover, assigning distinct CANNs to different action types naturally yields disentangled representations, facilitating generalization and compositionality.
+
+The overall GCQ pipeline follows an encoder-quantizer-decoder architecture, adapted for action-conditioned sequence compression (Fig. 2). Specifically, the model processes an observation-action sequence, where the action sequence is used to construct an action-conditioned codebook, and the observation sequence is passed through the encoder to produce a corresponding latent sequence. This latent sequence is then quantized via template matching with the action-conditioned codebook. The matched codewords are passed to the decoder, which reconstructs the original observation sequence. Since the codebook is fixed, training requires only a commitment loss and a reconstruction loss. To enable gradient flow through the discrete quantization step, we use a straight-through estimator (STE).
+
+GCQ is a dynamic compression approach that operates on observation-action sequences, and therefore serves as a form of world model [19, 20]. Unlike prior world models that rely on a two-stage design to separately compress space and time—typically using models like VQ-VAE for static spatial observations and autoregressive models [21] for temporal dynamics—GCQ performs spatial and temporal compression jointly.
+
+In summary, our contributions are as follows:
+
+- To the best of our knowledge, GCQ is the first model to unify spatial and temporal compression through an action-conditioned quantization process. This enables direct compression of observation-action sequences, offering an integrated alternative to conventional two-stage world models. (Sec. 4)  
+- GCQ's spatiotemporal compression yields a cognitive map, which supports long-horizon prediction, goal-directed planning, and the derivation of an inverse model. In particular, goal-directed planning becomes computationally simple, as it reduces to finding a sequence of valid bump transitions on the map. (Sec. 5).  
+- GCQ offers insights into the formation of GCs in the brain, enhancing our understanding of neural representations (Sec. 6).
+
+# 2 Related Work
+
+VQ methods Vanilla VAEs [22] often suffer from posterior collapse in their latent spaces when compressing high-dimensional data [23], impairing downstream tasks. VQ-VAEs [1] address this by enforcing a structured latent space through discretization. Due to their superior compression efficiency and tokenization paradigm, VQ has become a standardized module in single-modal preprocessing pipelines in machine learning [3, 4]. In multimodal settings, these compressed tokens further act as a universal interface across modalities [5]. Meanwhile, numerous studies have proposed diverse codebook designs to enhance compression rates [24, 25, 26]. Unlike most learnable codebooks, FSQ uses a predefined codebook. Similarly, our GCQ utilizes a fixed codebook derived from continuous attractor dynamics. Critically, our method diverges from conventional VQ approaches by performing sequence-to-sequence template matching rather than single-frame matching.
+
+World models [19, 20] provide a framework for predicting future observations conditioned on actions. Most world models based on encoder-decoder architectures first compress observations using a VAE, and then model temporal dynamics in the latent space using temporal predictors such as RNNs [27, 28], Transformers [29], S4 models [30], or continuous Hopfield networks [31]. These approaches typically follow a two-stage design, with spatial and temporal compression handled separately. In contrast, GCQ is also an encoder-decoder world model, but it performs spatial and temporal compression jointly. There also exist decoder-only world models [32] that skip explicit compression and directly predict future observations. However, these models often struggle with planning due to the high computational cost of operating in the raw observation space. GCQ, by
+
+![](images/be081db885eca71eddc8db78173f27d54a24ceac393e60a2ddeca233ea295cba.jpg)
+
+![](images/aa439d90de83873d61f03972418303bc6cfb72e8dd8a986691d67d2a73be4610.jpg)
+
+![](images/035b4110edc65f22ebb35ccdc73f5dd95f0a3f43bba6b5035f1335e3be7f2d20.jpg)
+
+![](images/3a40dadc06885d4235c547b626ac215275cd0a962a3cc9c74a5b24451ff17eba.jpg)  
+Figure 1: (A) Schematic of a CANN: Each green dot represents a neuron uniformly distributed on a torus. The neurons receive external input  $I$ . (B) Energy landscape of CANN dynamics: Each local minimum in the energy landscape corresponds to an attractor state, which manifests as a 2D Gaussian bump on the torus. (C) Template matching via CANN dynamics: The CANN inherently performs template matching between the external input  $I$  and its attractor states. The input  $I$  is matched to the attractor that maximizes their inner product. (D) Attractor transition: Under four distinct actions, the attractor initially at position  $(0,0)$  stabilizes to four new attractor states. (E) Due to the periodic boundary conditions of the CANN, bump movements along the two axes naturally form grid-like patterns.
+
+![](images/01291a29c6360ee3bee8cc6c4678ff97e99e84e5a446714e265dbdffe18cb094.jpg)
+
+![](images/83eeba1c531df176bf816904de629b27cfbcbd345a8372da147db8cf3793f44c.jpg)
+
+![](images/d07a6b03df982b6b9fbb8887abe7d82eeecacb32dcf31d3d363a81f9b22dc3c4.jpg)
+
+![](images/2eb84c02aa17cb635b0b29c625596efd403d744d384f1088a672a658ad814618.jpg)
+
+![](images/41da4a1275c5fd577e871e0240442e47e5944a3388d52d8630053812fad3462d.jpg)
+
+![](images/d4d9471a80bca97a144b856e5083cb6fa541eea2a81bb4334e0699fa9149b243.jpg)
+
+compressing both space and time into a compact latent representation, enables more efficient planning and inference.
+
+Cognitive map with CANNs Unlike classical attractor networks [33]—which store discrete, unstructured patterns—CANNs encode structured patterns organized by metric relationships. This geometric regularity facilitates flexible state transitions through predefined operators [34], enabling operations like metric-based navigation and relational inference. Recent advances [35] have harnessed predefined CANNs as structured latent states for representation learning, empirically validating their ability to model neural population dynamics. Further work [36] proposes that structured latent spaces can map biologically to the entorhinal-hippocampal loop, a core circuit for spatial and episodic memory. However, existing implementations rely on biologically constrained online learning, which limits scalability. Our GCQ framework uses offline learning, enhancing parallelism and enabling application to large-scale datasets.
+
+# 3 CANNs and Template Matching
+
+In this section, we will briefly introduce CANNs, explain how they can form bumps as attractor states. In parallel, for VQ, the latent state obtained by the encoder must undergo template matching with codewords. We will demonstrate that CANNs inherently implement template matching between representations and bump states through their intrinsic dynamics. Finally, we will show how transitions between distinct attractor states can be mediated by actions.
+
+GCs can naturally be modeled by bumps in CANNs (Fig. 1E). The formation of CANNs does not require complex optimization but relies on translation-invariant connectivity and periodic boundary conditions. CANNs have been widely used as canonical models to elucidate the encoding of features in neural systems, including, for example, the encoding of orientation [37], head direction [38] and spatial location [17, 39]. CANNs can be expressed through various mathematical formulations. Here, we adopt a relatively concise form [16] to demonstrate their principles. We consider  $N^2$  neurons distributed on a toroidal  $(S^1 \times S^1)$  surface. These neurons are indexed by their positions on the torus  $\theta \in \{\theta_i\}_{i=1}^N$  and  $\varphi \in \{\varphi_j\}_{j=1}^N$ , where  $\theta_i$  and  $\varphi_j$  are uniformly distributed over  $(-\pi, \pi]$  (Fig. 1A). Let  $U_{\theta,\varphi}(t)$  and  $r_{\theta,\varphi}(t)$  denote the synaptic input and firing rate, respectively, of the neuron located
+
+at  $(\theta, \varphi)$  at time  $t$ . The dynamics of the CANN are governed by:
+
+$$
+\tau \frac {\partial U _ {\theta , \varphi} (t)}{\partial t} = - U _ {\theta , \varphi} (t) + \rho \sum_ {\theta^ {\prime}, \varphi^ {\prime}} W _ {\theta , \varphi} \left(\theta^ {\prime}, \varphi^ {\prime}\right) r _ {\theta^ {\prime}, \varphi^ {\prime}} (t) + I _ {\theta , \varphi} (t), \tag {1}
+$$
+
+where  $\tau$  is the synaptic time constant and  $\rho$  is the neuronal density.  $W_{\theta, \varphi}(\theta', \varphi')$  is the recurrent neuronal connections weights between neuron  $(\theta, \varphi)$  and neuron  $(\theta', \varphi')$ ,
+
+$$
+W _ {\theta , \varphi} (\theta^ {\prime}, \varphi^ {\prime}) = \frac {J}{2 \pi a ^ {2}} \exp \left[ - \frac {| | \theta - \theta^ {\prime} | | _ {S} ^ {2} + | | \varphi - \varphi^ {\prime} | | _ {S} ^ {2}}{2 a ^ {2}} \right]. \tag {2}
+$$
+
+The norm  $\| \cdot \| _S$  denotes the shortest path between two points on the circle, ensuring periodic boundary and translation-invariant conditions. The parameters  $J$  and  $a$  control the strength and width of the Gaussian connectivity, respectively. The nonlinear relationship between the firing rate  $r_{\theta ,\varphi}(t)$  and the synaptic input  $U_{\theta ,\varphi}(t)$  is implemented by divisive normalization, which is written as,
+
+$$
+r _ {\theta , \varphi} (t) = \frac {U _ {\theta , \varphi} ^ {2} (t)}{1 + k \rho \sum_ {\theta^ {\prime} , \varphi^ {\prime}} W _ {\theta , \varphi} \left(\theta^ {\prime} , \varphi^ {\prime}\right) U _ {\theta^ {\prime} , \varphi^ {\prime}} ^ {2} (t)}, \tag {3}
+$$
+
+where  $k$  controls the normalization strength. In reality, divisive normalization could be implemented by shunting inhibition [40].
+
+Previous studies [16, 41] have established that the CANN dynamics governed by Eq. (1) possess  $N^2$  stationary states (attractors) when the external input  $I_{\theta, \varphi}(t) = 0$  (Fig. 1B). Each state corresponds to a 2D Gaussian bump on the torus, centered at coordinates  $(\theta, \varphi)$ , with the firing rate of the neuron at position  $(\theta', \varphi')$  given by:
+
+$$
+e _ {\theta , \varphi} \left(\theta^ {\prime}, \varphi^ {\prime}\right) = A \exp \left[ - \frac {\| \theta - \theta^ {\prime} \| _ {S} ^ {2} + \| \varphi - \varphi^ {\prime} \| _ {S} ^ {2}}{2 a ^ {2}} \right], \tag {4}
+$$
+
+where  $A = \left[1 + (1 - 32\pi a^2 k / J^2\rho)^{1 / 2}\right] / (4\pi a^2 k\rho)$  is the amplitude. When  $I_{\theta, \varphi}(t)$  is a constant input, prior work [42] demonstrated that after its removal, the network converges to an attractor determined by:
+
+$$
+\theta^ {*}, \varphi^ {*} = \max  _ {\theta , \varphi} \sum_ {\theta^ {\prime}, \varphi^ {\prime}} e _ {\theta , \varphi} \left(\theta^ {\prime}, \varphi^ {\prime}\right) I _ {\theta^ {\prime}, \varphi^ {\prime}}, \tag {5}
+$$
+
+which demonstrates that CANN dynamics effectively perform template matching between the input  $I$  and the  $N^2$  attractors according to their inner product. (Fig. 1C).
+
+The bump in CANNs exhibit high mobility, enabling controlled movement through mechanisms such as: anti-symmetric connections [43], negative feedback [44, 43], velocity neurons [45]. Such bump displacements correspond to transitions between attractor states. For the toroidal CANN described above, each attractor can undergo local two-dimensional displacements in the  $\theta$ ,  $\varphi$  plane. We define two orthogonal action bases aligned with the  $\theta$  and  $\varphi$  axes (Fig. 1D),
+
+$$
+a _ {\theta} ^ {\pm} = e _ {\theta \pm \Delta \theta , \varphi} - e _ {\theta , \varphi}, \quad a _ {\varphi} ^ {\pm} = e _ {\theta , \varphi \pm \Delta \varphi} - e _ {\theta , \varphi}. \tag {6}
+$$
+
+where  $\Delta \varphi$  and  $\Delta \theta$  denote a small displacement step.
+
+# 4 Grid-like Code Quantization
+
+In this section, we first introduce the action-conditioned codebook in GCQ and the template matching process for sequences. We then describe how GCQ enables bidirectional mapping between real-world actions and latent transitions, and propose a greedy operator for measuring distances on the cognitive map to support inverse modeling and planning.
+
+# 4.1 Action-conditioned codebook and sequence matching
+
+We first introduce the key difference between GCQ and VQ from a high-level perspective. In the VQ method, the encoder first compresses the observation  $o$  into  $s$ , which is then matched to the closest codes in the codebook through template matching, producing  $\hat{s}$ . The decoder then reconstructs  $\hat{o}$  from  $\hat{s}$ . In GCQ, the input consists of an action-observation sequence  $\{o_1, o_1, o_2, o_2, \dots, o_n\}$ . The encoder
+
+![](images/fe27c638681a42a2e65bea1de2a1eb77c2bc74fd4fdfaed1f5cfd4e37f8874e6.jpg)  
+Figure 2: Schematic of GCQ: The action-observation sequence is encoded by the encoder into a latent state composed of  $m = 3$  codes. Through sequence template matching with the action-conditioned codebook, the decoder reconstructs the predicted observations. The gray arrows indicate the template matching process, and the gray dashed boxes represent the matching targets.
+
+compresses the observation sequence  $o_{1:n} = \{o_1, o_2, \dots, o_n\}$  into  $s_{1:n}$ , which is then matched to the closest codes in the action-conditioned codebook via template matching, yielding  $\hat{s}_{1:n}$ . The decoder then reconstructs  $\hat{o}_{1:n}$  from  $\hat{s}_{1:n}$ .
+
+In GCQ, each code corresponds to an attractor in the CANN (Fig. 1B). In the previous section, we used  $\theta$  and  $\varphi$  to index different attractors; for simplicity, we will now use natural numbers as attractor indices. Each code consists of  $d$  neurons, and the codebook contains  $K$  attractors. A simple implementation sets  $K = d$ , where each attractor's center coincides with a single neuron. Alternatively, we can set  $K > d$ , causing some attractor centers to fall between two neurons. In practice, different combinations of  $K$  and  $d$  can be selected. The state representation  $s_i \in \mathbb{R}^{m \times d}$ , meaning that  $s_i$  is composed of  $m$  codes.
+
+Additionally, we manually define a mapping between the action sequence  $a_{i} \in \mathcal{A}$  from the dataset and the action combinations applied to the CANNs. For notational simplicity, we hereafter use  $a_{i}$  to refer to an action in either the original space or the CANN space. In the latter context,  $a_{i} \in \mathbb{R}^{m \times d}$  represents the composite action over  $m$  bumps, with its component  $a_{i}^{j} \in \mathbb{R}^{d}$  denoting the action applied to the  $j$ -th bump in Eq.(6). Each CANN supports five distinct actions, resulting in up to  $5^{m}$  possible action combinations across  $m$  CANNs. Since this mapping is injective, the discrete action space must satisfy  $|\mathcal{A}| \leq 5^{m}$ . For continuous actions, a CANN can define transitions in two directions, imposing the constraint  $\dim(\mathcal{A}) \leq 2m$ .
+
+After establishing the mapping, we quantize the latent representation  $s_{1:n} = \{s_{1:n}^j\}_{j=1}^m$ . This representation consists of a set of  $m$  parallel sequences, where each  $s_{1:n}^j$  corresponds to a sequence from one of the  $m$  CANNs (as depicted by the dashed lines in Fig. 2). The quantization process is performed independently for each of these  $m$  sequences. For each latent sequence  $s_{1:n}^j$ , we perform a template matching procedure. This involves comparing  $s_{1:n}^j$  against a set of  $K$  candidate trajectories. Each candidate trajectory is generated by applying the known action sequence  $a_{1:n-1}^j$  to a base bump state  $e_i$ . We denote this operation as:
+
+$$
+e _ {i} \oplus a _ {1: n - 1} ^ {j} = \left\{e _ {i}, e _ {i} + a _ {1} ^ {j}, \dots , e _ {i} + a _ {1: n - 1} ^ {j} \right\}, \tag {7}
+$$
+
+where  $e_i + a_{1:n-1}^j = e_i + \sum_{t=1}^{n-1} a_t^j$ . The index  $k$  of the best-matching codeword for the  $j$ -th latent sequence is found by minimizing a distance metric (e.g., the L2 norm) between the latent sequence and each of the  $K$  candidate trajectories:
+
+$$
+k _ {j} = \arg \min  _ {i \in \{1,.., K \}} | | s _ {1: n} ^ {j} - \left(e _ {i} \oplus a _ {1: n - 1} ^ {j}\right) | | \tag {8}
+$$
+
+![](images/43129556252e69c743a8a5e42f8b267ff0b985e972c9e550ae7dbdeac42cebfc.jpg)  
+(A) Long-horizon prediction
+
+![](images/26801bc85de1a23be5ca855b333bab277477ee7178220259da056db27398055f.jpg)  
+(B) Goal-directed planning  
+Figure 3: (A) Schematic of long-horizon prediction. The figure illustrates the process of initializing with a sequence of length 3 and predicting two future observations. (B) Schematic of goal-directed planning. Some gray arrows are omitted for clarity. The blue arrows represent the mapping from actions in the latent space to real agent actions. Through iterative action generation, environment interaction, and observation, the agent continues until it outputs a no-op action  $a_{\emptyset}$ , indicating that the goal has been reached.
+
+Finally, the quantized sequence  $\hat{s}_{1:n}^{j}$  is constructed using this optimal codeword  $e_{k_j}$ . The complete quantized representation  $\hat{s}_{1:n}$  is the collection of these individually quantized sequences:
+
+$$
+\hat {s} _ {1: n} ^ {j} = e _ {k _ {j}} \oplus a _ {1: n - 1} ^ {j}, \quad \text {a n d} \quad \hat {s} _ {1: n} = \left\{\hat {s} _ {1: n} ^ {j} \right\} _ {j = 1} ^ {m} \tag {9}
+$$
+
+When computing the loss in GCQ using backpropagation (BP), we adopt the same straight-through estimator (STE) as in the VQ method, copying gradients from the decoder input to the encoder output to enable gradient flow to the encoder. GCQ uses two loss terms: a reconstruction loss and a commitment loss:
+
+$$
+L = \left\| o _ {1: n} - \hat {o} _ {1: n} \right\| ^ {2} + \beta \left\| s _ {1: n} - \operatorname {s g} \left[ \hat {s} _ {1: n} \right] \right\| ^ {2} \tag {10}
+$$
+
+where  $\mathrm{sg}[\cdot ]$  denotes the stop-gradient operation and  $\beta$  adjusts the strength of the commitment loss.
+
+In GCQ, the encoder and decoder are not designed in the same way as in conventional VQ models. Traditional VQ architectures often use ResNet-based building blocks, which provide each code with only a limited receptive field. As a result, modifying a single code typically leads to only local changes in the reconstructed observation. In contrast, GCQ assigns each code to an action, and altering the action can result in global changes to the observation. This necessitates that each code has access to global information during encoding and decoding. To address this, we explore three architectural variants for the encoder and decoder: (1) ResNet followed by a fully connected layer, (2) ViT [46], and (3) a hybrid of ResNet and ViT. Among these, ViT achieves the best trade-off in terms of parameter efficiency, training stability, and overall performance (Table.1).
+
+# 4.2 Operations on cognitive map
+
+GCQ uses a structured latent space, allowing an agent's actions in the real environment to correspond to simple movements of bumps within the latent space. In effect, GCQ constructs a space defined by bump dynamics, which can be interpreted as a cognitive map. By establishing a mapping between observations and this map, actions in the real space can be projected onto the map to determine position changes, and conversely, movements within the map can be mapped back to real-space actions. This bidirectional mapping enables GCQ to support both inverse modeling and goal-directed planning. Specifically, to compute the distance between two states  $s_i$  and  $s_j$ , we define an operation on the cognitive map. Since bump movements are action-driven and only valid actions produce feasible transitions, we introduce the following operation:
+
+$$
+s _ {i} \ominus s _ {j} = \underset {a \in \mathcal {A}} {\arg \min } \left| s _ {j} + a - s _ {i} \right|. \tag {11}
+$$
+
+This operation represents a greedy step: it selects the best valid action  $a$  that moves  $s_j$  one step closer to  $s_i$ .
+
+![](images/7769bd96ea8322d2959665a692ac03e988e602d19d274f9f1be2538a6b9999ce.jpg)  
+(A)
+
+![](images/699d7b1eceec452c017a6b376904bc84bb456fe73ae5e0291589410727eec460.jpg)  
+(B)
+
+![](images/803a4d0c202a3500634ebe62833acb9421ee9bb31019f71370ac1559087ae201.jpg)  
+(C)
+
+![](images/f5db7b32ed77172d72d4c1c14d8d331a9874099c50aeb0375448cd37c3dd736e.jpg)  
+(D)
+
+![](images/ed03e3805c960beeb20eeb1e65d2406f0d6a03b615a9b7bca7b6e2d47a0f1242.jpg)  
+(E)  
+Figure 4: (A)(B) Reconstruction FID and prediction FID for GCQ and VQ+UNet across different model sizes. (C) Prediction FID of GCQ varies with changes in the initialization length. (D) Prediction FID of GCQ (#Para:112M), VQ+UNet (96M) and VQ+Transformer (121M) changes as the prediction length increases. (E) Predictions on GSV dataset. The first patch in each row represents the trajectory drawn by the action. The first three rows correspond to movement actions, while the last two rows correspond to rotation actions. In practice, different actions are encoded within a single GCQ, enabling their use. For visualization convenience, they are plotted separately here. Rows 1, 2, and 4 show GCQ predictions with different initialization lengths (orange frames), predicting subsequent observations (blue frames). Rows 3 and 5 show VQ+UNet predictions under the same conditions. As the prediction length increases, the images become blurry.
+
+# 5 Experiment
+
+As a spatiotemporal compression model, GCQ is first evaluated in ablation studies to demonstrate its ability to compress and reconstruct observations. We then show that GCQ, when used as a world model, supports long-horizon prediction, goal-directed planning, and inverse modeling. Compared to traditional two-stage models, GCQ exhibits superior performance in long-range prediction tasks.
+
+Datasets. We evaluate GCQ on four datasets, all of which contain image-based observations. The 2DMaze [47] dataset is a virtual environment where actions correspond to the agent's movements. Each observation contains a full view of the maze, providing complete information. The Google Street View (GSV) dataset represents real-world environments with partial observations; the actions include both translational movements and rotational head turns in two directions. In the MPI3D [48] and 3DShapes [49] datasets, actions are defined as abstract feature-level changes.
+
+Baselines. We compare GCQ with traditional two-stage world models. VQ-VAE is used in the first stage for spatial compression. The codebook size in GCQ and VQ-VAE is kept the same for a fair comparison. For modeling temporal relationships, we use a UNet that predicts the next latent state  $s_{t+1}$  based on the current latent state  $s_t$  and action  $a_t$ . We refer to this baseline as 'VQ+UNet.' For action embedding in the UNet, we follow the approach from LAPO [7]. To further model temporal dependencies, we also adopt a Transformer-based architecture following TransDreamer [29]. We refer to this baseline as 'VQ+Transformer.'
+
+Evaluation Metrics. To evaluate the quality of the model-generated observations, we report peak signal-to-noise ratio (PSNR) for pixel-level reconstruction fidelity, and use the Fréchet Inception Distance (FID) [50] to assess the quality of generated images.
+
+Ablations We first conducted ablation experiments on the GSV dataset. Table 1 presents the performance of three different encoder-decoder network building blocks. It can be observed that the ViT and Hybrid models achieve better performance with fewer parameters. However, during the
+
+experiments, we found that the Hybrid model was less stable in training and converged more slowly than ViT. Therefore, unless otherwise specified, all subsequent experiments utilized the ViT-structured network. The GCQ exhibits scalability with model size similar to VQ+UNet, both in reconstruction and prediction. (Fig. 4A,B).
+
+<table><tr><td>Model type</td><td>Image size</td><td>#Para.</td><td>FIDr↓</td><td>FIDp↓</td><td>PSNRt↑</td><td>PSNRp↑</td></tr><tr><td>Resnet</td><td>3 × 80 × 40</td><td>330M</td><td>48.05</td><td>48.57</td><td>25.70</td><td>25.59</td></tr><tr><td rowspan="2">Hybrid</td><td>3 × 80 × 40</td><td>64M</td><td>21.29</td><td>22.31</td><td>29.07</td><td>28.54</td></tr><tr><td>3 × 128 × 128</td><td>140M</td><td>41.55</td><td>41.91</td><td>26.31</td><td>25.59</td></tr><tr><td rowspan="2">ViT</td><td>3 × 80 × 40</td><td>90M</td><td>13.27</td><td>13.92</td><td>31.32</td><td>31.34</td></tr><tr><td>3 × 128 × 128</td><td>112M</td><td>42.56</td><td>43.41</td><td>27.82</td><td>27.77</td></tr></table>
+
+We also make the bump-like codes in the codebook learnable by using the following loss function:
+
+$$
+L = \left\| o _ {1: n} - \hat {o} _ {1: n} \right\| ^ {2} + \beta \left\| s _ {1: n} - \operatorname {s g} \left[ \hat {s} _ {1: n} \right] \right\| ^ {2} + \gamma \left\| \operatorname {s g} \left[ s _ {1: n} \right] - \hat {s} _ {1: n} \right\| ^ {2} \tag {12}
+$$
+
+However, our experiments show that making the codes learnable actually degrades performance. We attribute this to the fact that, unlike the relatively simple codes in VQ, our codes exhibit more complex dynamic relationships. Allowing the codes themselves to be trained may therefore reduce training stability.
+
+Table 1: Model comparisons on the Street View dataset. FIDr, FIDp, PSNR, and PSNRp represent the FID and PSNR scores for reconstruction and prediction, respectively. ↓ and ↑ indicate that lower or higher values are better. The ResNet model was not trained on higher resolution images because its architecture includes a fully connected layer after the convolutional backbone, resulting in an excessively large number of parameters.  
+
+<table><tr><td>Model</td><td>FIDp↓</td><td>PSNRp↑</td></tr><tr><td>GCQ (fixed)</td><td>43.41</td><td>27.77</td></tr><tr><td>GCQ (learnable)</td><td>47.76</td><td>24.48</td></tr></table>
+
+Table 2: Effect of learnable vs. fixed codes in GCQ.
+
+Long-horizon prediction. After being initialized with an observation-action sequence, GCQ can perform actions directly in the latent space to predict future observations (Fig. 3A). Notably, its prediction performance remains stable regardless of the length of the initialization sequence (Fig. 4C; Fig. 4E, rows 1-2). As shown in Fig. 4D, the performance of the VQ method degrades as the prediction horizon increases, whereas GCQ maintains robust predictive quality due to its stable latent structure (Fig. 4E, rows 2-3). This is a key advantage of GCQ: by constructing a consistent cognitive map, it effectively addresses the instability issues commonly seen in current world models [51]—such as inaccurate predictions after completing a full rotation in the environment (Fig. 4E, rows 4-5). GCQ also demonstrates strong zero-shot prediction capabilities on relatively simple datasets. As shown in Fig. 5, rows 1-2, the model produces reasonable predictions in environments it has never encountered during training. Furthermore, by treating abstract feature transitions as a form of action, GCQ can also be used to predict observation changes driven by abstract-level variations. These predictions likewise exhibit long-range stability (Fig. 5, rows 3-4).
+
+Goal-directed planning. Given a goal and an initial position, the GCQ can utilize the distance in the cognitive map to generate the most desirable action for the current step. After executing the action, a new observation is obtained, and this process is iterated, continuously reducing the distance to the goal in the cognitive map until the goal is reached (Fig. 3B, Fig. 6 rows, 1-2). The computation of the action at each step is of constant complexity.
+
+Inverse model. Given a sequence of observations, the GCQ can first map them onto the cognitive map and then use goal-directed planning to determine the action or sequence of actions between adjacent observations, thus implementing the inverse model. The corresponding action sequence can be applied to the latent representation of another observation, and using the prediction capability, the generated sequence under this set of actions can be obtained (Fig. 6 rows, 3-4).
+
+![](images/47082170b70593401ae8857afa9d5ebb32ffbd79bb3cfb1340f3a5af3c476af2.jpg)  
+Figure 5: With the same setup as Fig. 4E, Rows 1-2: Prediction on the 2DMaze dataset. Rows 3-4: Prediction on the 3DShapes dataset.
+
+# Goal-directed planning
+
+![](images/c4e0bdc04019c6d1273c3de2b9fbd0bd71c73e984c52ee52b1ceb63453c21a39.jpg)  
+Figure 6: Rows 1-2: Goal-directed planning. The first patch shows the trajectory after planning, with orange indicating the starting point, red indicating the endpoint, and blue representing the planned route. The subsequent red-framed patches represent the endpoint, orange-framed patches represent the starting point, and blue-framed patches represent intermediate observations encountered during the process. Rows 3-4: Inverse modeling. The top row displays the given observation sequence. The bottom row starts with the first patch showing the action trajectory inferred from the observation sequence, with orange indicating the given initial observation, followed by the sequence generated based on the action trajectory.
+
+# 6 Discussion
+
+In this work, we introduced GCQ, a brain-inspired framework for compressing observation-action sequences into discrete, structured representations. GCQ uses continuous attractor dynamics to generate grid-like codewords, and selects them in an action-conditioned manner to capture both spatial and temporal dependencies. This spatiotemporal quantization process produces compact latent representations that serve as cognitive maps, enabling long-horizon prediction, goal-directed planning, and inverse modeling. Our experiments demonstrate that GCQ supports generalization across tasks while offering interpretability through its structured latent space.
+
+Insights for Neuroscience. Beyond its practical performance, GCQ also offers a new computational hypothesis for the emergence of GCs in the brain. Traditionally, the formation of neurons with structured tuning properties was approached through handcrafted models [52], which provided only limited explanatory power. In contrast, the machine learning paradigm offers a data-driven framework: artificial neural networks are optimized to perform cognitive tasks, and their internal representations are analyzed to reveal emergent coding principles. Following this approach, prior studies have shown that GCs can arise when networks are trained to perform path integration under biologically constraints [53, 54, 55]. Subsequent work has emphasized the importance of predictive rather than reconstructive objectives [56] and extended the analysis to more general frameworks such as world models and predictive learning [57]. Efforts to induce disentangled representations through architectural or loss function constraints have further refined these insights [58, 59]. However, the robustness of grid-like pattern emergence remains debated [60], with some studies [61] suggesting that specific architectural features (e.g., one-hot inputs) are necessary.
+
+Recent developmental findings add a new dimension to this discussion. Experiments show that toroidal activity patterns emerge in the medial entorhinal cortex even before sensory experience [62]. Intriguingly, such toroidal structures can be naturally modeled by bump attractors in CANNs. This suggests that the brain may possess preconfigured low-dimensional structures capable of bump activity, even prior to learning. Consistent with this, recent work has argued that GCs likely emerge from internal CANN mechanisms rather than from purely feedforward architectures [63].
+
+This leads us to a novel hypothesis inspired by GCQ: GCs may arise not from optimizing networks, but from learning to map sensory experience onto a set of preexisting bump-based activity patterns. In GCQ, the codebook is defined by CANN-generated bumps before learning begins. The learning process then consists of associating observation-action sequences with combinations of these fixed codewords. Similarly, we speculate that the brain may use a fixed set of toroidal patterns—produced by CANNs—as a biological codebook. Through experience, the brain learns to map external sensory inputs onto these internal structures, endowing them with meaning and interpretability, allowing for the decoding of grid-like patterns. This perspective suggests a unified model of how the brain may simultaneously achieve compression and semantic organization of sensory information.
+
+Static Setting. We also evaluated GCQ in a static setting, where the sequence length is 1. In this case, we compared GCQ with VQ-VAE on ImageNet [64], finding that GCQ suffered minimal performance degradation. This suggests that the use of a fixed codebook does not significantly harm performance on static tasks. For further details, refer to Appendix A.
+
+Scalability of Action. Our current work utilizes 2D attractors, where each has 5 potential transitions (four shifts and one stationary). With such  $m$  CANNs, the model can represent  $5^m$  distinct actions. If we use  $P$ -dimensional attractors, the number of states per CANN will become  $2P + 1$ , yielding a total action space of  $(2P + 1)^5$ . Therefore, GCQ can be scaled to higher-dimensional action spaces by adjusting both  $m$  and  $P$ . For the experiments in this paper, which involve relatively low-dimensional actions, 2D attractors are sufficient.
+
+Future Work. A promising direction for advancing GCQ lies in enabling the encoder and decoder to process entire sequences holistically, rather than treating each sequence element independently. Incorporating ViTs with spatial-temporal attention could serve as an effective approach toward this goal. Moreover, scaling GCQ to larger and more diverse datasets would facilitate a deeper investigation into its generalization capabilities and robustness across a broader range of tasks and domains.
+
+# Acknowledgments
+
+This work was supported by the National Natural Science Foundation of China (no. T2421004 to S.W.), the Science and Technology Innovation 2030-Brain Science and Brain-inspired Intelligence Project (no. 2021ZD0200204, S.W.).
+
+# References
+
+[1] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017.  
+[2] Robert Gray. Vector quantization. IEEE Assp Magazine, 1(2):4-29, 1984.  
+[3] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022.  
+[4] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021.  
+[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563-22575, 2023.  
+[6] Tuan Vu Ho, Quoc Huy Nguyen, Masato Akagi, and Masashi Unoki. Vector-quantized variational autoencoder for phase-aware speech enhancement. 2022.  
+[7] Dominik Schmidt and Minqi Jiang. Learning to act without actions. arXiv preprint arXiv:2312.10812, 2023.  
+[8] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024.  
+[9] Torkel Hafting, Marianne Fyhn, Sturla Molden, May-Britt Moser, and Edvard I Moser. Microstructure of a spatial map in the entorhinal cortex. Nature, 436(7052):801-806, 2005.  
+[10] Christian F Doeller, Caswell Barry, and Neil Burgess. Evidence for grid cells in a human memory network. Nature, 463(7281):657-661, 2010.  
+[11] Alexandra O Constantinescu, Jill X O'Reilly, and Timothy EJ Behrens. Organizing conceptual knowledge in humans with a gridlike code. Science, 352(6292):1464-1468, 2016.  
+[12] Joshua Jacobs, Christoph T Weidemann, Jonathan F Miller, Alec Solway, John F Burke, Xue-Xin Wei, Nanthia Suthana, Michael R Sperling, Ashwini D Sharan, Itzhak Fried, et al. Direct recordings of grid-like neuronal activity in human spatial navigation. Nature neuroscience, 16(9):1188-1190, 2013.  
+[13] KiJung Yoon, Sam Lewallen, Amina A Kinkhabwala, David W Tank, and Ila R Fiete. Grid cell responses in 1d environments assessed as slices through a 2d lattice. Neuron, 89(5):1086-1099, 2016.  
+[14] Dmitriy Aronov, Rhino Nevers, and David W Tank. Mapping of a non-spatial dimension by the hippocampal-entorhinal circuit. Nature, 543(7647):719–722, 2017.  
+[15] Shun-ichi Amari. Dynamics of pattern formation in lateral-inhibition type neural fields. Biological cybernetics, 27(2):77-87, 1977.  
+[16] Si Wu, Kosuke Hamaguchi, and Shun-ichi Amari. Dynamics and computation of continuous attractors. Neural Computation, 20(4):994–1025, 2008.  
+[17] Yoram Burak and Ila R Fiete. Accurate path integration in continuous attractor network models of grid cells. PLoS computational biology, 5(2):e1000291, 2009.  
+[18] Marcella Noorman, Brad K Hulse, Vivek Jayaraman, Sandro Romani, and Ann M Hermundstad. Maintaining and updating accurate internal representations of continuous variables with a handful of neurons. Nature Neuroscience, 27(11):2207-2217, 2024.
+
+[19] Jiirgen Schmidhuber. Making the world differentiable: On using self-supervised fully recurrent n eu al networks for dynamic reinforcement learning and planning in non-stationary environments, 1990.  
+[20] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 62(1), 2022.  
+[21] Grégoire Deletang, Anian Ruoss, Paul-Ambroise Duquenne, Elliot Catt, Tim Genewein, Christopher Mattern, Jordi Grau-Moya, Li Kevin Wenliang, Matthew Aitchison, Laurent Orseau, et al. Language modeling is compression. arXiv preprint arXiv:2309.10668, 2023.  
+[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.  
+[23] Ananya Harsh Jha, Saket Anand, Maneesh Singh, and VS Rao Veeravasarapu. Disentangling factors of variation with cycle-consistent variational auto-encoders. In Proceedings of the European Conference on Computer Vision (ECCV), pages 805–820, 2018.  
+[24] Sander Dieleman, Charlie Nash, Jesse Engel, and Karen Simonyan. Variable-rate discrete representation learning. arXiv preprint arXiv:2103.06089, 2021.  
+[25] Aurko Roy, Ashish Vaswani, Arvind Neelakantan, and Niki Parmar. Theory and experiments on vector quantized autoencoders. arXiv preprint arXiv:1805.11063, 2018.  
+[26] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023.  
+[27] Junyoung Chung, Kyle Kastner, Laurent Dinh, Kratarth Goel, Aaron C Courville, and Yoshua Bengio. A recurrent latent variable model for sequential data. Advances in neural information processing systems, 28, 2015.  
+[28] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International conference on machine learning, pages 2555-2565. PMLR, 2019.  
+[29] Chang Chen, Yi-Fu Wu, Jaesik Yoon, and Sungjin Ahn. Transdreamer: Reinforcement learning with transformer world models. arXiv preprint arXiv:2202.09481, 2022.  
+[30] Mohammad Reza Samsami, Artem Zholus, Janarthanan Rajendran, and Sarath Chandar. Mastering memory tasks with world models. arXiv preprint arXiv:2403.04253, 2024.  
+[31] James Whittington, Timothy Muller, Shirely Mark, Caswell Barry, and Tim Behrens. Generalisation of structural knowledge in the hippocampal-entorhinal system. Advances in neural information processing systems, 31, 2018.  
+[32] Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models. arXiv preprint arXiv:2412.03572, 2024.  
+[33] S-I Amari. Learning patterns and pattern sequences by self-organizing nets of threshold elements. IEEE Transactions on computers, 100(11):1197-1206, 1972.  
+[34] Junfeng Zuo, Ying Nian Wu, Si Wu, and Wenhao Zhang. The motion planning neural circuit in goal-directed navigation as lie group operator search. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024.  
+[35] Xingsi Dong, Xiangyuan Peng, and Si Wu. Predictive learning in energy-based models with attractor structures. arXiv preprint arXiv:2501.13997, 2025.  
+[36] Sarthak Chandra, Sugandha Sharma, Rishidev Chaudhuri, and Ila Fiete. Episodic and associative memory from spatial scaffolds in the hippocampus. Nature, pages 1-13, 2025.  
+[37] R Ben-Yishai, R Lev Bar-Or, and H Sompolinsky. Theory of orientation tuning in visual cortex. Proceedings of the National Academy of Sciences, 92(9):3844–3848, 1995.
+
+[38] Kechen Zhang. Representation of spatial orientation by the intrinsic dynamics of the head-direction cell ensemble: a theory. The Journal of Neuroscience, 16(6):2112-2126, 1996.  
+[39] Bruce L McNaughton, Francesco P Battaglia, Ole Jensen, Edvard I Moser, and May-Britt Moser. Path integration and the neural basis of the 'cognitive map'. Nature Reviews Neuroscience, 7(8):663-678, 2006.  
+[40] Simon J Mitchell and R Angus Silver. Shunting inhibition modulates neuronal gain during synaptic excitation. Neuron, 38(3):433-445, 2003.  
+[41] C. C Alan Fung, K. Y. Michael Wong, and Si Wu. A moving bump in a continuous manifold: A comprehensive study of the tracking dynamics of continuous attractor neural networks. *Neural Computation*, 22(3):752-792, 2010.  
+[42] Sophie Deneve, Peter E Latham, and Alexandre Pouget. Reading population codes: a neural implementation of ideal observers. Nature Neuroscience, 2(8):740-745, 1999.  
+[43] Yuanyuan Mi, CC Fung, KY Wong, and Si Wu. Spike frequency adaptation implements anticipative tracking in continuous attractor neural networks. Advances in neural information processing systems, 27, 2014.  
+[44] K Wong, He Wang, Si Wu, and Chi Fung. Attractor dynamics with synaptic depression. Advances in Neural Information Processing Systems, 23, 2010.  
+[45] Wenhao Zhang, Ying Nian Wu, and Si Wu. Translation-equivariant representation in recurrent networks with a continuous manifold of attractors. Advances in Neural Information Processing Systems, 35:15770-15783, 2022.  
+[46] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.  
+[47] Karl Cobbe, Christopher Hesse, Jacob Hilton, and John Schulman. Leveraging procedural generation to benchmark reinforcement learning. arXiv preprint arXiv:1912.01588, 2019.  
+[48] Muhammad Waleed Gondal, Manuel Wuthrich, Djordje Miladinovic, Francesco Locatello, Martin Breidt, Valentin Volchkov, Joel Akpo, Olivier Bachem, Bernhard Schölkopf, and Stefan Bauer. On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019.  
+[49] Chris Burgess and Hyunjik Kim. 3d shapes dataset. https://github.com/deepmind/3dshapes-dataset/, 2018.  
+[50] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017.  
+[51] Fei Deng, Junyeong Park, and Sungjin Ahn. Facing off world model backbones: Rnns, transformers, and s4. Advances in Neural Information Processing Systems, 36:72904-72930, 2023.  
+[52] David H Hubel and Torsten N Wiesel. Receptive fields and functional architecture of monkey striate cortex. The Journal of physiology, 195(1):215-243, 1968.  
+[53] Christopher J Cueva and Xue-Xin Wei. Emergence of grid-like representations by training recurrent neural networks to perform spatial localization. arXiv preprint arXiv:1803.07770, 2018.  
+[54] Andrea Banino, Caswell Barry, Benigno Uria, Charles Blundell, Timothy Lillicrap, Piotr Mirowski, Alexander Pritzel, Martin J Chadwick, Thomas Degris, Joseph Modayil, et al. Vector-based navigation using grid-like representations in artificial agents. Nature, 557(7705):429-433, 2018.
+
+[55] Ben Sorscher, Gabriel Mel, Surya Ganguli, and Samuel Ocko. A unified theory for the origin of grid cells through the lens of pattern formation. Advances in neural information processing systems, 32, 2019.  
+[56] Stefano Recanatesi, Matthew Farrell, Guillaume Lajoie, Sophie Deneve, Mattia Rigotti, and Eric Shea-Brown. Predictive learning as a network mechanism for extracting low-dimensional latent space representations. Nature communications, 12(1):1417, 2021.  
+[57] James CR Whittington, Timothy H Muller, Shirley Mark, Guifen Chen, Caswell Barry, Neil Burgess, and Timothy EJ Behrens. The tolman-eichenbaum machine: unifying space and relational memory through generalization in the hippocampal formation. Cell, 183(5):1249-1263, 2020.  
+[58] James CR Whittington, Will Dorrell, Surya Ganguli, and Timothy EJ Behrens. Disentanglement with biological constraints: A theory of functional cell types. arXiv preprint arXiv:2210.01768, 2022.  
+[59] Yusi Chen, Huanqiu Zhang, Mia Cameron, and Terrence Sejnowski. Predictive sequence learning in the hippocampal formation. Neuron, 112(15):2645-2658, 2024.  
+[60] Ben Sorscher, Gabriel C Mel, Aran Nayebi, Lisa Giocomo, Daniel Yamins, and Surya Ganguli. When and why grid cells appear or not in trained path integrators. bioRxiv, pages 2022-11, 2022.  
+[61] Ryan Schaeffer, Mikail Khona, and Ila Fiete. No free lunch from deep learning in neuroscience: A case study through models of the entorhinal-hippocampal circuit. Advances in neural information processing systems, 35:16052–16067, 2022.  
+[62] M. GUARDAMAGNA, E. HERMANSE, J. CARPENTER, C. LYKKEN, B. DUNN, E. I. MOSER, and M.-B. MOSER. Experience-independent emergence of toroidal and ring manifolds in the entorhinal cortex. *Neuroscience Meeting Planner*. Chicago, IL.: Society for Neuroscience, 2024. Online, (PSTR190.07. 2024), 2024.  
+[63] Richard J Gardner, Erik Hermansen, Marius Pachitariu, Yoram Burak, Nils A Baas, Benjamin A Dunn, May-Britt Moser, and Edvard I Moser. Toroidal topology of population activity in grid cells. Nature, 602(7895):123-128, 2022.  
+[64] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009.
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes]
+
+Justification: Our abstract and introduction accurately reflect the paper's contributions and scope.
+
+Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: See Section 6.
+
+Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory assumptions and proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [Yes]
+
+Justification: See Section3.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.  
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental result reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+Answer: [Yes]
+
+Justification: See Section 5 and the code in supplementary material.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.  
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example  
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+# Answer: [Yes]
+
+Justification: We have included the code for reproducing the main results in supplementary material.
+
+# Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.  
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental setting/details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+# Answer: [Yes]
+
+Justification: See Section 5.
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment statistical significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+# Answer: [Yes]
+
+Justification: We used fid and PSNR to evaluate our work.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.  
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).  
+- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).  
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.
+
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments compute resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [Yes]
+
+Justification: See Section B
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code of ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes]
+
+Justification: Our work conform with the NeurIPS Code of Ethics.
+
+Guidelines:
+
+- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [NA]
+
+Justification: There is no societal impact of the work performed.
+
+Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.  
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.  
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.  
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to
+
+generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.
+
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [NA]
+
+Justification: The paper poses no such risks.
+
+Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: We cite the related paper.
+
+Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URL.  
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.  
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.  
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.  
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+# Answer: [Yes]
+
+Justification: See the documentation in supplementary material.
+
+# Guidelines:
+
+- The answer NA means that the paper does not release new assets.  
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and research with human subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+# Answer: [NA]
+
+Justification: The paper does not involve crowdsourcing nor research with human subjects.
+
+# Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+# Answer: [NA]
+
+Justification: The paper does not involve crowdsourcing nor research with human subjects.
+
+# Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.  
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.
+
+# 16. Declaration of LLM usage
+
+Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required.
+
+# Answer: [NA]
+
+Justification: LLM is used only for writing.
+
+# Guidelines:
+
+- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.  
+- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described.
+
+# A Static Setting
+
+![](images/8b72f9b42fd426ac5c802adc942ab42a4005efa5f38e9baa29df0da28c9f855b.jpg)  
+(C)
+
+![](images/73f7c0e259bfb34487477b131827b6a1e2929785ac5cfb9009efd164356b7d6b.jpg)
+
+![](images/af3d832d3bf75245e013c3de007c15c5132d399bf56008dd23ed48a03039f91e.jpg)  
+Figure 7: To evaluate the performance of GCQ in a static setting, we conducted experiments on the CIFAR-10 and ImageNet datasets. Figures (A) and (B) show that, for both GCQ and VQ, the FIDr decreases as the codebook size K increases. At 20 training epochs, GCQ outperforms VQ, whereas at 50 epochs VQ achieves better results. Figure (C) compares the reconstructions produced by GCQ and VQ on ImageNet: the first row shows the original images, the second row the reconstructions obtained with GCQ, and the third row those obtained with VQ.
+
+# B Experiment Details
+
+Here are the hyperparameters used in the experiment. All programs run on an NVIDIA A100-SXM4-80GB. The experiments reported in this paper, including the ViT, ResNet, and Hybrid networks, required 8-12 hours of training each. For ViT and hybrid architectures, we trained for 40 epochs with a learning rate of 1e-4; for the ResNet network, we trained for 100 epochs with a learning rate of 3e-4. All training runs used the Adam optimizer.
+
+# C Python Implementation
+
+Our Python implementation of GCQ is fully vectorized, relying exclusively on matrix operations without any for loops. This design makes it highly amenable to parallelization.
+
+```python
+def forward(self, latents: Tensor, label: Tensor) -> Tuple[Tensor, Tensor]:  
+    '''  
+Vector quantization for sequence data with parallel processing support for batch dimensions  
+Args:  
+    latents: Tensor of shape [B x S x D x H x W]  
+    label: Tensor of shape [B x S x 4] representing the change from current frame to next frame  
+Returns:  
+    quantized: Quantized tensor of shape [B x S x D x H x W]  
+    vq_loss: Vector quantization loss
+```
+
+111
+
+B，S，D，H，W  $=$  latents.shape
+
+N = H * W // 4
+
+device = latents(device
+
+Reshape to processable form
+
+latents  $=$  latents.view(B，S，D，4，N).permute(0，3，4，1，2）.
+
+contiguous() # [B, 4, N, S, D]
+
+factor_latents = latents.view(B, 4, N, S * D)
+
+Construct expanded embeddings without loops
+
+label: [B, S, 4] -> rearrange to [B, 4, S]
+
+shift_amounts = (label.permute(0, 2, 1).long() * self.step) # [B,
+
+4, S]
+
+Construct indices for rolling: for each shift_amount against
+
+selfembedding_weight
+
+First generate indices [K], then calculate new indices via
+
+broadcasting: new_idx = (arange(K) - shift) % K
+
+kidx = torch.arange(self.K, device=device).view(1, 1, 1, self.K)
+
+# [1,1,1,K]
+
+After expanding shift_amounts: [B,4,S,1]
+
+rolled Indices = (kidx - shift_amounts.unsqueeze(-1)) % self.K #
+
+[B，4，S，K]
+
+Using rolled Indices to extract new embeddings from
+
+embedding_weight, resulting shape [B,4,S,K,D]
+
+expanded_embedding = self embeddingweight[rolled Indices] # [B,4,S,K, D]
+
+Adjust dimensions: exchange [K] and [S] positions before merging S and D
+
+expanded_embedding = expanded_embedding.permute(0, 1, 3, 2, 4).contiguous() # [B,4,K,S,D]
+
+Finally reshape to [B, 4, K, S*D]
+
+expanded_embedding = expanded_embedding.view(B, 4, self.K, S * D)
+
+Calculate nearest neighbor indices
+
+factor_latents: [B,4,N,S*D]; expanded_embedding: [B,4,K,S*D]
+
+A = factor_latents # [B,4,N,S*D]
+
+B Expand = expanded_embedding # [B,4,K,S*D]
+
+A_sq = (A ** 2).sum(dim=-1, keepdim=True) # [B,4,N,1]
+
+B_sq = (B Expand ** 2). sum(dim=-1). unsqueeze(-2) # [B,4,1,K]
+
+cross = 2 * torch/matmul(A, B Expand.transpose(-1, -2)) # [B,4,N, K]
+
+dist = A_sq + B_sq - cross # [B,4,N,K]
+
+encoding_inds = dist.argmax(dim=-1) # [B,4,N]
+
+Sample vectors from expanded_embedding according to indices
+
+using torch.gather
+
+expanded_embedding shape [B,4,K,S*D], sampling on dim=2
+
+encoding_inds_exp = encoding_inds.unsqueeze(-1).expand(-1, -1, -1,
+
+S\*D） #[B,4,N,S\*D]
+
+embedding_results = torch.gather(expanded_embedding, 2,
+
+encoding_indsp) # [B,4,N,S*D]
+
+commitment_loss = Fmse_loss(embeding_resultsdetach(),
+
+factor_latents)
+
+embedding_results = factor_latents + (embedding_results -
+
+factor_latents).detach() # [B,4,N,S*D]
+
+embedding_results = embedding_results.view(B, 4, N, S, D)
+
+embedding_results = embedding_results.permute(0, 3, 4, 1, 2).
+
+contiguous().view(B, S, D, H, W)
+
+return embedding_results,commitment_loss

2025/Vector Quantization in the Brain_ Grid-like Codes in World Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be24d8cba4e22a22e366dc0f351da0fc203a7dbafb0c25f134c44eae34310fa7
+size 585633

2025/Venus-MAXWELL_ Efficient Learning of Protein-Mutation Stability Landscapes using Protein Language Models/d6e2fc33-adea-4d36-b95e-087103bd5d6d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df800ead936307af87af44cb61c4f459da64eec0848aea2b8e8d0f5e1a6a3658
+size 2179898

2025/Venus-MAXWELL_ Efficient Learning of Protein-Mutation Stability Landscapes using Protein Language Models/full.md

@@ -0,0 +1,719 @@
+# Venus-MAXWELL: Efficient Learning of Protein-Mutation Stability Landscapes using Protein Language Models
+
+Yuanxi  $\mathbf{Yu}^{1,2\dagger}$  Fan Jiang $^{1,2\dagger}$  Xinzhu Ma $^{3}$  Liang Zhang $^{1,2}$  Bozitao Zhong $^{1,2}$   
+Wanli Ouyang $^{3}$  Guisheng Fan $^{4}$  Huiqun Yu $^{4}$  Liang Hong $^{1,2*}$  Mingchen Li $^{1,2,3,4*}$ $^{1}$  Institute of Natural Sciences, Shanghai Jiao Tong University, China  
+ $^{2}$  Zhangjiang Institute for Advanced Study, Shanghai Jiao Tong University, China  
+ $^{3}$  Shanghai Artificial Intelligence Laboratory, China  
+ $^{4}$  School of Information Science and Engineering, East China University of Science and Technology, China
+
+# Abstract
+
+In-silico prediction of protein mutant stability, measured by the difference in Gibbs free energy change  $(\Delta \Delta G)$ , is fundamental for protein engineering. Current sequence-to-label methods typically employ the two-stage pipeline: (i) encoding mutant sequences using neural networks (e.g., transformers), followed by (ii) the  $\Delta \Delta G$  regression from the latent representations. Although these methods have demonstrated promising performance, their dependence on specialized neural network encoders significantly increases the complexity. Additionally, the requirement to individually compute latent representations for each mutant site negatively impacts computational efficiency and poses the risk of overfitting. This work proposes the Venus-MAXWELL framework, which reformulates mutation  $\Delta \Delta G$  prediction as a sequence-to-landscape task. In Venus-MAXWELL, mutations of a protein and their corresponding  $\Delta \Delta G$  values are organized into a landscape matrix, allowing our framework to learn the  $\Delta \Delta G$  landscape of a protein with a single forward and backward pass during training. Besides, to facilitate future works, we also curated a large-scale  $\Delta \Delta G$  dataset with strict controls on data leakage and redundancy to ensure robust evaluation. Venus-MAXWELL is compatible with multiple protein language models and enables these models for accurate and efficient  $\Delta \Delta G$  prediction. For example, when integrated with the ESM-IF, Venus-MAXWELL achieves higher accuracy than ThermoMPNN with  $10\times$  faster in inference speed (despite having  $50\times$  more parameters than ThermoMPNN). The training codes, model weights, and datasets are publicly available at https://github.com/ai4protein/Venus-MAXWELL.
+
+# 1 Introduction
+
+Protein stability, determined by the change in Gibbs free energy  $(\Delta G)$  of protein sequences between the native and unfolded states, plays a critical role in various applications, ranging from therapeutic protein design to industrial enzyme engineering [1, 2, 3, 4, 5, 6]. In protein engineering, accurately predicting mutation-induced stability changes  $(\Delta \Delta G)$  is critical, and current methods for assessing stability across large mutant libraries remain prohibitively expensive and resource-intensive, creating a pressing need for efficient computational solutions. Consequently, developing accurate and efficient
+
+computational tools for protein stability prediction is a longstanding focus of research, spanning physics-based and statistical methods [7, 8, 9] to machine learning approaches [10, 11, 12, 13, 14].
+
+Deep learning, exemplified by AlphaFold2 [15], has revolutionized protein-related tasks [16, 17, 18, 19, 20, 21] with protein language models (PLMs), emerging as powerful tools for capturing co-evolutionary patterns from vast sequence and structural datasets [22, 23, 24, 25, 26]. These models have shown powerful capabilities in wide-range downstream tasks, including stability prediction. Recent methods, such as Stability Oracle [13] and ThermoMPNN [14], leverage PLMs to achieve impressive performance in  $\Delta \Delta G$  estimation. However, their reliance on complex architectures and massive computational requirements leads to prohibitively high training/inference costs, fundamentally limiting their scalability for large-scale protein engineering applications. These methods can be broadly categorized as sequence-to-label approaches. Specifically, these methodologies involve a two-stage process: first, a PLM or a similar encoder model transforms a mutant sequence into a latent representation; subsequently, a regression head predicts the  $\Delta \Delta G$  value from the representation.
+
+The sequence-to-label paradigm, however, presents three main challenges: (i) Computational Inefficiency: This design requires computing a distinct latent vector for each mutant site (as shown in Figure 1 (A)), which is computationally expensive and heavily dependent on the encoder's design; (ii) Knowledge Misutilization: The regression head is often trained from scratch, failing to fully harness the rich evolutionary patterns PLMs learn during pre-training, which are implicitly captured in the model's output likelihoods or logits. (iii) Architectural Rigidity: The reliance on a specific regression head design limits versatility and complicates the integration of different PLM types. These limitations highlight the need for a generalizable and efficient approach for stability prediction.
+
+![](images/9a935698063fa92fd72e70520605c8733070536c471d36940669416d3d2f5618.jpg)  
+Figure 1: Comparison of Venus-MAXWELL and sequence-to-label models. (A): The sequence-to-label models compute a latent vector for each mutant site. (B): The Venus-MAXWELL outputs a landscape in one forward pass.
+
+![](images/72517e89bc0bffc525b113b1274499d78ceed3e8b2f9127417cf4e90fb28a943.jpg)
+
+In this work, we introduce Venus-MAXWELL (Matrix-wise landscape learning), an efficient sequence-to-landscape framework that reframes  $\Delta \Delta G$  prediction. Venus-MAXWELL is not merely a faster model but a new fine-tuning paradigm that directly addresses the limitations of sequence-to-label approaches.
+
+First, Venus-MAXWELL maximizes the utilization of pre-trained PLM knowledge. Our key insight is to formulate the framework as a differentiable, matrix-based version of the PLM's native zero-shot scoring function. Many PLMs, pre-trained on millions of sequences, can estimate the fitness like  $\Delta \Delta G$  of mutant in a "zero-shot" manner by calculating the log-likelihood difference between mutant and wild-type sequences[27, 28]. Venus-MAXWELL is designed to inherit this powerful capability. Critically, its "initial performance" (before any fine-tuning) is mathematically equivalent to the underlying PLM's zero-shot prediction performance. Therefore, our fine-tuning process does not learn from scratch but starts from a significantly advanced foundation, directly enhancing the rich evolutionary information already captured by the PLM.
+
+Second, Venus-MAXWELL introduces a novel matrix-based fine-tuning methodology. Instead of adding a task-specific regression head, our method directly fine-tunes the PLM's output logits in a matrix-wise fashion (as depicted in Figure 1 (B)), elegantly adapting the PLM for  $\Delta \Delta G$  estimation. This paradigm is highly versatile, serving as a universal fine-tuning framework compatible with diverse PLMs, including masked language models and inverse folding models.
+
+Third, as a direct benefit of this architecture, Venus-MAXWELL is exceptionally computationally efficient. To predict all single-site mutations for a protein of 200 amino acids (3800 mutants), a sequence-to-label model must execute 3800 separate computations. In stark contrast, Venus-
+
+MAXWELL accomplishes all predictions in one computational pass using the wild-type sequence, enabling high-throughput  $\Delta \Delta G$  prediction at minimal cost. This capability is particularly important for protein engineering workflows such as early-stage directed evolution, which often begin by comprehensively screening single-site mutations to identify stabilizing candidates. While complex multi-site mutant prediction is typically handled by specialized models (e.g., ECNET [29], ProteinNPT [30] and PRIME[31]), these models could offer multi-site mutant predictions based on training on single-site mutants. Venus-MAXWELL is thus positioned as a crucial tool to provide high-quality, early-stage single-site predictions, which can then be used for experimental validation and as input for downstream multi-mutant prediction models, thereby accelerating the entire directed evolution pipeline.
+
+Besides, we have observed that existing mutation  $\Delta \Delta G$  datasets often exhibit significant redundancy between training and test sets, as well as misalignment between mutant and wild-type sequences. To address these issues, we constructed a high-quality dataset derived from public resources, which underwent rigorous de-duplication and redundancy reduction based on sequence alignment. As a result, we obtained strictly separated training and test sets, containing over 226K and 12K curated mutation entries, respectively, ensuring a fair evaluation of generalization. The difference between this dataset and those in previous studies [13] lies in an intensive manual verification process to validate the sequence-mutation consistency.
+
+Utilizing this meticulously constructed  $\Delta \Delta G$  dataset for evaluation, we demonstrate that fine-tuning with our Venus-MAXWELL framework significantly enhances the predictive capabilities of various PLMs. For example, when using ESM-IF as the base model, the fine-tuned ESM-IF not only outperforms the sequence-to-label paradigm (which uses ESM-IF to generate features per mutant for regression) on the  $\Delta \Delta G$  prediction task but also surpasses ThermoMPNN, a current state-of-the-art model designed explicitly for  $\Delta \Delta G$  prediction, while achieving over 10 times faster inference.
+
+To summarize, our main contributions are as follows:
+
+A matrix-based sequence-to-landscape framework: Venus-MAXWELL offers a new efficient and broadly applicable approach for fine-tuning diverse PLMs on  $\Delta \Delta G$  prediction task. This approach uniquely maximizes the utilization of pre-trained evolutionary knowledge to enable superior stability assessment with minimal computational overhead.
+
+A qualified mutation  $\Delta \Delta G$  dataset: We provide a cleaned protein mutation  $\Delta \Delta G$  dataset. The sequences within the training and test sets have been processed to remove duplicates and reduce redundancy, ensuring that their sequence identity remains below  $30\%$ .
+
+# 2 Related Work
+
+Protein language models (PLMs). Similar to advances in natural language processing, PLMs have evolved into two dominant architectural paradigms: masked language models and auto-regressive models. Masked language models, such as ESM series [22, 24, 25, 32, 33], employ bidirectional attention to predict masked residues, capturing rich contextual representations of protein sequences. Auto-regressive models, like ProGen2 [34] and Tranception [35], predict amino acids sequentially, excelling in generative tasks. These pre-trained models have demonstrated success across various downstream applications, particularly in mutation effect prediction and protein design. Through fine-tuning, PLMs can effectively leverage their encoded evolutionary information to enhance protein fitness prediction, supporting directed evolution [29, 31, 36, 37]. Moreover, PLMs often exhibit zero-shot capabilities [22, 24], but this ability is non-directional and not specifically tailored to stability prediction. Our work offers an efficient way to convert this non-directional prediction into a directional one by fine-tuning on  $\Delta \Delta G$  mutation datasets.
+
+Physics-based methods for  $\Delta \Delta G$  prediction. Early efforts in protein stability prediction relied on physics-based methods or statistical analysis [7, 8, 9]. Rosetta [7], a widely adopted framework, employs Cartesian space sampling methods to predict stability changes based on high-resolution structural data. Similarly, FoldX [8] estimates  $\Delta \Delta G$  of mutants using an empirically derived energy function calibrated against experimental mutagenesis data. Such methods are often highly interpretable but are limited by computational complexity and input data precision, leading to lower accuracy and speed compared to contemporary deep learning methods, as we present in this work.
+
+Deep learning-based methods for  $\Delta \Delta G$  prediction. Prior to the development of pre-trained PLMs, several deep learning models have been designed to reduce the bias in predicting stabilizing mutations [10, 11, 12]. The application of PLM further advanced  $\Delta \Delta G$  prediction. For example, Stability Oracle [13] integrates a pretrained graph transformer model with a data augmentation approach based on thermodynamic permutations to address the challenges of data bias and model generalization. Leveraging the knowledge overlap between sequence recovery and stability optimization tasks, ThermoMPNN [14] utilizes the embeddings from ProteinMPNN [38] for transfer learning, resulting in a fast and robust  $\Delta \Delta G$  prediction. Despite their success, these models often require complex architectures and extensive computational resources. Our approach, however, offers a more general and efficient means of facilitating stability prediction with enhanced speed and accuracy, consequently improving its applicability to a range of directed evolution contexts.
+
+# 3 Method
+
+# 3.1 Preliminary
+
+Protein Language Model (PLM). The PLM, whether it is a GPT-style generative model or a BERT-style de-noising model, can be viewed as a parameterized function  $f_{\theta}$  that maps a protein sequence  $\mathbf{A} = (a_1, a_2, \ldots, a_n)$  to a probability distribution matrix  $\mathbf{P} \in \mathbb{R}^{n \times |\mathcal{A}|}$ :
+
+$$
+\mathbf {P} = f _ {\theta} (\mathbf {A}), \tag {1}
+$$
+
+where  $n$  is the sequence length and  $|\mathcal{A}|$  is the size of the amino acid alphabet (typically 20 for standard amino acids). The element in the  $i$ -th row and  $j$ -th column of  $\mathbf{P}$  is the probability that the  $i$ -th amino acid in the protein sequence is the  $j$ -th amino acid in the alphabet. Besides, PLMs can be extended by incorporating protein structural condition  $\mathbf{S}$ , modifying the Equation 1 to  $\mathbf{P} = f_{\theta}'(\mathbf{A}, \mathbf{S})$ . For simplicity, we omit the structural condition in the following, as it does not alter the framework.
+
+PLM for mutation scoring. PLMs are typically trained on large-scale protein databases, primarily composed of naturally occurring proteins, which exhibit higher stability due to evolutionary selection. This enables PLMs to predict mutation effects using a "zero-shot" approach by comparing the probabilities of wild-type and mutant amino acids [24]. For a mutation at position  $t$  from amino acid  $a_{t}$  to  $b_{t}$ , the predicted zero-shot score is:
+
+$$
+\operatorname {S c o r e} \left(a _ {t} \rightarrow b _ {t}\right) = \log f _ {\theta} (\mathbf {A}) _ {t, b _ {t}} - \log f _ {\theta} (\mathbf {A}) _ {t, a _ {t}}, \tag {2}
+$$
+
+where a higher score indicates a positive mutant. However, the score is often undirected (it could reflect activity, stability, or other properties). Notably, for inverse folding models, this specific log-likelihood ratio has been mathematically derived to possess a free energy interpretation directly related to  $\Delta \Delta G$  [39]. This provides a strong theoretical foundation for its use. We therefore fine-tune the PLM using  $\Delta \Delta G$  data to enable it to explicitly predict mutation  $\Delta \Delta G$ .
+
+# 3.2  $\Delta \Delta G$  Landscape Construction
+
+The  $\Delta \Delta G$  landscape of a given protein is a real-valued matrix  $\mathbf{Y} \in \mathbf{R}^{n \times |\mathsf{A}|}$ , where  $\mathbf{Y}_{i,j}$  is the  $\Delta \Delta G$  value resulting from mutating the amino acid at position  $i$  to the  $j$ -th amino acid. However, in practice, it is usually infeasible to experimentally measure  $\Delta \Delta G$  for all possible mutants, leading to a sparse landscape. To track which entries have been experimentally measured, we define a binary mask matrix  $\mathbf{M} \in \{0,1\}^{n \times |\mathsf{A}|}$ , where  $\mathbf{M}_{i,j} = 1$  if  $\mathbf{Y}_{i,j}$  is available, and  $\mathbf{M}_{i,j} = 0$  otherwise.
+
+# 3.3 Framework Architecture
+
+The overview of Venus-MAXWELL is shown in Figure 2. Specifically, by reformulating zero-shot mutation scoring (Equation 2) as matrix-driven scoring, Venus-MAXWELL enables PLMs to compute the mutation landscape in a single forward pass. Crucially, the process is differentiable, allowing PLMs to be fine-tuned on protein mutation  $\Delta \Delta G$  landscapes through backpropagation.
+
+For a protein sequence  $\mathbf{A} = (a_{1},a_{2},\dots ,a_{n})$  Venus-MAXWELL computes the landscape as follows:
+
+Sequence encoding. The sequence  $\mathbf{A}$  is first tokenized into a one-hot encoding matrix  $\mathbf{X} \in \{0,1\}^{n \times |\mathbf{A}|}$ . For each position  $i$ , the row  $\mathbf{X}_i$  is a one-hot vector where the entry corresponding to the wild-type amino acid  $a_i$  is set to 1, and all other entries are set to 0.
+
+![](images/061441a7792a8f54fb84fbd972e623fca20b867808ae6ba38401ece219739fbd.jpg)  
+Figure 2: Overview of Venus-MAXWELL. Venus-MAXWELL utilizes matrix-driven scoring to enable PLMs to learn the mutation landscape of a protein through a single forward-backward computation process.
+
+Probability prediction. The PLM computes a log-probability matrix  $\mathbf{Q} \in \mathbb{R}^{n \times |\mathcal{A}|}$ :
+
+$$
+\mathbf {Q} = \log (\mathbf {P}) = \log f _ {\theta} (\mathbf {A}), \tag {3}
+$$
+
+where  $f_{\theta}$  is the parameterized PLM function, and the logarithm is applied element-wise. Each entry  $\mathbf{Q}_{i,j}$  represents the log-probability of amino acid  $a_j$  at position  $i$ .
+
+Wild-type probability extraction. To isolate the log-probabilities associated with the wild-type amino acids, we compute the Hadamard product (element-wise multiplication) between the log-probability matrix  $\mathbf{Q}$  and the one-hot encoding matrix  $\mathbf{X}$ , yielding a matrix  $\mathbf{Q}_{\mathrm{wt}} \in \mathbf{R}^{n \times |\mathcal{A}|}$ :
+
+$$
+\mathbf {Q} _ {\mathrm {w t}} = \mathbf {Q} \circ \mathbf {X}, \tag {4}
+$$
+
+where  $\circ$  is the Hadamard product. Since  $\mathbf{X}_{i,j} = 1$  only for the wild-type amino acid  $a_{i}$  at position  $i$ ,  $\mathbf{Q}_{\mathrm{wt}}$  retains log-probabilities at wild-type positions and zeros elsewhere.
+
+Wild-type vector construction. The matrix  $\mathbf{Q}_{\mathrm{wt}}$  is then aggregated to produce a compact representation of the wild-type probabilities by summing  $\mathbf{Q}_{\mathrm{wt}}$  over the amino acid dimension:
+
+$$
+\boldsymbol {v} = \mathbf {Q} _ {\mathrm {w t}} \mathbf {1} _ {| \mathcal {A} |}, \tag {5}
+$$
+
+where  $\mathbf{1}_{|A|} \in \mathbb{R}^{|A|}$  is a unity vector and  $v_{i}$  is the log-probability of the wild-type amino acid  $a_{i}$ .
+
+Wild-type matrix replication. To enable comparison between wild-type and mutant probabilities, the wild-type log-probability vector  $\pmb{v}$  is replicated along the column dimension  $|\mathcal{A}|$  times to form a matrix  $\mathbf{V}_{\mathrm{wt}}^{\mathrm{rep}} \in \mathbb{R}^{n \times |\mathcal{A}|}$ . Specifically, we form  $\mathbf{V}_{\mathrm{wt}}^{\mathrm{rep}}$  through multiplication of  $\pmb{v}$  with the unity vector  $\mathbf{1}_{|\mathcal{A}|}^{T} \in \mathbb{R}^{1 \times |\mathcal{A}|}$ :
+
+$$
+\mathbf {V} _ {\mathrm {w t}} ^ {\operatorname {r e p}} = \boldsymbol {v} \cdot \mathbf {1} _ {| \mathcal {A} |} ^ {T}. \tag {6}
+$$
+
+In this matrix, each row  $i$  contains the log-probability of the wild-type amino acid  $a_{i}$ , repeated across all columns  $j$ .
+
+Mutation landscape computation. Finally, we subtract the replicated wild-type matrix  $\mathbf{V}_{\mathrm{wt}}^{\mathrm{rep}}$  from the log-probability matrix  $\mathbf{Q}$  to obtain the mutation landscape matrix  $\mathbf{L} \in \mathbb{R}^{n \times |\mathcal{A}|}$ :
+
+$$
+\mathbf {L} = \mathbf {Q} - \mathbf {V} _ {\mathrm {w t}} ^ {\mathrm {r e p}}. \tag {7}
+$$
+
+Each entry  $\mathbf{L}_{i,j}$  represents the log-probability difference between the  $j$ -th amino acid and the wild-type amino acid at position  $i$ . To formalize this computation, we define a function  $g_{\theta}$  as follows:
+
+$$
+\begin{array}{l} g _ {\theta} (\mathbf {A}) = \mathbf {L} = \mathbf {Q} - \mathbf {V} _ {\mathrm {w t}} ^ {\mathrm {r e p}} (8) \\ = \log f _ {\theta} (\mathbf {A}) - \left(\left(\log f _ {\theta} (\mathbf {A}) \circ \mathbf {X}\right) \cdot \mathbf {1} _ {| \mathcal {A} |}\right) \cdot \mathbf {1} _ {| \mathcal {A} |} ^ {T}. (9) \\ \end{array}
+$$
+
+It is clear that  $g_{\theta}$  is differentiable because the protein language model  $f_{\theta}$  is differentiable. Additionally, the mutation score is equal to the zero-shot scoring formula shown in Equation 2. We also provide the algorithm and pseudo-code of Venus-MAXWELL for training and landscape prediction in Appendix A.1 and Appendix A.2.
+
+# 3.3.1 Training objective
+
+The training objective ensures the predicted  $\Delta \Delta G$  landscape closely aligns with the ground-truth  $\Delta \Delta G$  landscape. Given a protein sequence  $\mathbf{X}$ , the experimentally measured mutation stability landscape  $\mathbf{Y} \in \mathbb{R}^{n \times |\mathcal{A}|}$ , and a binary mask  $\mathbf{M} \in \{0,1\}^{n \times |\mathcal{A}|}$  indicating observed entries, the model predicts a landscape  $\mathbf{L} = g_{\theta}(\mathbf{A})$ .
+
+The training loss comprises two components: Ranking loss  $\mathcal{L}_{\mathrm{Ranking}}$  and mean squared error loss  $\mathcal{L}_{\mathrm{MSE}}$ . The first component encourages correct ranking of mutations by minimizing the negative Pearson correlation between the predicted and measured  $\Delta \Delta G$  values:
+
+$$
+\mathcal {L} _ {\text {R a n k i n g}} (\mathbf {Y}, \mathbf {M}, \mathbf {L}) = 1 - \frac {\sum_ {i , j} \mathbf {M} _ {i , j} \left(\mathbf {L} _ {i , j} - \bar {\mathrm {L}}\right) \left(\mathbf {Y} _ {i , j} - \bar {\mathrm {Y}}\right)}{\sqrt {\sum_ {i , j} \mathbf {M} _ {i , j} \left(\mathbf {L} _ {i , j} - \bar {\mathrm {L}}\right) ^ {2} \cdot \sum_ {i , j} \mathbf {M} _ {i , j} \left(\mathbf {Y} _ {i , j} - \bar {\mathrm {Y}}\right) ^ {2}}}, \tag {10}
+$$
+
+where  $\bar{\mathrm{L}}$  and  $\bar{\mathrm{Y}}$  represent the mean predicted and ground-truth  $\Delta \Delta G$  values over the masked positions. The second component, a mean squared error (MSE) loss, aligns absolute values:
+
+$$
+\mathcal {L} _ {\mathrm {M S E}} (\mathbf {Y}, \mathbf {M}, \mathbf {L}) = \sum_ {i, j} \mathbf {M} _ {i, j} \left(\mathbf {Y} _ {i, j} - \tilde {\mathbf {L}} _ {i, j}\right) ^ {2}. \tag {11}
+$$
+
+Here,  $\tilde{\mathbf{L}} = \mathrm{MLP}(\mathbf{L}) = \mathbf{W}_2\sigma (\mathbf{W}_1\mathbf{L} + \boldsymbol {b}_1) + \boldsymbol {b}_2$  is the output of a two-layer multilayer perceptron (MLP), where  $\sigma (\cdot)$  denotes SELU activation function [40]. This transformation adjusts for the scale mismatch between log-probability scores and experimental  $\Delta \Delta G$  values. The parameters  $\mathbf{W}_1,\mathbf{W}_2,\pmb {b}_1,$  and  $b_{2}$  are learnable. Specifically, the learnable weights have dimensions  $\mathbf{W}_1\in \mathbb{R}^{V\times D}$  and  $\mathbf{W}_2\in \mathbb{R}^{D\times 1}$ , where  $V$  is the PLM's vocabulary size and  $D$  is the hidden dimension.
+
+The total loss is a weighted sum of both components:
+
+$$
+\mathcal {L} (\theta) = \sum_ {(\mathbf {A}, \mathbf {Y}, \mathbf {M}) \in \mathcal {D}} \mathcal {L} _ {\text {R a n k i n g}} (\mathbf {Y}, \mathbf {M}, g _ {\theta} (\mathbf {A})) + \lambda \cdot \mathcal {L} _ {\mathrm {M S E}} (\mathbf {Y}, \mathbf {M}, g _ {\theta} (\mathbf {A})), \tag {12}
+$$
+
+where  $\lambda$  is a hyperparameter weighting factor that balances the two loss components.  $\mathcal{D}$  is the training dataset, comprising multiple protein sequences annotated with a sparse mutation landscape and a corresponding binary mask. Each data point in  $\mathcal{D}$  is a tuple  $(\mathbf{A},\mathbf{Y},\mathbf{M})$ . The training objective is to minimize the total loss:  $\theta^{*} = \arg \min_{\theta}\mathcal{L}(\theta)$ . Since  $g_{\theta}$  is differential, we can utilize backpropagation and gradient descent to optimize the parameters  $\theta$ , which are initialized from the pre-trained model weights. The specific optimization details, hyperparameter selection, and implementation settings are described in Section 5.1 and Appendix A.3.
+
+# 4 Datasets Building
+
+We collected a large protein mutation  $\Delta \Delta G$  dataset, which includes 573 protein sequences and 239,408 mutations (average 418 mutations per protein) to evaluate the performance of Venus-MAXWELL. All data have been manually verified to ensure consistency between mutant and wild-type sequences, and duplicates were removed to maintain data integrity.
+
+Test set. The test set is a compilation of mutation  $\Delta \Delta G$  data that we are currently able to collect, including P53 [17], Myoglobin [17], SSym [17], S669 [41], S8754 [42], M1261 [42], vb1432 [43], Fireprotdb [44] and Thermomutdb [45]. Duplicates within the datasets are addressed by selecting the  $\Delta \Delta G$  measurement with the highest absolute value. This was intended to address the neutral bias and move the  $\Delta \Delta G$  predictions from a model away from 0. After removing duplicates, the test set contains 12,443 mutations across 308 proteins, namely 308 sparse mutation  $\Delta \Delta G$  landscape. This dataset, named Test12K, is organized by protein to enable per-protein prediction and evaluation, consistent with benchmarks like ProteinGym [27]. To ensure sufficient data for meaningful per-protein analysis, only proteins with at least five mutations were included.
+
+Training Set. The training set is derived from a large-scale dataset containing 272K protein mutation sequences, denoted as cDNA272K [46]. This dataset is a commonly used training dataset in the task of protein stability prediction [13]. To prevent potential data leakage issues, we used the MMSeqs2 tool [47] to filter the training set, as illustrated in Figure 3 (A). Specifically, we removed all sequences from the training set that share a sequence identity of  $30\%$  or higher with any sequence in the test set. This threshold  $30\%$  is chosen to maintain a significant diversity of sequences, as it lies within the "twilight zone", a region where  $95\%$  of protein pairs have distinct structural folds [48]. The resulting training set, Train226K, consists of 226K sequences with less than  $30\%$  sequence similarity to any sequences in the test set. All datasets are stored in a protein-specific format to facilitate model training and evaluation, and all sequences were folded using ColabFold 2.5 [49] to generate structural input for PLMs if needed.
+
+To further validate the separation between training and test sets, we performed pairwise global sequence alignment using the Needleman-Wunsch algorithm [50]. As shown in Figure 3 (B), the similarity distribution has a mean of  $6.17\%$ , with over  $88\%$  of test-training pairs exhibiting less than  $10\%$  similarity. This low overlap confirms that our dedduplication effectively reduces data leakage risks and enhances the model's ability to generalize to unseen sequences.
+
+# 5 Experiments and Results
+
+# 5.1 Experiments Setup
+
+The ability of Venus-MAXWELL to improve protein stability prediction  $(\Delta \Delta G)$  was systematically evaluated across a wide range of PLMs. We strategically selected representative PLMs from two distinct categories: (1) structure-aware PLMs (ESM-IF and ProSST), which incorporate structural information either explicitly or through hybrid architectures, with ESM-IF demonstrating superior zero-shot performance in stability prediction tasks [22] and ProSST representing an emerging paradigm that efficiently combines sequence and structural information [23]; and (2) sequence-only PLMs (ESM-1b, ESM2, ESM-1v), which are widely adopted in protein engineering applications and serve as pure sequence-based model representatives [24, 51, 52, 33, 53].
+
+Hyper-parameter settings. For all Venus-MAXWELL enhanced models, we utilized the Adam optimizer [54]. All key hyperparameters were selected through a rigorous 5-fold cross-validation procedure on the training set (Train226K) to prevent test set leakage. Based on this analysis, we set the initial learning rate to  $5 \times 10^{-5}$ , the loss weighting factor  $\lambda$  to 0.1, and the MLP hidden dimension  $D$  to  $V$  (the PLM's vocabulary size). The optimal number of training epochs (approximately 7, determined via early stopping with a patience of 5 epochs) was also identified during this cross-validation process. The final models were then trained on the complete Train226K dataset for the determined optimal number of epochs. A detailed description of the cross-validation methodology,
+
+![](images/a355e88db3b74528dfc61ed68e30a39c503505201327b09e729fbd98cf4d28bc.jpg)  
+Figure 3: Dataset details. (A): Dataset construction process of Train226K and Test12K. (B): Global sequence alignment using Needleman-Wunsch confirmed the low sequence similarity between Test12K and Train226K.
+
+![](images/de924d335ecba8d835855d2080db76dd5533fa0135cb31a14bb4c1f07687fe51.jpg)
+
+hyperparameter search spaces, and full ablation results for each choice is provided in Appendix A.3. All experiments were conducted on a PC with an NVIDIA RTX 4090 GPU.
+
+Evaluation metrics. To rigorously assess model performance, we employed per-protein evaluation followed by averaging across the dataset. For each protein, we calculated Pearson  $(\rho_{p})$  and Spearman  $(\rho_{s})$  correlation coefficients between predicted and experimental  $\Delta \Delta G$  values, capturing linear and monotonic relationships, respectively. For classification of stabilizing vs. destabilizing mutations (threshold at  $\Delta \Delta G = 0$ ), we computed AUC-ROC and F1 scores. This per-protein averaging approach prevents bias toward proteins with more mutations, ensuring balanced assessment across diverse protein structures [27, 28].
+
+Baselines. We compared Venus-MAXWELL against: (1) ThermoMPNN, the current state-of-the-art neural network model [55], used with its publicly available checkpoint $^2$ . Our training set is a subset of the training set of ThermoMPNN, so there are no data leakage issues. (2) FoldX and Rosetta, two representative physics-based computational methods; and (3) Embedding transfer baselines using ProSST and ESM-IF, denoted as "ProSST + MLP" and "ESM-IF + MLP". The hyperparameter settings of baseline models are shown in Appendix A.4.
+
+# 5.2 Model Performance
+
+We evaluated Venus-MAXWELL's performance on the Test12K dataset (Section 4), comprising 12,443 mutations across 308 diverse proteins. As shown in Figure 4 (A), Venus-MAXWELL enhances zero-shot performance across multiple PLMs, achieving an average Spearman correlation improvement of 0.143 (Wilcoxon signed-rank test,  $p < 0.01$ ). With ESM-IF, Venus-MAXWELL reaches a Spearman correlation of 0.517, followed by ProSST at 0.479, highlighting the importance of structure-aware representations in protein stability prediction tasks [27, 28]. This improvement stems from matrix-driven scoring (Equation 8), which maps sequences to mutation landscapes in one forward pass.
+
+![](images/b3917492eb9921450cf04093b57e43b350a2b976b8e6ab21c4f8761c0f7a9104.jpg)  
+Figure 4: Performance comparison of Venus-MAXWELL. (A): Comparison of Spearman correlation scores between zero-shot predictions and Venus-MAXWELL enhanced predictions across different PLMs. The improvement from zero-shot to Venus-MAXWELL is statistically significant (Wilcoxon signed-rank test,  $p < 0.01$ ). (B): Comparison with existing mutation prediction methods. Venus-MAXWELL (ESM-IF) outperforms ThermoMPNN (Wilcoxon signed-rank test,  $p < 0.01$ ). All evaluations were performed on the Test12K dataset, with per-protein correlations averaged across the entire set.
+
+![](images/05c625df861b0a4e82b73475c13539464405c0fa7d98bf2928b847c8425ad9db.jpg)
+
+When compared to existing methods (Figure 4 (B)), embedding transfer methods struggle under the low similarity between Train226K and Test12K (Spearman: 0.302/0.234). The physics-based Rosetta and FoldX method also demonstrates moderate performance. In contrast, the specialized architecture of ThermoMPNN delivers strong results (Spearman: 0.508). Notably, Venus-MAXWELL with ESM-IF surpasses ThermoMPNN (Spearman: 0.517,  $p < 0.01$ ) while Venus-MAXWELL with ProSST achieves comparable performance, without requiring architectural modifications to the base PLMs. To
+
+further validate these findings on established public benchmarks, we also benchmarked performance on specific, well-known subsets within Test12K (e.g., p53, s669, Myoglobin) against ThermoMPNN and Stability Oracle. As detailed in Appendix A.5, Venus-MAXWELL consistently outperforms both SOTA models on these individual datasets, confirming its robust and state-of-the-art performance. These results demonstrate that Venus-MAXWELL's matrix-based landscape approach effectively captures global stability patterns, outperforming both traditional regression methods and physics-based approaches by leveraging evolution pattern knowledge from pre-trained PLMs. For more details of comparison on additional metrics, please refer to Table A4 in Appendix A.5.
+
+Furthermore, we investigated whether the inherent sparsity of the training landscapes adversely affects generalization. We conducted an ablation study by progressively removing random portions of the training data. The results show that Venus-MAXWELL's performance is exceptionally robust, maintaining a superior performance (Spearman: 0.501) even when  $90\%$  of the training mutants are removed. This demonstrates that the model learns generalizable stability principles from the diverse collection of proteins, rather than overfitting to sparse local landscapes. The full details of this sparsity analysis are also presented in Appendix A.5.
+
+To ensure that the superior performance of Venus-MAXWELL with ESM-IF and ProSST, both of which utilize structural information, does not stem from structural data leakage, we analyzed structural similarity between Train226K and Test12K using TM-align [56]. The average TM-score between training and test set structures is approximately 0.3, indicating low structural similarity. Furthermore, the weak correlation between prediction accuracy and structural similarity (see more details in Figure A1 and Appendix A.6) confirms Venus-MAXWELL's robust generalization ability across diverse protein structures.
+
+# 5.3 Prediction Speed Analysis
+
+We evaluated Venus-MAXWELL's inference efficiency in the context of directed evolution, where rapid screening of single-site mutants is crucial. Testing on single-site saturation mutagenesis of 10 random proteins from Test12K (averaged over five sampling runs), Venus-MAXWELL demonstrates remarkable speed improvements over baseline methods.
+
+Table 1: Efficiency comparison of stability prediction methods.  
+
+<table><tr><td>Model</td><td>Trainable parameters</td><td>Prediction time (s)</td><td>Mutants per second</td></tr><tr><td>Venus-MAXWELL (ESM-IF)</td><td>142 M</td><td>2.11 �� 0.04</td><td>20050 ± 714</td></tr><tr><td>Venus-MAXWELL (ProSST)</td><td>119 M</td><td>0.53 ± 0.02</td><td>85471 ± 3512</td></tr><tr><td>ESM-IF + MLP</td><td>132 K</td><td>4744.76 ± 335.9</td><td>9 ± 1</td></tr><tr><td>ProSST + MLP</td><td>296 K</td><td>807.77 ± 7.53</td><td>51 ± 1</td></tr><tr><td>ThermoMPNN</td><td>2.68 M</td><td>20.27 ± 1.81</td><td>2088 ± 7</td></tr></table>
+
+Despite maintaining full model capacity (142 million parameters for Venus-MAXWELL (ESM-IF) and 119 million parameters for Venus-MAXWELL (ProSST)), Venus-MAXWELL achieves exceptional inference efficiency. As shown in Table 1, Venus-MAXWELL(ESM-IF) predicts 20,050 mutants per second, while Venus-MAXWELL(ProSST) reaches 85,471 mutants per second, outperforming ThermoMPNN (2,088 mutants/s) by over 10-fold and MLP-based baselines (9-51 mutants/s) by orders of magnitude. The efficiency improvement remained robust across sequence lengths (100-1000 residues), as detailed in Table A7 and A8 in Appendix A.7, where we further analyze inference speed variations with sequence length, underscoring Venus-MAXWELL's capability for parallel prediction of mutation landscapes. Notably, the training efficiency is similarly improved: Venus-MAXWELL (ESM-IF) completes one training epoch on Train226K in 41 seconds, over 200 times faster than embedding transfer methods, despite fine-tuning the entire PLM (see Figure A2 in Appendix A.8 for details). These efficiency gains, combined with the predictive accuracy reported in Section 5.2, highlight Venus-MAXWELL's scalability for high-throughput directed evolution.
+
+# 5.4 Ablation Study
+
+To dissect the contribution of Venus-MAXWELL's components, we conducted ablation studies using ESM-IF as the base model, evaluating performance on Test12K (Table 2). Our analysis reveals:
+
+Table 2: Ablation studies of Venus-MAXWELL (ESM-IF) on Test12K datasets  
+
+<table><tr><td>Model</td><td>ρs</td><td>ρp</td><td>F1</td><td>AUC</td></tr><tr><td>Venus-MAXWELL (ESM-IF)</td><td>0.517</td><td>0.542</td><td>0.427</td><td>0.678</td></tr><tr><td>Venus-MAXWELL (ESM-IF) (No L MSE)</td><td>0.506</td><td>0.528</td><td>0.330</td><td>0.638</td></tr><tr><td>Venus-MAXWELL (ESM-IF) (No LRanking)</td><td>0.496</td><td>0.518</td><td>0.410</td><td>0.671</td></tr><tr><td>Venus-MAXWELL (ESM-IF) (ESMFold)</td><td>0.492</td><td>0.519</td><td>0.389</td><td>0.662</td></tr><tr><td>Venus-MAXWELL (ESM-IF without pretrain)</td><td>0.202</td><td>0.234</td><td>0.382</td><td>0.501</td></tr><tr><td>ESM-IF + MLP</td><td>0.302</td><td>0.331</td><td>0.383</td><td>0.627</td></tr><tr><td>Zero-shot</td><td>0.375</td><td>0.376</td><td>0.218</td><td>0.575</td></tr><tr><td>Zero-shot (ESMFold)</td><td>0.354</td><td>0.360</td><td>0.217</td><td>0.574</td></tr><tr><td>ThermoMPNN</td><td>0.508</td><td>0.529</td><td>0.371</td><td>0.646</td></tr></table>
+
+Ablation on training objective. Removing the MSE module ("No  $\mathcal{L}_{\mathrm{MSE}}$ ) maintained strong mutation ranking (Spearman: 0.506) but decreased classification performance (F1: 0.330, AUC: 0.638). Conversely, removing the ranking loss component ("No  $\mathcal{L}_{\mathrm{Ranking}}$ ) better preserved classification metrics (F1: 0.410, AUC: 0.671) while showing reduced correlation scores (Spearman: 0.496). Notably, both variants remained competitive with ThermoMPNN, demonstrating the robustness of our matrix-based approach even with partial architecture.
+
+Ablation on structure quality. Using ESMFold-generated structures instead of ColabFold slightly reduced performance [51]. As a result, if the structure is predicted using ESMFold instead of ColabFold, both our model (Venus-MAXWELL (ESMFold)) and the zero-shot method experienced a slight decrease in scoring accuracy. Specifically, the Spearman correlation dropped from 0.517 to 0.492 for our model and from 0.375 to 0.354 for the zero-shot method, respectively.
+
+Ablation on pre-training impact. Random initialization of ESM-IF severely degraded performance (Spearman: 0.202), demonstrating the crucial role of pre-trained weights in capturing evolutionary patterns. All models with pre-trained weights outperformed both embedding transfer (Spearman: 0.302) and zero-shot baselines (0.375/0.354 for ESM-IF/ESMFold), confirming the effectiveness of our matrix-based landscape approach.
+
+# 6 Conclusion and Limitations
+
+In this work, we introduced Venus-MAXWELL, a universal and efficient framework that enhances protein stability prediction while preserving PLMs' inherent capabilities. Through elegant matrix transformations, our approach enables rapid modeling of the complete mutation landscape, demonstrating superior performance in both ranking and classification metrics. Extensive evaluation demonstrates a computational speedup of three orders of magnitude, establishing a scalable paradigm for protein engineering and stability optimization.
+
+While Venus-MAXWELL excels in stability prediction, its matrix-based approach holds some potential for broader applications. For example, the current matrix-based implementation is inherently limited to PLMs using single amino acids as vocabulary tokens, excluding structure-aware tokenization schemes (e.g., SaProt [57]) or BPE tokenizer [58] (e.g., ProGPT [59]). Moreover, the framework focuses on single-site saturation mutations now, and extending it to model combinatorial mutation effects poses a key challenge. Future work could address these limitations by developing adaptive tokenization strategies, incorporating multi-site mutation modeling, and evaluating Venus-MAXWELL on diverse protein function prediction tasks.
+
+# Acknowledgments and Disclosure of Funding
+
+This work was supported by the National Key Research and Development Program of China (2024YFA0917603), Computational Biology Key Program of Shanghai Science and Technology Commission (23JS1400600), and Science and Technology Innovation Key R&D Program of Chongqing (CSTB2022TIAD-STX0017).
+
+# References
+
+[1] Sasha B Ebrahimi and Devleena Samanta. Engineering protein-based therapeutics through structural and chemical design. Nature Communications, 14(1):2411, 2023.  
+[2] Steven M Jay and Richard T Lee. Protein engineering for cardiovascular therapeutics: untapped potential for cardiac repair. Circulation research, 113(7):933-943, 2013.  
+[3] Michaela Gebauer and Arne Skerra. Engineered protein scaffolds as next-generation therapeutics. Annual review of pharmacology and toxicology, 60(1):391-415, 2020.  
+[4] Elizabeth L Bell, William Finnigan, Scott P France, Anthony P Green, Martin A Hayes, Lorna J Hepworth, Sarah L Lovelock, Haruka Niikura, Silvia Osuna, Elvira Romero, et al. Biocatalysis. Nature Reviews Methods Primers, 1(1):1-21, 2021.  
+[5] Zhoutong Sun, Qian Liu, Ge Qu, Yan Feng, and Manfred T Reetz. Utility of b-factors in protein science: interpreting rigidity, flexibility, and internal motion and engineering thermostability. Chemical reviews, 119(3):1626-1665, 2019.  
+[6] Fan Jiang, Jiahao Bian, Hao Liu, Song Li, Xue Bai, Lirong Zheng, Sha Jin, Zhuo Liu, Guang-Yu Yang, and Liang Hong. Creatinase: Using increased entropy to improve the activity and thermostability. The Journal of Physical Chemistry B, 127(12):2671-2682, 2023.  
+[7] Rebecca F Alford, Andrew Leaver-Fay, Jeliazko R Jeliazkov, Matthew J O'Meara, Frank P DiMaio, Hahnbeom Park, Maxim V Shapovalov, P Douglas Renfrew, Vikram K Mulligan, Kalli Kappel, et al. The rosetta all-atom energy function for macromolecular modeling and design. Journal of chemical theory and computation, 13(6):3031–3048, 2017.  
+[8] Joost Schymkowitz, Jesper Borg, Francois Stricher, Robby Nys, Frederic Rousseau, and Luis Serrano. The foldx web server: an online force field. *Nucleic acids research*, 33(suppl_2):W382-W388, 2005.  
+[9] Shuangye Yin, Feng Ding, and Nikolay V Dokholyan. Eris: an automated estimator of protein stability. Nature methods, 4(6):466-467, 2007.  
+[10] Bian Li, Yucheng T Yang, John A Capra, and Mark B Gerstein. Predicting changes in protein thermodynamic stability upon point mutation with deep 3d convolutional neural networks. PLoS computational biology, 16(11):e1008291, 2020.  
+[11] Huali Cao, Jingxue Wang, Liping He, Yifei Qi, and John Z Zhang. Deepddg: predicting the stability change of protein point mutations using neural networks. Journal of chemical information and modeling, 59(4):1508-1514, 2019.  
+[12] Lijun Quan, Qiang Lv, and Yang Zhang. Strum: structure-based prediction of protein stability changes upon single-point mutation. Bioinformatics, 32(19):2936-2946, 2016.  
+[13] Daniel J Diaz, Chengyue Gong, Jeffrey Ouyang-Zhang, James M Loy, Jordan Wells, David Yang, Andrew D Ellington, Alexandros G Dimakis, and Adam R Klivans. Stability oracle: a structure-based graph-transformer framework for identifying stabilizing mutations. Nature Communications, 15(1):6170, 2024.  
+[14] Henry Dieckhaus, Michael Brocidiacono, Nicholas Z Randolph, and Brian Kuhlman. Transfer learning to leverage larger datasets for improved prediction of protein stability changes. Proceedings of the National Academy of Sciences, 121(6):e2314853121, 2024.  
+[15] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021.  
+[16] Ziyi Zhou, Liang Zhang, Yuanxi Yu, Banghao Wu, Mingchen Li, Liang Hong, and Pan Tan. Enhancing efficiency of protein language models with minimal wet-lab data through few-shot learning. Nature Communications, 15(1):5566, 2024.
+
+[17] Bian Li, Yucheng T Yang, John A Capra, and Mark B Gerstein. Predicting changes in protein thermodynamic stability upon point mutation with deep 3d convolutional neural networks. PLoS computational biology, 16(11):e1008291, 2020.  
+[18] Lasse M Blaabjerg, Maher M Kassem, Lydia L Good, Nicolas Jonsson, Matteo Cagiada, Kristoffer E Johansson, Wouter Boomsma, Amelie Stein, and Kresten Lindorff-Larsen. Rapid protein stability prediction using deep learning representations. *Elife*, 12:e82593, 2023.  
+[19] Shuyu Wang, Hongzhou Tang, Peng Shan, Zhaoxia Wu, and Lei Zuo. Pros-gnn: predicting effects of mutations on protein stability using graph neural networks. Computational Biology and Chemistry, 107:107952, 2023.  
+[20] Yuanxi Yu, Fan Jiang, Bozitao Zhong, Liang Hong, and Mingchen Li. Entropy-driven zero-shot deep learning model selection for viral proteins. Physical Review Research, 7(1):013229, 2025.  
+[21] Liang Zhang, Kuan Luo, Ziyi Zhou, Yuanxi Yu, Fan Jiang, Banghao Wu, Mingchen Li, and Liang Hong. A deep retrieval-enhanced meta-learning framework for enzyme optimum ph prediction. Journal of Chemical Information and Modeling, 65(7):3761-3770, 2025.  
+[22] Chloe Hsu, Robert Verkuil, Jason Liu, Zeming Lin, Brian Hie, Tom Sercu, Adam Lerer, and Alexander Rives. Learning inverse folding from millions of predicted structures. In International conference on machine learning, pages 8946-8970. PMLR, 2022.  
+[23] Mingchen Li, Yang Tan, Xinzhu Ma, Bozitao Zhong, Huiqun Yu, Ziyi Zhou, Wanli Ouyang, Bingxin Zhou, Pan Tan, and Liang Hong. ProSST: Protein language modeling with quantized structure and disentangled attention. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024.  
+[24] Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu, and Alex Rives. Language models enable zero-shot prediction of the effects of mutations on protein function. Advances in neural information processing systems, 34:29287-29303, 2021.  
+[25] Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637):1123-1130, 2023.  
+[26] Ahmed Elnaggar, Michael Heinzinger, Christian Dallago, Ghalia Rehawi, Yu Wang, Lion Jones, Tom Gibbs, Tamas Feher, Christoph Angerer, Martin Steinegger, et al. Protrans: Toward understanding the language of life through self-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 44(10):7112-7127, 2021.  
+[27] Pascal Notin, Aaron Kollasch, Daniel Ritter, Lood Van Niekerk, Steffanie Paul, Han Spinner, Nathan Rollins, Ada Shaw, Rose Orenbuch, Ruben Weitzman, et al. Proteingym: Large-scale benchmarks for protein fitness prediction and design. Advances in Neural Information Processing Systems, 36:64331-64379, 2023.  
+[28] Liang Zhang, Hua Pang, Chenghao Zhang, Song Li, Yang Tan, Fan Jiang, Mingchen Li, Yuanxi Yu, Ziyi Zhou, Banghao Wu, et al. Venusmuthub: A systematic evaluation of protein mutation effect predictors on small-scale experimental data. Acta Pharmaceutica Sinica B, 2025.  
+[29] Yunan Luo, Guangde Jiang, Tianhao Yu, Yang Liu, Lam Vo, Hantian Ding, Yufeng Su, Wesley Wei Qian, Huimin Zhao, and Jian Peng. Ecnet is an evolutionary context-integrated deep learning framework for protein engineering. Nature communications, 12(1):5743, 2021.  
+[30] Pascal Notin, Ruben Weitzman, Debora Marks, and Yarin Gal. Protein: Improving protein property prediction and design with non-parametric transformers. Advances in Neural Information Processing Systems, 36:33529-33563, 2023.  
+[31] Fan Jiang, Mingchen Li, Jiajun Dong, Yuanxi Yu, Xinyu Sun, Banghao Wu, Jin Huang, Liqi Kang, Yufeng Pei, Liang Zhang, et al. A general temperature-guided language model to design proteins of enhanced stability and activity. Science Advances, 10(48):eadr2641, 2024.
+
+[32] Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q Tran, Jonathan Deaton, Marius Wiggert, et al. Simulating 500 million years of evolution with a language model. Science, page eads0018, 2025.  
+[33] Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C Lawrence Zitnick, Jerry Ma, et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proceedings of the National Academy of Sciences, 118(15):e2016239118, 2021.  
+[34] Erik Nijkamp, Jeffrey A Ruffolo, Eli N Weinstein, Nikhil Naik, and Ali Madani. Progen2: exploring the boundaries of protein language models. Cell systems, 14(11):968-978, 2023.  
+[35] Pascal Notin, Mafalda Dias, Jonathan Frazer, Javier Marchena-Hurtado, Aidan N Gomez, Debora Marks, and Yarin Gal. Trancection: protein fitness prediction with autoregressive transformers and inference-time retrieval. In International Conference on Machine Learning, pages 16990-17017. PMLR, 2022.  
+[36] Kaiyi Jiang, Zhaoqing Yan, Matteo Di Bernardo, Samantha R Sgrizzi, Lukas Villiger, Alisan Kayabolen, BJ Kim, Josephine K Carscadden, Masahiro Hiraizumi, Hiroshi Nishimasu, et al. Rapid in silico directed evolution by a protein language model with evolvepro. Science, page eadr6006, 2024.  
+[37] Mingchen Li, Liqi Kang, Yi Xiong, Yu Guang Wang, Guisheng Fan, Pan Tan, and Liang Hong. Sesnet: sequence-structure feature-integrated deep learning method for data-efficient protein engineering. Journal of Cheminformatics, 15(1):12, 2023.  
+[38] Justas Dauparas, Ivan Anishchenko, Nathaniel Bennett, Hua Bai, Robert J Ragotte, Lukas F Milles, Basile IM Wicky, Alexis Courbet, Rob J de Haas, Neville Bethel, et al. Robust deep learning-based protein sequence design using proteinmpnn. Science, 378(6615):49-56, 2022.  
+[39] Jes Frellsen, Maher M Kassem, Tone Bengtsen, Lars Olsen, Kresten Lindorff-Larsen, Jesper Ferkinghoff-Borg, and Wouter Boomsma. Zero-shot protein stability prediction by inverse folding models: a free energy interpretation. arXiv preprint arXiv:2506.05596, 2025.  
+[40] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. Advances in neural information processing systems, 30, 2017.  
+[41] Corrado Pancotti, Silvia Benevenuta, Giovanni Birolo, Virginia Alberini, Valeria Repetto, Tiziana Sanavia, Emidio Capriotti, and Piero Fariselli. Predicting protein stability changes upon single-point mutation: a thorough comparison of the available tools on a new dataset. *Briefings in Bioinformatics*, 23(2):bbab555, 2022.  
+[42] Yunxin Xu, Di Liu, and Haipeng Gong. Improving the prediction of protein stability changes upon mutations by geometric learning and a pre-training strategy. Nature Computational Science, pages 1-11, 2024.  
+[43] Yang Yang, Siddhaling Urolagin, Abhishek Niroula, Xuesong Ding, Bairong Shen, and Mauno Vihinen. Pon-tstab: protein variant stability predictor. importance of training data quality. International journal of molecular sciences, 19(4):1009, 2018.  
+[44] Jan Stourac, Juraj Dubrava, Milos Musil, Jana Horackova, Jiri Damborsky, Stanislav Mazurenko, and David Bednar. Fireprotdb: database of manually curated protein stability data. Nucleic acids research, 49(D1):D319–D324, 2021.  
+[45] Joicymara S Xavier, Thanh-Binh Nguyen, Malancha Karmarkar, Stephanie Portelli, Pamela M Rezende, Joao PL Velloso, David B Ascher, and Douglas EV Pires. Thermomutdb: a thermodynamic database for missense mutations. Nucleic acids research, 49(D1):D475-D479, 2021.  
+[46] Kotaro Tsuboyama, Justas Dauparas, Jonathan Chen, Elodie Laine, Yasser Mohseni Behbahani, Jonathan J Weinstein, Niall M Mangan, Sergey Ovchinnikov, and Gabriel J Rocklin. Mega-scale experimental analysis of protein folding stability in biology and design. Nature, 620(7973):434-444, 2023.
+
+[47] Martin Steinegger and Johannes Söding. Mmseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nature biotechnology, 35(11):1026-1028, 2017.  
+[48] Burkhard Rost. Twilight zone of protein sequence alignments. Protein engineering, 12(2):85-94, 1999.  
+[49] Milot Mirdita, Konstantin Schütze, Yoshitaka Moriwaki, Lim Heo, Sergey Ovchinnikov, and Martin Steinegger. Colabfold: making protein folding accessible to all. Nature methods, 19(6):679-682, 2022.  
+[50] Vladimir Likic. The needleman-wunsch algorithm for sequence alignment. Lecture given at the 7th Melbourne Bioinformatics Course, Bi021 Molecular Science and Biotechnology Institute, University of Melbourne, pages 1-46, 2008.  
+[51] Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, et al. Language models of protein sequences at the scale of evolution enable accurate structure prediction. BioRxiv, 2022:500902, 2022.  
+[52] Lin Tang. Large model predicts variant effects. Nature Methods, 20(10):1448-1448, 2023.  
+[53] Benjamin J Livesey and Joseph A Marsh. Advancing variant effect prediction using protein language models. Nature genetics, 55(9):1426-1427, 2023.  
+[54] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.  
+[55] Antoni Beltran, Xiang'er Jiang, Yue Shen, and Ben Lehner. Site-saturation mutagenesis of 500 human protein domains. Nature, pages 1-10, 2025.  
+[56] Yang Zhang and Jeffrey Skolnick. Tm-align: a protein structure alignment algorithm based on the tm-score. *Nucleic acids research*, 33(7):2302-2309, 2005.  
+[57] Jin Su, Chenchen Han, Yuyang Zhou, Junjie Shan, Xibin Zhou, and Fajie Yuan. Saprot: Protein language modeling with structure-aware vocabulary. In The Twelfth International Conference on Learning Representations, 2024.  
+[58] Philip Gage. A new algorithm for data compression. The C Users Journal, 12(2):23-38, 1994.  
+[59] Noelia Ferruz, Steffen Schmidt, and Birte Höcker. Protgpt2 is a deep unsupervised language model for protein design. Nature communications, 13(1):4348, 2022.  
+[60] Jinrui Xu and Yang Zhang. How significant is a protein structure similarity with tm-score= 0.5? Bioinformatics, 26(7):889-895, 2010.
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes] We state clearly the scope and contribution of both communities of protein stability prediction and computer science in the abstract and introduction.
+
+# Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes] Section 7 discusses limitations.
+
+# Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory assumptions and proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [NA] This paper does not include theoretical results.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.
+
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental result reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+[Yes] This paper has disclosed all the necessary information to reproduce the main experimental results, including the dataset (See Section 4), model architecture (See Section 3.3), and hyper-parameters (See Section 5.1).
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.  
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example
+
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+Answer: [Yes] Our training code and datasets will be made publicly available.
+
+Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.  
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental setting/details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+Answer: [Yes] We provide pseudo-code and algorithm details of our framework in Section 5.1, Appendix A.1 and Appendix A.2. Baseline methods are provided by their authors, which requires no further supervision or learning procedures.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment statistical significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+[Yes] In Section5.2, we employ Wilcoxon signed-rank test to assess the statistical significance of performance differences between methods. In Section5.3 and AppendixA.7, we report error bars for inference speed across different models, which represent the standard deviation computed over five randomly sampled test sets. Relevant details are provided in the main text.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.  
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).  
+- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).  
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.
+
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments compute resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [Yes] Details for conducting experiments are provided at Section 5.1.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code of ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes] The general ethical conduct has been carefully gone through to make sure the listed concerns are avoided or not applicable in this research.
+
+Guidelines:
+
+- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [NA] There is no societal impact of the work performed.
+
+Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.  
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.  
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.  
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.
+
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [NA] The paper poses no such risks.
+
+Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes] We have cited all the papers for baseline methods and datasets..
+
+Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URL.  
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.  
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.  
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.  
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer: [NA] The paper does not release new assets.
+
+Guidelines:
+
+- The answer NA means that the paper does not release new assets.
+
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and research with human subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [NA] The paper does not involve crowdsourcing nor research with human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA] The paper does not involve crowdsourcing nor research with human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.  
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.
+
+# 16. Declaration of LLM usage
+
+Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required.
+
+Answer: [NA] LLM is only used only for text editing and formatting purposes.
+
+Guidelines:
+
+- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.  
+- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described.
+
+# A Appendix
+
+# A.1 Algorithm
+
+Algorithm 1 outlines the training procedure for Venus-MAXWELL. The process iterates through the dataset, computing the predicted mutation landscape  $L_{i}$  from the PLM's log-probabilities for each protein  $A_{i}$ . The model parameters  $\theta$  are then updated using gradient descent based on a combined loss function that incorporates both ranking  $(\mathcal{L}_{\mathrm{Ranking}})$  and absolute value  $(\mathcal{L}_{\mathrm{MSE}})$  objectives.
+
+Algorithm 1 Venus-MAXWELL for mutation  $\Delta \Delta G$  landscape learning  
+Require: Dataset  $D = \{(A_i,Y_i,M_i)\}_{i = 1}^N$  , where  $A_{i}$  is the  $i$  -th wild-type protein sequence,  $Y_{i}$  is the  $i$  -th mutation  $\Delta \Delta G$  landscape, and  $M_{i}$  is the  $i$  -th binary mask matrix   
+Require: Protein language model  $f_{\theta}$  with parameters  $\theta$    
+Require: Alphabet size  $|\mathcal{A}|$  of the protein language model   
+Require: Learning rate  $\eta$  ; number of epochs  $T$    
+1:  $\theta_0 = \theta$    
+2: for  $t = 1$  to  $T$  do   
+3: for  $i = 1$  to  $N$  do   
+4: Tokenize protein sequence into one-hot encoding matrix  $X_{i}$    
+5: Compute log-probability matrix  $Q_{i} = \log f_{\theta}(A_{i})$  (Equation 3)   
+6: Compute scoring matrix  $L_{i} = \log f_{\theta}(A_{i}) - ((\log f_{\theta}(A_{i})\circ X_{i})\cdot \mathbf{1}_{|\mathcal{A}|})\cdot \mathbf{1}_{|\mathcal{A}|}^{T}$  (Equation 8)   
+7: Compute loss  $\mathcal{L}(\theta) = \mathcal{L}_{\mathrm{ranking}}(Y_i,M_i,L_i) + \lambda \cdot \mathcal{L}_{\mathrm{MSE}}(Y_i,M_i,L_i)$    
+8: Update parameters:  $\theta_t\gets \theta_{t - 1} - \eta \nabla_\theta \mathcal{L}$  (Equation 10 and 11)   
+9: end for   
+10: end for   
+11:  $\theta^{\prime} = \theta_{T}$    
+12: return  $f_{\theta^{\prime}}$
+
+Algorithm 2 details the prediction process using a PLM fine-tuned by Venus-MAXWELL. Given a protein sequence  $A$ , the model performs a single forward pass to compute the log-probability matrix and subsequently derives the mutation  $\Delta \Delta G$  landscape  $L$  based on Equation 8.
+
+Algorithm 2 Venus-MAXWELL for mutation  $\Delta \Delta G$  landscape prediction  
+Require: Protein sequence A   
+Require: Protein language model  $f_{\theta}$  with fine-tuned parameters  $\theta^{\prime}$    
+1: Tokenize protein sequence into one-hot encoding matrix  $\pmb{X}$    
+2: Compute log-probability matrix  $Q_{i} = \log f_{\theta}(A_{i})$  (Equation 3)   
+3: Compute scoring matrix  $\pmb {L} = \log f_{\theta}(\pmb {A}) - ((\log f_{\theta}(\pmb {A})\circ \pmb {X})\cdot \mathbf{1}_{|\mathcal{A}|})\cdot \mathbf{1}_{|\mathcal{A}|}^{T}$  (Equation 8)   
+4: return Mutation  $\Delta \Delta G$  landscape  $\pmb{L}$
+
+# A.2 Pseudo Code
+
+Algorithm 3 provides a Pytorch-like pseudocode implementation of the core training loop for Venus-MAXWELL. It illustrates how input sequences are processed, logits are computed, and transformed into the mutation landscape matrix, and how the ranking and MSE losses are calculated and combined for backpropagation and parameter updates.
+
+Algorithm 3 Pseudocode of Venus-MAXWELL training in a pytorch-like style.  
+```python
+# model : protein language model
+for (seq, landscape, matrix) in loader:
+    # Shape [seq_len]
+    input_ids = tokenizer(seq, return_tensors="pt")
+    # Shape [seq_len, alphabet_size]
+    one-hot = F-one-hot(input_ids, num_classes=alphabet_size)
+    # Shape [seq_len]
+    logits = model(input_ids)
+    # Shape [seq_len, alphabet_size]
+    logits = logits.log softmax(dim=1)
+    # Shape [seq_len, alphabet_size] - [seq_len, 1] -> [seq_len, alphabet_size] (broadcast)
+    logits = logits - (logits * one-hot).sum(dim=-1, keep_dim=True)
+    # Shape [num_mutants, ]
+    predictions = logits[mask]
+    scores = landscape[mask]
+    rank_loss = 1 - Pearson(predictions, scores)
+    mse_loss = MSE(predictions, scores)
+    loss = rank_loss + 0.1 * mse_loss
+    lossbackward()
+    optimizer.step()
+```
+
+# A.3 Hyperparameter Selection and Ablation Studies
+
+To ensure a fair and robust selection of hyperparameters without test set leakage, we performed 5-fold cross-validation (CV) on the training dataset. The training set was split into five folds  $\{T_{1}, T_{2}, T_{3}, T_{4}, T_{5}\}$ . For each set of hyperparameters (e.g., a specific learning rate), we followed this procedure:
+
+- For  $i = 1$  to 5:  
+- Set the validation fold:  $\mathcal{D}_{\mathrm{val}} = T_{i}$ .  
+- Set the internal training set:  $\mathcal{D}_{\mathrm{train\_internal}} = \mathcal{D}_{\mathrm{train}} \setminus T_i$ .  
+- Train the model on  $\mathcal{D}_{\mathrm{train\_internal}}$  for a fixed number of epochs (e.g., 10 epochs for the initial LR search).  
+- Record the best Spearman correlation  $\rho$  achieved on the validation fold  $\mathcal{D}_{\mathrm{val}}$ .
+
+- Calculate the average Spearman  $\rho$  across all 5 folds.
+
+We define the "optimal performance" as this 5-fold mean per-protein Spearman correlation  $\bar{\rho}_s$ . A higher average score indicates a better hyperparameter choice. After identifying the optimal hyperparameters (e.g., learning rate  $5 \times 10^{-5}$ ), we determined the average optimal training epochs (approximately 7) from the CV folds. Finally, we trained a single model on the entire training dataset for 7 epochs and reported its performance on the held-out test set (Test12K). The test set was used exclusively for this final evaluation.
+
+# Ablation for Loss Weighting Factor  $\lambda$
+
+In our loss function (defined in Equation 12),  $\lambda$  balances the ranking loss  $(\mathcal{L}_{\mathrm{Ranking}})$  and the MSE loss  $(\mathcal{L}_{\mathrm{MSE}})$ . We conducted a sensitivity analysis by training Venus-MAXWELL (ESM-IF) with different values of  $\lambda$ .
+
+As shown in Table A1, the model achieved the best performance on the test dataset when  $\lambda = 0.1$ . This suggests the framework performs optimally when the ranking loss serves as the primary training objective, while the MSE loss acts as a beneficial auxiliary objective to align the scale of the predictions. Based on this analysis, we set  $\lambda = 0.1$  for all other experiments.
+
+# Ablation for Learning Rate (lr)
+
+Table A1: Ablation study on the weighting factor  $\lambda$ . We report the average Spearman  $\bar{\rho}_s$  and Pearson  $\bar{\rho}_p$  correlation coefficients of Venus-MAXWELL (ESM-IF) on the Test12K dataset.  
+
+<table><tr><td>Weighting Factor (λ)</td><td>ρs</td><td>ρp</td></tr><tr><td>0.0</td><td>0.506</td><td>0.528</td></tr><tr><td>0.1</td><td>0.517</td><td>0.542</td></tr><tr><td>0.2</td><td>0.514</td><td>0.540</td></tr><tr><td>0.4</td><td>0.504</td><td>0.538</td></tr><tr><td>0.8</td><td>0.494</td><td>0.523</td></tr><tr><td>1.0</td><td>0.486</td><td>0.519</td></tr></table>
+
+Following the 5-fold CV methodology, we performed a grid search for the optimal learning rate (lr). The results are shown in Table A2. We selected  $5 \times 10^{-5}$  as the optimal learning rate, as it demonstrated the best average performance across the validation sets.
+
+Table A2: Performance of Venus-MAXWELL (ESM-IF) with different learning rates during 5-fold cross-validation. "Avg Spearman  $\rho$  " refers to the mean Spearman correlation across the 5 validation folds. Best performance is in bold.  
+
+<table><tr><td>Learning Rate (lr)</td><td>ρs</td></tr><tr><td>5 × 10-3</td><td>0.184</td></tr><tr><td>1 × 10-3</td><td>0.278</td></tr><tr><td>5 × 10-4</td><td>0.416</td></tr><tr><td>1 × 10-4</td><td>0.499</td></tr><tr><td>5 × 10-5</td><td>0.517</td></tr><tr><td>1 × 10-5</td><td>0.501</td></tr><tr><td>5 × 10-6</td><td>0.508</td></tr><tr><td>1 × 10-6</td><td>0.492</td></tr></table>
+
+# Ablation for MLP Hidden Layer Dimension
+
+The MLP (used in Equation 11) is a two-layer network where the learnable weights have dimensions  $\mathbf{W}_1\in \mathbb{R}^{V\times D}$  and  $\mathbf{W}_2\in \mathbb{R}^{D\times 1}$ .  $V$  is the PLM vocabulary size, and  $D$  is a tunable hidden dimension. We investigated the model's sensitivity to  $D$ , using ESM-IF ( $V = 35$ ) as the base model.
+
+As shown in Table A3, the results indicate that the model achieves its best performance when the hidden dimension  $D$  is set to  $V$  (35). While performance is relatively stable for  $D$  up to 512, this analysis validates our choice of  $D = V$ , showing that increasing the MLP's complexity does not yield further benefits.
+
+Table A3: Performance of Venus-MAXWELL (ESM-IF) with varying MLP hidden layer dimensions  $(D)$ .  
+
+<table><tr><td>Hidden Dimension (D)</td><td>ρs</td></tr><tr><td>V (35)</td><td>0.517</td></tr><tr><td>64</td><td>0.506</td></tr><tr><td>128</td><td>0.516</td></tr><tr><td>256</td><td>0.515</td></tr><tr><td>512</td><td>0.505</td></tr><tr><td>1024</td><td>0.487</td></tr></table>
+
+# A.4 Evaluation Metrics and Embedding Transfer Baseline Models Configuration
+
+For embedding transfer, features from frozen PLMs were fed into a three-layer MLP with SELU activations and 0.1 dropout, trained with Adam optimizer (learning rate  $1 \times 10^{-3}$ , batch size 32). Following the similar training protocol as Venus-MAXWELL, the optimal number of training epochs
+
+was determined through five-fold cross-validation on the training set, with early stopping during validation (patience of 5 epochs based on Pearson correlation).
+
+# A.5 Model Performance Details
+
+Table A4: Model performance across different metrics  
+
+<table><tr><td>Model Type</td><td>Model</td><td>ρs</td><td>ρp</td><td>F1</td><td>AUC</td></tr><tr><td rowspan="5">PLM zero-shot</td><td>ESM-IF (Zero-shot)</td><td>0.375</td><td>0.376</td><td>0.217</td><td>0.577</td></tr><tr><td>ProSST (Zero-shot)</td><td>0.369</td><td>0.370</td><td>0.372</td><td>0.560</td></tr><tr><td>ESM-1b (Zero-shot)</td><td>0.263</td><td>0.278</td><td>0.281</td><td>0.524</td></tr><tr><td>ESM-1v (Zero-shot)</td><td>0.251</td><td>0.256</td><td>0.295</td><td>0.527</td></tr><tr><td>ESM-2 (Zero-shot)</td><td>0.268</td><td>0.284</td><td>0.289</td><td>0.533</td></tr><tr><td rowspan="3">Baseline models</td><td>ThermoMPNN</td><td>0.508</td><td>0.529</td><td>0.371</td><td>0.647</td></tr><tr><td>Rosetta</td><td>0.368</td><td>0.239</td><td>0.346</td><td>0.503</td></tr><tr><td>FoldX</td><td>0.410</td><td>0.431</td><td>0.322</td><td>0.581</td></tr><tr><td rowspan="2">Embedding transfer</td><td>ESM-IF + MLP</td><td>0.302</td><td>0.331</td><td>0.383</td><td>0.628</td></tr><tr><td>ProSST + MLP</td><td>0.234</td><td>0.263</td><td>0.360</td><td>0.608</td></tr><tr><td rowspan="5">Venus-MAXWELL (PLM)</td><td>Venus-MAXWELL (ESM-IF)</td><td>0.517</td><td>0.542</td><td>0.427</td><td>0.678</td></tr><tr><td>Venus-MAXWELL (ProSST)</td><td>0.479</td><td>0.502</td><td>0.405</td><td>0.604</td></tr><tr><td>Venus-MAXWELL (ESM-1b)</td><td>0.384</td><td>0.402</td><td>0.334</td><td>0.578</td></tr><tr><td>Venus-MAXWELL (ESM-1v)</td><td>0.358</td><td>0.371</td><td>0.324</td><td>0.555</td></tr><tr><td>Venus-MAXWELL (ESM-2)</td><td>0.427</td><td>0.450</td><td>0.366</td><td>0.545</td></tr></table>
+
+Table A4 provides a detailed breakdown of the performance for all evaluated models across the four primary metrics: Mean Spearman correlation  $(\bar{\rho}_s)$ , Mean Pearson correlation  $(\bar{\rho}_p)$ , Mean F1 score  $(F1)$ , and Mean Area Under the Curve  $(AUC)$ . Models are categorized by type (PLM zero-shot, baseline SOTA model, and physical-based method, embedding transfer, and Venus-MAXWELL) to facilitate comparison within and across different approaches. The best performance for each metric is highlighted in bold. This comprehensive table allows for a granular assessment of each model's strengths and weaknesses on the Test12K dataset.
+
+In addition to this comprehensive performance comparison, we further validated our model on specific, established public benchmarks. We clarify that Test12K is a curated meta-benchmark composed of these benchmarks (e.g., p53, s669). To directly address concerns about evaluation solely on our constructed dataset, Table A5 presents a detailed comparison against SOTA models on several individual subsets.
+
+The results show that Venus-MAXWELL consistently outperforms both ThermoMPNN and Stability Oracle on these well-known datasets. We deliberately excluded the popular ProteinGym benchmark, as its significant overlap with the training data of both our model and ThermoMPNN would lead to data leakage and invalidate the conclusions.
+
+Table A5: Model Performance on Public Benchmarks (Mean per-protein Spearman  ${\bar{\rho }}_{s}$  ).  
+
+<table><tr><td>Benchmark</td><td>Venus-MAXWELL (ESM-IF)</td><td>ThermoMPNN</td><td>Stability Oracle</td></tr><tr><td>p53</td><td>0.751</td><td>0.702</td><td>0.727</td></tr><tr><td>s669</td><td>0.565</td><td>0.545</td><td>0.503</td></tr><tr><td>Myoglobin</td><td>0.724</td><td>0.620</td><td>0.669</td></tr><tr><td>Ssym</td><td>0.602</td><td>0.516</td><td>0.567</td></tr></table>
+
+Finally, to address concerns about the inherent sparsity of the training data and its potential to adversely affect generalization, we conducted an ablation study on training set sparsity. We evaluated Venus-MAXWELL (ESM-IF) after randomly removing  $p$  percent of the mutation entries from the Train226K dataset. Performance (Mean per-protein Spearman  $\bar{\rho}_s$ ) on the Test12K set remains stable even with extreme data removal.
+
+As shown in Table A6, the framework demonstrates exceptional robustness. Performance remains remarkably stable, with a Spearman's correlation of 0.501 even when using only  $10\%$  (23K mutants) of the training data. A significant performance drop only occurs at the extreme of removing  $95\%$  of the data. This analysis confirms that our model learns generalizable stability principles from the diverse collection of proteins, rather than overfitting to sparse local landscapes.
+
+Table A6: Effect of training set sparcity on model performance.  
+
+<table><tr><td>Data Removed (p)</td><td>Training Mutants</td><td>ρs</td><td>ρp</td><td>F1</td><td>AUC</td></tr><tr><td>0%</td><td>226K</td><td>0.517</td><td>0.542</td><td>0.427</td><td>0.678</td></tr><tr><td>20%</td><td>181K</td><td>0.517</td><td>0.541</td><td>0.423</td><td>0.673</td></tr><tr><td>40%</td><td>136K</td><td>0.515</td><td>0.541</td><td>0.421</td><td>0.675</td></tr><tr><td>60%</td><td>90K</td><td>0.512</td><td>0.540</td><td>0.416</td><td>0.669</td></tr><tr><td>80%</td><td>45K</td><td>0.509</td><td>0.537</td><td>0.408</td><td>0.667</td></tr><tr><td>90%</td><td>23K</td><td>0.501</td><td>0.523</td><td>0.402</td><td>0.660</td></tr><tr><td>95%</td><td>11K</td><td>0.485</td><td>0.506</td><td>0.388</td><td>0.647</td></tr></table>
+
+# A.6 Generalization Ability Analysis Across Structural identity
+
+In constructing the training and test sets, we ensured low sequence similarity via MMseqs2 deduplication and Needleman-Wunsch validation (See Section 4). Since the top-performing Venus-MAXWELL enhanced models, ESM-IF and ProSST, leverage protein structure information, we further assessed the structural similarity between the sets to evaluate the models' extrapolation capability to structurally novel proteins. We employed TM-align [56] to compute TM-scores for each protein in Test12K against all structures in Train226K, using the maximum score as the structural similarity metric.
+
+As illustrated in Figure A1 (A), the maximum TM-scores of the test proteins displayed a tightly clustered distribution, with the overall probability density peaking around a mean value of approximately 0.31. According to established conventions in structural biology, TM-scores below 0.5 indicate distinct folding topologies [60]. This confirms the structural diversity between Train226K and Test12K, establishing a rigorous benchmark for extrapolation.
+
+Further analysis of Venus-MAXWELL (ESM-IF) performance relative to structural similarity revealed only a weak correlation  $(\rho_{p} = 0.208$ , as shown in Figure A1 (B)). This finding indicates that the performance enhancement of structure-aware protein language models within the Venus-MAXWELL framework is independent of structural similarity to training samples, demonstrating strong generalization capabilities to unseen protein architectures.
+
+![](images/ba8728885bc60cfad51c4ac767302d0e2c3e41cff222182afbd84b5c57a30405.jpg)  
+Figure A1: Model generalization of Venus-MAXWELL (ESM-IF) against structural dissimilarity. (A): Distribution of maximum TM-scores for Test12K proteins relative to Train226K, overlaid with average Pearson correlation per TM-score bin. (B): Relationship between maximum TM-score and prediction performance, showing minimal dependence on structural similarity to training proteins.
+
+![](images/9fa1d95b895e725bc4e7e7e3ba848f9e3d1031303fd1a53b2e79fe4490bad929.jpg)
+
+# A.7 Prediction Speed Scaling with Sequence Length
+
+Table A7: Prediction time (s) across sequence length ranges for single-site mutagenesis  
+
+<table><tr><td>Model</td><td>0-200</td><td>200-400</td><td>400-600</td><td>600+</td></tr><tr><td>Venus-MAXWELL (ProSST)</td><td>0.30±0.10</td><td>0.37±0.13</td><td>0.27±0.02</td><td>0.30±0.07</td></tr><tr><td>Venus-MAXWELL (ESM-IF)</td><td>1.33±0.10</td><td>1.65±0.10</td><td>1.78±0.13</td><td>2.19±0.16</td></tr><tr><td>ThermoMPNN</td><td>4.41±0.04</td><td>13.64±0.01</td><td>20.64±0.11</td><td>30.89±0.13</td></tr><tr><td>ProSST + MLP</td><td>162.19±0.87</td><td>539.72±0.67</td><td>844.52±3.68</td><td>1347.73±6.83</td></tr><tr><td>ESM-IF + MLP</td><td>421.44±1.63</td><td>2846.02±1.13</td><td>6255.54±13.25</td><td>13554.11±7.27</td></tr></table>
+
+Table A8: Mutants per second across different sequence length ranges for single-site mutagenesis  
+
+<table><tr><td>Model</td><td>0-200</td><td>200-400</td><td>400-600</td><td>600+</td></tr><tr><td>Venus-MAXWELL (ProSST)</td><td>30451±8141</td><td>82336±27736</td><td>159005±8935</td><td>217882±43378</td></tr><tr><td>Venus-MAXWELL (ESM-IF)</td><td>6513±501</td><td>16887±973</td><td>23993±1705</td><td>29425±2155</td></tr><tr><td>ThermoMPNN</td><td>1957±17</td><td>2038±2</td><td>2058±11</td><td>2080±9</td></tr><tr><td>ProSST + MLP</td><td>53.22±0.29</td><td>51.51±0.06</td><td>50.29±0.22</td><td>47.67±0.24</td></tr><tr><td>ESM-IF + MLP</td><td>20.48±0.08</td><td>9.77±0.01</td><td>6.79±0.01</td><td>4.74±0.00</td></tr></table>
+
+To assess the scalability of Venus-MAXWELL's inference efficiency across diverse protein sizes, we evaluated its prediction speed for single-site saturation mutagenesis on proteins with different sequence length ranges from the Test12K dataset. For each sequence length range (0-200, 200-400, 400-600, and  $600+$  residues), we randomly selected five proteins, performed single-site saturation mutagenesis, and measured inference speed across all models. This process was repeated five times to compute the mean and standard deviation of prediction times and throughput.
+
+As shown in Table A7, prediction times for all methods except Venus-MAXWELL (ProSST) increase with sequence length. For instance, ThermoMPNN's prediction time rises from  $4.41 \pm 0.04$  s (0–200 residues) to  $30.89 \pm 0.13$  s (600+ residues), and MLP-based baselines (ProSST-MLP and ESM-IF-MLP) exhibit even steeper increases, reaching up to  $13,554.11 \pm 7.27$  s for ESM-IF-MLP on sequences exceeding 600 residues. This trend is driven by two factors: (1) longer sequences result in a higher total number of single-site mutants, and (2) Transformer-based PLMs require significantly more time to encode longer sequences due to their quadratic complexity with respect to sequence length.
+
+In contrast, Venus-MAXWELL (ProSST) maintains rather stable prediction times (0.27–0.37 s across all ranges), likely due to its remarkable high throughput, which remains well above the demands of current sequence lengths and mutation counts. Table A8 further validates this observation. The throughput of embedding transfer methods decreases with sequence length—from  $53.22 \pm 0.29$  to  $47.67 \pm 0.24$  mutants/s for ProSST + MLP and from  $20.48 \pm 0.08$  to  $4.74 \pm 0.00$  mutants/s for ESM-IF + MLP—reflecting their need to re-encode each mutant sequence, making them highly sensitive to sequence length. ThermoMPNN, leveraging ProteinMPNN's structure to reduce encoding overhead, achieves a relatively stable throughput of approximately 2,000 mutants/s (1,957–2,080 mutants/s), but this appears to approach its practical upper limit.
+
+Conversely, both Venus-MAXWELL (ProSST) and Venus-MAXWELL (ESM-IF) exhibit significant throughput increases with sequence length, reaching orders-of-magnitude higher performance. This scalability is likely attributable to Venus-MAXWELL's optimized matrix processing architecture (see Section 3), which efficiently handles the whole mutation landscapes.
+
+# A.8 Model Training Speed
+
+This Figure A2 highlights the training efficiency of Venus-MAXWELL compared with other methods, using ESM-IF and ProSST as base models. Embedding transfer methods, which require encoding each mutant sequence during forward propagation, are significantly constrained in training speed. For ESM-IF and ProSST, processing 100,000 mutant sequences takes over 2,000 seconds, with a single training epoch exceeding one hour. ThermoMPNN substantially reduces training time by encoding
+
+![](images/f0127cab05d8df819b1c30d3c21485c15e0604888073b0254b85e78f88d1625e.jpg)  
+Figure A2: Training time analysis. (A): Comparison between Venus-MAXWELL models and their embedding transfer (MLP) counterparts for both ESM-IF and ProSST architectures. (B): Performance of Maxwell implementations contrasted with ThermoMPNN
+
+![](images/0beba3db6133a98984a23fdc60f9879aebcc15f395c76b270d87576aa32b9df1.jpg)
+
+only the wild-type sequence, processing 100,000 mutations in under 300 seconds. While the training efficiency of Venus-MAXWELLis further 20 times faster than ThermoMPNN, completing an epoch on the entire Train225K dataset in approximately 40 seconds.

2025/Venus-MAXWELL_ Efficient Learning of Protein-Mutation Stability Landscapes using Protein Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:706320da954f843e5e33e65e49511c0b7e52d174503f5e767a9f426a7beca41f
+size 755475

2025/VeriLoC_ Line-of-Code Level Prediction of Hardware Design Quality from Verilog Code/f0ee19e6-974b-4839-8629-990369c86a68_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ece0823382026039b78dbcbe05f08d6d0653afe0f4d3115c2afbc3166efa47f0
+size 8383423

2025/VeriLoC_ Line-of-Code Level Prediction of Hardware Design Quality from Verilog Code/full.md

@@ -0,0 +1,658 @@
+# VeriLoC: Line-of-Code Level Prediction of Hardware Design Quality from Verilog Code
+
+Raghu Vamshi Hemadri $^{1*}$  Jitendra Bhandari $^{1*}$  Andre Nakkab $^{1}$  Johann Knechtel $^{2}$  Badri P Gopalan $^{3}$  Ramesh Narayanaswamy $^{3}$  Ramesh Karri $^{1}$  Siddharth Garg $^{1}$
+
+$^{1}$ New York University Tandon School of Engineering
+
+$^{2}$ New York University Abu Dhabi
+
+Synopsys
+
+# Abstract
+
+Modern chip design is complex, and there is a crucial need for early-stage prediction of key design-quality metrics like timing and routing congestion directly from Verilog code (a commonly used programming language for hardware design). It is especially important yet complex to predict individual lines of code that cause timing violations or downstream routing congestion. Prior works have tried approaches like converting Verilog into an intermediate graph representation and using LLM embeddings alongside other features to predict module-level quality, but did not consider line-level quality prediction. We propose VeriLoC, the first method that predicts design quality directly from Verilog at both the line- and module-level. To this end, VeriLoC leverages recent Verilog code-generation LLMs to extract local line-level and module-level embeddings, and trains downstream classifiers/regressors on concatenations of these embeddings. VeriLoC achieves high F1-scores of 0.86–0.95 for line-level congestion and timing prediction, and reduces the mean average percentage error from  $14\% - 18\%$  for SOTA methods down to only  $4\%$ . We believe that VeriLoC embeddings and insights from our work will also be of value for other predictive and optimization tasks for complex hardware design.
+
+# 1 Introduction
+
+Modern chip design is highly complex. It begins with devising a description of the chip's behavior in a hardware description language (HDL) like Verilog. This is followed by a series of automated steps, including synthesis (where RTL code is converted into a circuit of Boolean logic and its gate implementation), placement (which arranges gates on the chip canvas), and routing (which connects gates using metal wires). This process transforms the RTL code into a manufacturable chip layout.
+
+Key metrics for design quality, like area, timing, power, routing congestion, etc., can only be verified from final layouts, but obtaining these layouts can take hours or days as synthesis, placement, routing, and other steps in the design flow are extremely complex and time-consuming. Designers often iterate multiple times till specifications and quality targets are met; these iterations can take anywhere from weeks to months, impacting time-to-market. Timing and routing congestion, in particular, are difficult to manage and are frequently the main impediment to design closure [1-5].
+
+![](images/600288906088e0e394f590f8fff314e02c079740754247b4f4298d55876d7194.jpg)  
+Figure 1: Conventional flow vs proposed for an exemplary AES design. ① The RTL code is converted into a synthesized netlist, e.g., represented by an AIG. ② The netlist is converted into a layout, with congestion arising in green areas (bounded in red). ③ Congestion information is annotated and traced back to the RTL. With VerilogC, we directly predict congestion and timing for the RTL at run-time, bypassing the time-consuming conventional steps.
+
+To address this issue, a body of recent work has proposed early-stage prediction of design quality from RTL code, typically via intermediate representations of RTL like and-inverter graphs (AIGs) $^3$  [6-10]. However, intermediate representations might lose rich semantic information available in the RTL code in compact form. For example, a 64-bit multiplier is a single line of Verilog, but corresponds to hundreds or thousands of gates in an AIG, where information must then be reverse engineered by the ML model. Recent work has leveraged large language model (LLM) based encodings of Verilog modules for accurate module-level power, performance, and area (PPA) prediction [11, 12].
+
+Aside from early-stage module-level predictions, designers can greatly from identifying individual lines of code (LoC) responsible for inducing timing violations or routing congestion. While electronic design automation (EDA) tools like RTL Architect [13] provide the capability of back-annotating the lines of code from a final layout (Step 3 in Fig. 1), these too are complex and time-consuming. Here, we pose a new research question that has not been addressed in literature: can we predict design quality, specifically, timing and routing congestion, from RTL code at the module and the individual line-of-code level?
+
+A key challenge in addressing this question is how to obtain informative RTL embeddings—here, we leverage the recent emergence of LLMs trained specifically for Verilog code generation like CL-Verilog 13B [14]. Although CL-Verilog is a decoder-only model, recent studies [15, 16] have demonstrated that internal activations from these models can yield effective embeddings. Using penultimate layer outputs from CL-Verilog as embeddings, we propose VeriLoC, a novel architecture for line-level classification of Verilog code. To the best of our knowledge, the LoC-level prediction problem has not been addressed in literature before.
+
+The key idea in the proposed VeriLoC architecture (Figure 2) is to concatenate embeddings of each line-of-code in a Verilog module with an embedding of the entire module, thus obtaining both a local and global context. In practice, we find that additionally concatenating embeddings from up to two neighboring lines further improves performance. VeriLoC then trains a supervised classifier (or regressor) on ground-truth data from Synopsys RTL Architect [17] on the OpenABCD dataset [18],
+
+using models like XGBoost [19] and LightGBM [20] tailored for scarce and imbalanced data. Our contributions are as follows:
+
+- We propose and evaluate VeriLoC, a novel LLM-based architecture for early-stage prediction of hardware design quality directly from RTL code, both at the individual lines of code and entire modules. Prior work only performs module-level predictions and converts RTL to an intermediate representation, thereby losing rich semantic information.  
+- We identify the importance of capturing both the local context, i.e., neighboring lines of code, and global context, i.e., the entire Verilog module, in enabling line-level timing and congestion predictions. VeriLoC's architecture concatenates both local and global embeddings before the final classification step.  
+- VeriLoC achieves F1-scores of 0.86 in line-level congestion prediction, 0.95 in line-level timing prediction, and also outperforms state-of-art in module-level timing prediction, reducing the mean average percentage error from  $18\%$  and  $14\%$  to only  $4\%$ .  
+- We demonstrate the usefulness of LLMs specialized for RTL code generation to also generate powerful RTL code embeddings that can be used for challenging downstream prediction tasks, specifically, timing and routing congestion prediction.
+
+Overall, VeriLoC<sup>4</sup> establishes an entirely new approach for early-stage prediction from RTL code, which might be of value not only to other prediction tasks, but also for code and design optimization.
+
+# 2 Background and Related Work
+
+We discuss relevant background on hardware design and contrast VeriLoC with related work on predicting design quality and on LLM-based prediction of code quality.
+
+# 2.1 Hardware Design: Quality Metrics and Prediction
+
+# 2.1.1 Routing Congestion
+
+What is Routing Congestion? Routing is one of the most complex and time-consuming steps in hardware design. Routing entails interconnecting logic gates with wires after the gates are placed on the chip canvas. Modern chips have tens of different routing layers, where any two wires that need to cross without connecting electrically can be routed above another, akin to a flyover in a traffic network. Almost every problem related to routing is known to be intractable [21]. Thus, like most processes in EDA, routing is heavily reliant on heuristic optimization, which cannot guarantee best quality in one go. In this context, managing routing congestion—or congestion for short—is important. This arises when multiple wires pass through the same small area of a chip, such that, in the worst case, the number of routing layers is insufficient to route all wires correctly, i.e., without at least two wires crossing paths. When this happens, the entire design might need to be undergo placement again, or might even necessitate an RTL rewrite.
+
+Predicting Routing Congestion. Traditional methods commonly integrate actual routing processes [22-25] or analytical models that estimate congestion [26-29]. However, routing-based methods are plagued by considerable runtime cost while analytical-based approaches suffer from relatively low accuracy. To address these challenges, more recent works have employed ML techniques. For example, [30] utilize convolutional neural networks (CNNs) to predict the overall routability of placement solutions. In a follow-up work, [31] employ deep neural networks (DNNs) to achieve better performance for congestion prediction and guide toward less-congested placement. Furthermore, [32-37] all use graph neural networks (GNNs) using synthesized and/or placed netlists as inputs, which require running time-consuming synthesis and placement tools, respectively.
+
+# 2.1.2 Timing
+
+What is Timing? Timing is a critical aspect of chip design and determines the fastest frequency at which the chip can operate. Timing is affected by every process in the design cycle, but the most critical impact is within the RTL stage, as this dictates the architecture and data flow of the IC. Roughly, timing refers to the time it takes for data to propagate from a circuit's input to its output;
+
+thus, the longest sequence of gates from the input to the output is referred to as the critical path. Often, timing is measured by worst negative slack (WNS), i.e., the difference/slack between the desired critical-path delay and the actual critical-path delay. The goal for designers is to push WNS above zero, i.e., to keep delays within the desired budget.
+
+Timing Prediction. Prior works predict timing by employing various ML techniques at various design stages. Closest to our work, [1-4] predict timing for entire modules at the RTL stage, but use either AIGs or other intermediate representations. Of these, [4, 1] are the current state-of-the-art methods. VeriLoC demonstrates substantial accuracy improvements compared to both. Other methods propose timing prediction at later stages in the design, including after synthesis [5] and after placement [38, 39]. All of these methods are focused on module-level timing prediction; VeriLoC is the first method to provide line-of-code level predictions of WNS.
+
+# 2.2 LLMs for Code Generation and Quality Prediction
+
+# 2.2.1 LLMs for Software Code Quality
+
+LLMs have demonstrated significant potential for coding, with applications spanning bug detection, program synthesis, and performance optimization [40]. Models like CodeBERT, GraphCodeBERT, and CodeT5 effectively capture syntactic and semantic nuances in high-level programming languages, making them invaluable for tasks like code summarization, translation, and repair [41, 42]. The aforementioned LLMs excel at these generative tasks. Bug detection can be viewed as a line-level prediction task, and has been addressed via pattern matching using static analysis [43], enhanced with LLMs [44], or by using LLMs for test generation and fuzzing [45]. In most instances, given the massive amounts of open-source software and vulnerability datasets, these methods can leverage LLMs with careful prompt tuning, retrieval, and agentic frameworks. Indeed, state-of-art approaches like LLMSAN [46] utilize few-shot chain-of-thought prompting to extract structured data-flow paths for bug detection, but do not make any architectural modifications. Unfortunately, data is scarce in hardware, and concepts like routing congestion are barely mentioned. As we show later, prompting methods fail completely for line-level congestion and timing estimation.
+
+# 2.2.2 LLMs for Hardware
+
+While LLMs originally targeted software code, recent work has extended their use to hardware description languages such as Verilog. Generative Verilog models (e.g., VeriGen [47], CLVerilog [14], RTLCoder [48] and Others [49-52]) achieve impressive synthesis quality but do not provide downstream quality-of-results (QoR) metrics. Building on this trend, RTLRewriter applies LLM-guided rewriting for optimization [53], and RTLFixer employs LLM-driven debugging to correct syntax errors at scale [54]. Beyond generative tasks, LLMs have begun to assist QoR estimation for rapid design-space exploration. For PPA estimation, multimodal techniques—including CircuitFusion and VeriDistill hardware code with structural or graph-based embeddings to predict power, performance, and area [12, 11]. However, these methods operate at module or graph granularity, leaving line-level semantics unexplored. To the best of our knowledge, VeriLoC is the first ever line-level QoR predictor using a hardware-specialized LLM, enabling prediction of timing and congestion metrics directly at the statement level in Verilog code.
+
+# 3 Methodology
+
+Overview. VeriLoC builds on the premise that LLMs customized for RTL/Verilog code generation that have recently begun to emerge can be also be used as embeddings that capture the semantics of RTL code and used for downstream prediction tasks. VeriLoC is the first to demonstrate this property in the hardware context. We illustrate our methodology in Fig. 2. Our approach relies on embeddings generated by CL-Verilog [14], a variant of LLaMA-2 fine-tuned on Verilog code, extracted from its penultimate layer activations. We use these embeddings hierarchically, offering semantic representations at both the module-(Sec. 3.1) and line-level (Sec. 3.2), potentially with more context from neighboring lines (Sec. 3.2). These embeddings are projected to a lower dimension (Sec. 3.3), concatenated and a final classification/regression head outputs line- and module-level predictions (Sec. 3.5).
+
+![](images/57e7dcb870a238d2b8c64695f7f85a556cdf24442e2a13e702debfa054a37c9a.jpg)  
+Figure 2: The architecture of VeriLoC for line-level timing and congestion prediction from RTL. Module-level prediction uses module embedding. The context window is set to  $p = 5$  in this example.
+
+# 3.1 Module-Level Embeddings
+
+Modules in RTL designs are represented as sequences of lines of Verilog:  $M = \{l_1, l_2, \dots, l_n\}$ , where  $l_i$  is the  $i$ -th line of code in the module. To capture the global semantics of the module, the complete module is passed through CL-Verilog, which generates hidden states  $H = \{h_1, h_2, \dots, h_n\}$ , where  $h_i \in \mathbb{R}^k$  is the latent vector for the  $i$ -th token of CL-Verilog's tokenizer, and  $k$  is the dimensionality of the model's hidden state. Module embeddings  $e(M)$  are computed as the pooled dot products of the hidden states and attention mask, normalized by sum of the attention mask:
+
+$$
+e (M) = \frac {\sum_ {i = 1} ^ {n} \left(h _ {i} \cdot m _ {i}\right)}{\sum_ {i = 1} ^ {n} m _ {i}}.
+$$
+
+where  $m_i \in \{0,1\}$  is the attention mask that ensures only valid tokens contribute to the embedding, and the dot product  $h_i \cdot m_i$  highlights the importance of each hidden state relative to the mask. The resulting module embedding  $e(M) \in \mathbb{R}^k$  provides a condensed global representation of the module, enabling the detection of macro-level patterns such as resource utilization and timing violations.
+
+# 3.2 Line-Level Embeddings
+
+To capture localized semantics of Verilog and their impact on design quality, embeddings are also generated for individual lines of code. Each line  $l_{i}$  is passed independently through CL-Verilog, producing a hidden state  $h_{i}$ . Similar to module embeddings, the line embedding  $e(l_{i})$  is computed using the attention-weighted pooling mechanism:
+
+$$
+e (l _ {i}) = \frac {\sum_ {i = 1} ^ {n} \left(h _ {i} \cdot m _ {i}\right)}{\sum_ {i = 1} ^ {n} m _ {i}}.
+$$
+
+These embeddings  $e(l_i) \in \mathbb{R}^k$  focus on the specific performance characteristics of each line, such as whether it contributes to congestion or WNS.
+
+# 3.3 Dimensionality Reduction
+
+After extracting module-level and line-level embeddings, dimensionality reduction is applied to ensure computational efficiency and improve downstream task performance. The combined embedding  $x_{i} = [e(l_{i});e(M)]$  undergoes dimensionality reduction as follows. An encoder-decoder framework is trained to reconstruct the original concatenated embedding  $x_{i}$  from its reduced representation. The encoder maps the high-dimensional input  $x_{i}$  to a lower-dimensional space  $\mathbb{R}^d$ :  $z_{i} = \text{Encoder}(x_{i})$ , while the decoder reconstructs  $x_{i}$  from  $z_{i}$ :  $\Phi (x_{i}) = \text{Decoder}(\text{Encoder}(x_{i}))$ . The framework is optimized to minimize the reconstruction loss:  $\mathcal{L} = \| x_i - \Phi (x_i)\| ^2$ . Once training is complete, only the encoder is retained for dimensionality reduction. In short, the encoder provides compact embeddings  $z_{i}$  and serves as a pre-trained initialization for downstream classification and regression.
+
+# 3.4 Contextual Feature Augmentation
+
+Contextual feature augmentation enhances the representation of a target line by integrating dependencies from surrounding lines. It captures sequential patterns and inter-line relationships, which are essential for analyzing RTL code. The embeddings of a target line  $l_{i}$  are concatenated with those of its neighboring lines within a context window  $p$ . For any  $l_{i}$ , the augmented embedding is:  $z_{\mathrm{aug}}(l_i) = [z_{i - p};\ldots ;z_i;\ldots ;z_{i + p}]$ , where  $z_{i}$  is the reduced embedding and  $[\cdot ]$  denotes vector concatenation. This approach introduces local dependencies, enabling the classifier to capture the sequential nature of RTL designs.
+
+As shown in Fig. 3 (left), the statement always @(posedge clk) begins is not flagged, but in Fig. 3 (right), with maybe_full <= N17; introduced, the same line always @(posedge clk) begins becomes congestion-causing. This shows context-aware analysis can enhance detection by considering dependencies between neighbouring lines.
+
+![](images/ada949770b73e8ba09264eb04d695c22276f08f58155cdd102382d7a4eaef293.jpg)  
+Figure 3: Effect of neighbor embeddings in context-aware congestion and timing detection.
+
+# 3.5 Classification and Regression Heads
+
+The concatenated embeddings feed into a classification head for line-level classification of code that cause congestion and timing issues, and a regression head for WNS prediction. We compare three classification/regression heads: (1) Feedforward Neural Networks (FNNs), a single-layer and fully connected neural network that replaces CL-Verilog's original classification head, but with a single classification (or regression) output; (2) XGBoost, a gradient-boosted tree [19]; (3) LightGBM, a lightweight gradient-boosting framework optimized for speed and performance [55]. XGBoost and LightGBM are used because of their demonstrated performance on imbalanced datasets [56]. This is essential for our work as only a small number of lines of code cause congestion or timing issues.
+
+Although our primary focus is LoC-level prediction, we also use VeriLoC to estimate module-level WNS. Specifically, we first estimate WNS at the line level and then select the worst (smallest) value across all lines. This line-wise granular prediction strategy allows the model to capture granular WNS estimates, and improves upon state-of-art module-level WNS predictors that use only module-level embeddings or features. This formulation, to the best of our knowledge, is unique to VeriLoC.
+
+# 4 Empirical Evaluation
+
+# 4.1 Experimental Setting
+
+Dataset. We use the popular OpenABCD [18] RTL/Verilog code dataset for our experiments, using various Verilog modules from all projects in the dataset. We employed an 80/20 random split of the dataset to obtain training vs. test data. Dataset characteristics are shown in Table 1.
+
+To generate labels for timing and congestion, we use RTL Architect from Synopsys [13], transforming all designs to their physical layout, using the open-source Nangate 45nm standard-cell library. We ran all our designs with an aggressive timing constraint of 0.25 ns, because Synopsys RTL-A only reports WNS when a timing constraints are actually violated.
+
+Table 1: Characteristics of OpenABCD and extracted Verilog. Designs have between 300-30K LOC.  
+
+<table><tr><td>Design</td><td># of Modules</td><td># of Lines</td><td>Design</td><td># of Modules</td><td># of Lines</td></tr><tr><td>aes</td><td>2</td><td>301</td><td>coyote</td><td>114</td><td>176279</td></tr><tr><td>ariane</td><td>39</td><td>214930</td><td>dynamic_node</td><td>9</td><td>796</td></tr><tr><td>black_parrot</td><td>88</td><td>62948</td><td>ethmac</td><td>10</td><td>1168</td></tr><tr><td>bp_be_top</td><td>38</td><td>13289</td><td>jpeg</td><td>5</td><td>669</td></tr><tr><td>bp_fe_top</td><td>15</td><td>7363</td><td>microwatt</td><td>31</td><td>26033</td></tr><tr><td>bpmulti_top</td><td>89</td><td>32959</td><td>swerv wrapper</td><td>57</td><td>16496</td></tr><tr><td>bp_quad</td><td>252</td><td>293281</td><td>vanilla5</td><td>39</td><td>11577</td></tr></table>
+
+balance, max_depth=30, learning_rate=0.05 and n_estimators=500. LightGBM used is_unbalance=True for automatic class weight adjustment, and default settings of num_leaves=100, learning_rate=0.05, and feature_fraction=0.8. The regression head followed a similar training procedure using the 'XGBRegressor' with a squared error loss. The FNN consisted of a single sigmoid neuron trained with binary cross-entropy (BCE) loss and was optimized using Adam with a learning rate of  $1e^{-4}$ .
+
+Metrics. For congestion and timing classification tasks, we use the F1-score, precision, and recall to measure the balance between sensitivity and specificity. For the regression task of WNS prediction, we employ  $\mathbf{R}^2$  and mean absolute percentage error (MAPE), providing insights for goodness-of-fit and prediction error relative to the target.
+
+Hardware. CL-Verilog feature extraction was performed on a single NVidia H100, and downstream classifiers (XGBoost and LightGBM) were trained/evaluated on a CPU machine with 32GB RAM and 8 CPU cores. The FNN model was trained/evaluated using an NVidia RTX 8000 GPU.
+
+# 4.2 Line-level Classification Results
+
+We begin by discussing VeriLoC's performance on line-level classification for both congestion and timing prediction. Table 2 tabulates our results for three different classification heads, as well as different context lengths  $(p = \{1,3,5\})$
+
+Congestion Detection. As shown in Table 2, the highest F1-score for congestion detection is 0.86, achieved by the LightGBM classifier with a context length of 5. This result underscores the importance of contextual information in accurately identifying congestion-causing lines.
+
+Timing Detection. The highest F1-score of 0.95 is obtained using the XGBoost and LightGBM classifiers with a context length of 5, similar to the best performing model for congestion detection. Interestingly, timing prediction results are less sensitive to local context compared to congestion prediction. We achieve F1-scores of 0.83 for timing prediction even without local context  $(p = 0)$ .
+
+These results reinforce prior observations about the advantage of XGBoost and LightGBM over deep networks in handling imbalanced datasets and irregular feature distributions [56], albeit these results were in the context of tabular data. XGBoost and LightGBM improve F1-scores
+
+Table 2: Performance of VeriLoC on line-level congestion and timing detection. Best results are highlighted in blue.  
+
+<table><tr><td rowspan="2">Classifier</td><td rowspan="2">Context Length</td><td colspan="3">Congestion</td><td colspan="3">Timing</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td rowspan="3">FNN</td><td>0</td><td>0.38</td><td>0.74</td><td>0.50</td><td>0.67</td><td>0.88</td><td>0.76</td></tr><tr><td>3</td><td>0.41</td><td>0.77</td><td>0.54</td><td>0.71</td><td>0.89</td><td>0.80</td></tr><tr><td>5</td><td>0.86</td><td>0.70</td><td>0.77</td><td>0.76</td><td>0.92</td><td>0.83</td></tr><tr><td rowspan="3">XGB</td><td>0</td><td>0.38</td><td>0.74</td><td>0.50</td><td>0.76</td><td>0.92</td><td>0.83</td></tr><tr><td>3</td><td>0.42</td><td>0.78</td><td>0.55</td><td>0.91</td><td>0.93</td><td>0.92</td></tr><tr><td>5</td><td>0.94</td><td>0.78</td><td>0.85</td><td>0.94</td><td>0.94</td><td>0.94</td></tr><tr><td rowspan="3">LGBM</td><td>0</td><td>0.38</td><td>0.78</td><td>0.51</td><td>0.82</td><td>0.91</td><td>0.83</td></tr><tr><td>3</td><td>0.41</td><td>0.76</td><td>0.53</td><td>0.93</td><td>0.92</td><td>0.92</td></tr><tr><td>5</td><td>0.94</td><td>0.79</td><td>0.86</td><td>0.96</td><td>0.94</td><td>0.95</td></tr></table>
+
+from 0.77 to 0.86 for congestion prediction and 0.83 to 0.95 for timing prediction.
+
+As shown in Table 2, they consistently outperform FNNs in both congestion and timing detection, particularly with a larger context of 5. The highest F1-scores for congestion detection (0.86) and timing detection (0.95) were achieved by these models, reinforcing their robustness in handling
+
+![](images/58c952ec166c34b0c57f0e89c50c1eb972dc78cd79bdd59ac3887461b1221180.jpg)  
+Figure 4: Scatter plots of actual vs. predicted line-level WNS using VeriLoC for three Verilog projects. In most instances, the predictions follow the actual WNS closely.
+
+![](images/0a7b493bc1b371815af830f2aa92e36b1fb820683e39b7debc4cfd3da4a557a8.jpg)
+
+![](images/f180270ae64f49a79522091904e108fca2ffe62381fc4fa64204e03ee2633ffc.jpg)
+
+imbalanced datasets. The computational efficiency of XGBoost and its robustness in optimizing for minority class representation contributed to its superior performance compared to FNNs.
+
+# 4.3 Timing Prediction and Comaprisons with SoTA
+
+Table 3: Line-level prediction of WNS using VeriLoC. Module level prediction obtained from line-level ones.  
+
+<table><tr><td>Design</td><td>R²</td><td>MAPE</td><td>Design</td><td>R²</td><td>MAPE</td></tr><tr><td>aes</td><td>0.97</td><td>0.03</td><td>coyote</td><td>0.99</td><td>0.03</td></tr><tr><td>ariane</td><td>0.96</td><td>0.06</td><td>dynamic_node</td><td>0.76</td><td>0.18</td></tr><tr><td>black_parrot</td><td>0.99</td><td>0.01</td><td>ethmac</td><td>0.96</td><td>0.05</td></tr><tr><td>bp_be_top</td><td>0.76</td><td>0.27</td><td>jpeg</td><td>0.99</td><td>0.07</td></tr><tr><td>bp_fe_top</td><td>0.99</td><td>0.02</td><td>microwatt</td><td>0.93</td><td>0.21</td></tr><tr><td>bp-multi_top</td><td>0.99</td><td>0.02</td><td>swervwrapper</td><td>0.94</td><td>0.08</td></tr><tr><td>bp_quad</td><td>0.98</td><td>0.02</td><td>vanilla5</td><td>0.99</td><td>0.02</td></tr></table>
+
+Table 3 reports VeriLoC's performance for predicting WNS at the line-level. Across all designs, VeriLoC's achieves high  $R^2$  and low MAPE across all designs but two. The accuracy of our WNS predictions are also evident in Figure 4, where we compare actual vs. predicted WNS for all lines in three benchmarks.
+
+We now compare VeriLoC's module-level WNS prediction against state-of
+
+art approaches. Recall that VeriLoC's module-level WNS predictions are obtained by first predicting WNS of each line in the code and then picking the smallest WNS, while other methods only use entire module-level features. Table 4 compares VeriLoC with MasterRTL [4] and RTL-Timer [1], SoTA methods that use handcrafted features derived from RTL. Across designs, VeriLoC outperforms prior art, with higher  $R^2$  values (always  $>0.90$ ) and much lower MAPE (always  $<0.10$ ).
+
+Table 4: VeriLoC vs. SOTA on module-level WNS (timing) prediction. VeriLoC substantially improves  $\mathbb{R}^2$  and MAPE. Best results are in blue.  
+
+<table><tr><td rowspan="2">Design</td><td colspan="2">MasterRTL</td><td colspan="2">RTL-Timer</td><td colspan="2">VeriLoC</td></tr><tr><td>R2</td><td>MAPE</td><td>R2</td><td>MAPE</td><td>R2</td><td>MAPE</td></tr><tr><td>aes</td><td>0.67</td><td>0.26</td><td>0.76</td><td>0.23</td><td>0.97</td><td>0.08</td></tr><tr><td>ariane</td><td>0.71</td><td>0.15</td><td>0.79</td><td>0.11</td><td>0.93</td><td>0.07</td></tr><tr><td>black_parrot</td><td>0.75</td><td>0.10</td><td>0.83</td><td>0.07</td><td>0.99</td><td>0.02</td></tr><tr><td>bp_be_top</td><td>0.76</td><td>0.12</td><td>0.87</td><td>0.08</td><td>0.98</td><td>0.08</td></tr><tr><td>bp_fe_top</td><td>0.80</td><td>0.09</td><td>0.88</td><td>0.05</td><td>0.98</td><td>0.04</td></tr><tr><td>bp-multi_top</td><td>0.61</td><td>0.16</td><td>0.67</td><td>0.13</td><td>0.99</td><td>0.02</td></tr><tr><td>bp_quad</td><td>0.69</td><td>0.14</td><td>0.78</td><td>0.10</td><td>0.94</td><td>0.06</td></tr><tr><td>coyote</td><td>0.74</td><td>0.15</td><td>0.83</td><td>0.12</td><td>0.99</td><td>0.03</td></tr><tr><td>dynamic_node</td><td>0.73</td><td>0.20</td><td>0.80</td><td>0.16</td><td>0.99</td><td>0.06</td></tr><tr><td>ethmac</td><td>0.77</td><td>0.24</td><td>0.83</td><td>0.21</td><td>0.99</td><td>0.03</td></tr><tr><td>jpeg</td><td>0.82</td><td>0.29</td><td>0.90</td><td>0.26</td><td>0.98</td><td>0.04</td></tr><tr><td>microwatt</td><td>0.81</td><td>0.20</td><td>0.80</td><td>0.16</td><td>0.99</td><td>0.02</td></tr><tr><td>swerv wrapper</td><td>0.69</td><td>0.27</td><td>0.76</td><td>0.24</td><td>0.95</td><td>0.07</td></tr><tr><td>vanilla5</td><td>0.68</td><td>0.19</td><td>0.74</td><td>0.16</td><td>0.98</td><td>0.07</td></tr></table>
+
+Table 5 compares VeriLoC with three additional baselines on module-level WNS: GNN-based predictors from synthesized netlists [10], VeriDistill [11] that uses LLM-based RTL embeddings with GNN-based synthesized look-up-table (LUT) embeddings, and as an ablation, VeriLoC-mod, a version of VeriLoC (VeriLoC-mod) that only uses module-level but no line-level embeddings. VeriLoC achieves the best performance, notably improving upon VeriLoC-mod, demonstrating the value of line-level embeddings even for module-level timing prediction. VeriDistill is second on MAPE, but performs poorly on  $R^2$ .
+
+Direct comparisons of VeriLoC with
+
+CircuitFusion [12] method, a very recent, state-of-art, module-level multimodal PPA predictor, are challenging because they use a different dataset. Still, when comparing the respective improvements upon RTL-Timer, we see that CircuitFusion reports an increase in  $\mathbb{R}^2$  from  $0.81\rightarrow 0.83$  and a reduction in MAPE from  $16\% \to 11\%$  , whereas VeriLoC demonstrates larger relative gains, with  $\mathbb{R}^2$  increasing from  $0.86\rightarrow 0.94$  , and MAPE reducing from  $12\% \rightarrow 6\%$  . We caution against reading more
+
+Table 5: Comparison with SOTA methods for timing prediction on the OpenABCD benchmark. VeriLoC-mod uses only module embeddings but no line embeddings.  
+
+<table><tr><td>Metric</td><td>GNN [10]</td><td>MasterRTL [4]</td><td>RTL-Timer [1]</td><td>VeriDistill [11]</td><td>VeriLoC-Mod</td><td>VeriLoC</td></tr><tr><td>R2</td><td>0.69</td><td>0.74</td><td>0.86</td><td>0.728</td><td>0.91</td><td>0.94</td></tr><tr><td>MAPE</td><td>0.17</td><td>0.15</td><td>0.12</td><td>0.076</td><td>0.08</td><td>0.06</td></tr></table>
+
+into these comparisons because they are on different datasets, and also emphasize that VeriLoC's primary goal of LoC-level predictions is different from CircuitFusion's.
+
+# 4.4 Discussion
+
+We now comment on some interesting properties of VeriLoC, avenues for future improvement, and alternate baselines.
+
+Runtime Comparisons. We compare the runtime efficiency of VeriLoC over Synopsys RTL Architect [13] (our synthesis and PnR tool)—using CL-Verilog-13B as its base, VeriLoC achieves  $14 \times$  average speedup and a median speedup of  $61 \times$ . To explore the potential for even greater runtime improvement, we trained a 7B CL-Verilog model using the training procedure described in [14]. The 7B model has a  $22 \times$  on average and  $113 \times$  median speedup with a modest tradeoff in F1 score of 0.93 for timing and 0.84 for congestion. Thus, while the focus of VeriLoC is on accurate line-level PPA prediction, smaller models like the 7B variant or Mixture-of-Experts (MoE) based architectures can be adopted to further prioritize inference speed without significant degradation in predictive accuracy.
+
+Impact of Verilog Code Length on Accuracy. To assess the effect of Verilog file length on model performance, we conducted a stratified analysis by splitting test samples based on the number of lines per file. We observed a negligible degradation in performance for longer inputs. Specifically, the congestion F1-score dropped slightly from 0.862 for files with fewer than 2000 lines to 0.855 for files with more than 2000 lines. Similarly, the timing F1-score decreased from 0.953 to 0.946, and the Mean Absolute Percentage Error (MAPE) increased from  $5\%$  to  $6.5\%$ . These results indicate that the model exhibits strong robustness to variations in input length, with only minor performance degradation observed on longer files.
+
+Can Black-Box LLMs Predict Hardware Design Quality? As we have noted before, much of the work in the software community on line-level bug detection uses prompt engineering with pre-trained models. These strategies have been successful because of the abundance of training data in the software world. However, hardware data is scarce, especially so for complex concepts like routing congestion. Thus, we hypothesize that prompting strategies are unlikely to work in the hardware context for line-level detection of congestion and timing issues. To test this hypothesis, we picked 10 Verilog files from our test dataset and asked ChatGPT-4o [57] to identify lines of code that would cause timing or congestion issues. ChatGPT was unable to identify lines responsible for congestion in any of our trials. For timing, ChatGPT identified a line of code correctly in one instance, but also had a large number of false positives. Examples of chat responses are given in Appendix D (Fig. 6-9).
+
+# 5 Conclusion
+
+This paper presents a novel LLM-driven framework, VeriLoC, which provides real-time feedback to hardware designers for the impact of their RTL code on crucial design-quality metrics, timing and congestion. By leveraging embedding from LLMs, VeriLoC bridges the gap between RTL coding and downstream performance evaluation, enabling designers to make informed decisions early in the design process. Results demonstrate VeriLoC's capability in predicting design metrics at RTL stage, both for individual lines of code and at the module level.
+
+Limitations. Despite its strong performance, VeriLoC has several important limitations. First, it currently operates on leaf modules drawn from the Open-ABCD benchmark and has not been adapted to large-scale industrial designs with inter-module dependencies. Extensive validation on diverse, industry-grade RTL corpora will establish robustness and generalization, but is hindered by open access to such data.
+
+# References
+
+[1] W. Fang, S. Liu, H. Zhang, and Z. Xie, "Annotating slack directly on your verilog: Fine-grained RTL timing evaluation for early optimization," in Proceedings of the 61st ACM/IEEE Design Automation Conference, ser. DAC '24. New York, NY, USA: Association for Computing Machinery, 2024. [Online]. Available: https://doi.org/10.1145/3649329.3655671  
+[2] D. S. Lopera and W. Ecker, "Applying gnns to timing estimation at rtl : (invited paper)," in 2022 IEEE/ACM International Conference On Computer Aided Design (ICCAD), 2022, pp. 1-8.  
+[3] D. Sánchez Lopera, I. Subedi, and W. Ecker, "Using graph neural networks for timing estimations of rtl intermediate representations," in 2023 ACM/IEEE 5th Workshop on Machine Learning for CAD (MLCAD), 2023, pp. 1-6.  
+[4] W. Fang, Y. Lu, S. Liu, Q. Zhang, C. Xu, L. W. Wills, H. Zhang, and Z. Xie, "Masterrtl: A pre-synthesis ppa estimation framework for anyRTL design," in Proceedings of 2023 IEEE/ACM International Conference on Computer-Aided Design (ICCAD). IEEE, 2023, pp. 1-9.  
+[5] Z. Xie, R. Liang, X. Xu, J. Hu, C.-C. Chang, J. Pan, and Y. Chen, "Preplacement net length and timing estimation by customized graph neural network," IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 41, no. 11, pp. 4667-4680, 2022.  
+[6] A. Hosny, S. Hashemi, M. Shalan, and S. Reda, "Drills: Deep reinforcement learning for logic synthesis," 2020 25th Asia and South Pacific Design Automation Conference (ASP-DAC), pp. 581-586, 2019. [Online]. Available: https://apisemanticscholar.org/CorpusID:207853198  
+[7] K. Zhu, M. Liu, H. Chen, Z. Zhao, and D. Z. Pan, “Exploring logic optimizations with reinforcement learning and graph convolutional network,” in 2020 ACM/IEEE 2nd Workshop on Machine Learning for CAD (MLCAD), 2020, pp. 145–150.  
+[8] Y. V. Peruvemba, S. Rai, K. Ahuja, and A. Kumar, "Rl-guided runtime-constrained heuristic exploration for logic synthesis," in 2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD), 2021, pp. 1-9.  
+[9] A. Basak Chowdhury, B. Tan, R. Carey, T. Jain, R. Karri, and S. Garg, "Bulls-eye: Active few-shot learning guided logic synthesis," IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 42, no. 8, pp. 2580-2590, 2023.  
+[10] A. B. Chowdhury, M. Romanelli, B. Tan, R. Karri, and S. Garg, “Retrieval-guided reinforcement learning for boolean circuit minimization,” arXiv preprint arXiv:2401.12205, 2024.  
+[11] R. Moravej, S. Bodhe, Z. Zhang, D. Chetelat, D. Tsaras, Y. Zhang, H.-L. Zhen, J. Hao, and M. Yuan, “The graph's apprentice: Teaching an llm low level knowledge for circuit quality estimation,” 2025. [Online]. Available: https://arxiv.org/abs/2411.00843  
+[12] W. Fang, S. Liu, J. Wang, and Z. Xie, "Circuitfusion: Multimodal circuit representation learning for agile chip design," in International Conference on Learning Representations (ICLR), 2025. [Online]. Available: https://arxiv.org/abs/2505.02168  
+[13] "Synopsys rtl architect," https://www.synopsys.com/content/dam/synopsys/implementation&signoff/datasheets/rtl-architect-ds.pdf.  
+[14] A. Nakkab, S. Q. Zhang, R. Karri, and S. Garg, "Rome was not built in a single step: Hierarchical prompting for lvm-based chip design," in Proceedings of the 2024 ACM/IEEE International Symposium on Machine Learning for CAD, ser. MLCAD '24. New York, NY, USA: Association for Computing Machinery, 2024. [Online]. Available: https://doi.org/10.1145/3670474.3685964  
+[15] P. BehnamGhader, V. Adlakha, M. Mosbach, D. Bahdanau, N. Chapados, and S. Reddy, "Llm2vec: Large language models are secretly powerful text encoders," 2024. [Online]. Available: https://arxiv.org/abs/2404.05961  
+[16] Y. Luo, T. Zheng, Y. Mu, B. Li, Q. Zhang, Y. Gao, Z. Xu, P. Feng, X. Liu, T. Xiao, and J. Zhu, "Beyond decoder-only: Large language models can be good encoders for machine translation," 2025. [Online]. Available: https://arxiv.org/abs/2503.06594
+
+[17] "Synopsys rtl architect," https://www.synopsys.com/content/dam/synopsys/implementation&signoff/datasheets/rtl-architect-ds.pdf.  
+[18] A. B. Chowdhury, B. Tan, R. Karri, and S. Garg, "Openabc-d: A large-scale dataset for machine learning guided integrated circuit synthesis," 2021.  
+[19] T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” in Proceedings of the 22nd acm SIGkdd international conference on knowledge discovery and data mining, 2016, pp. 785–794.  
+[20] G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y. Liu, “Lightgbm: a highly efficient gradient boosting decision tree,” in Proceedings of the 31st International Conference on Neural Information Processing Systems, ser. NIPS'17. Red Hook, NY, USA: Curran Associates Inc., 2017, p. 3149–3157.  
+[21] A. B. Kahng, J. Lienig, I. L. Markov, and J. Hu, VLSI Physical Design: From Graph Partitioning to Timing Closure. Springer, 2011.  
+[22] M.-C. Kim, J. Hu, D.-J. Lee, and I. L. Markov, “A simplr method for routability-driven placement,” in 2011 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2011, pp. 67–73.  
+[23] X. He, T. Huang, W.-K. Chow, J. Kuang, K.-C. Lam, W. Cai, and E. F. Young, "Ripple 2.0: High quality routability-driven placement via global router integration," in 2013 50th ACM/EDAC/IEEE Design Automation Conference (DAC), 2013, pp. 1-6.  
+[24] W.-H. Liu, C.-K. Koh, and Y.-L. Li, “Optimization of placement solutions for routability,” in 2013 50th ACM/EDAC/IEEE Design Automation Conference (DAC), 2013, pp. 1-9.  
+[25] C.-C. Huang, H.-Y. Lee, B.-Q. Lin, S.-W. Yang, C.-H. Chang, S.-T. Chen, Y.-W. Chang, T.-C. Chen, and I. Bustany, "Ntuple4dr: A detailed-routing-driven placer for mixed-size circuit designs with technology and region constraints," IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 37, no. 3, pp. 669-681, 2018.  
+[26] P. Spindler and F. M. Johannes, "Fast and accurate routing demand estimation for efficient routability-driven placement," in 2007 Design, Automation & Test in Europe Conference & Exhibition, 2007, pp. 1-6.  
+[27] Y. Wei, C. Sze, N. Viswanathan, Z. Li, C. J. Alpert, L. Reddy, A. D. Huber, G. E. Tellez, D. Keller, and S. S. Sapatnekar, “Glare: Global and local wiring aware routability evaluation,” in DAC Design Automation Conference 2012, 2012, pp. 768–773.  
+[28] X. He, T. Huang, L. Xiao, H. Tian, and E. F. Y. Young, "Ripple: A robust and effective routability-driven placer," IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 32, no. 10, pp. 1546-1556, 2013.  
+[29] J.-M. Lin, C.-W. Huang, L.-C. Zane, M.-C. Tsai, C.-L. Lin, and C.-F. Tsai, "Routability-driven global placer target on removing global and local congestion for vlsi designs," in 2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD), 2021, pp. 1-8.  
+[30] Z. Xie, Y.-H. Huang, G.-Q. Fang, H. Ren, S.-Y. Fang, Y. Chen, and J. Hu, "Routenet: Routability prediction for mixed-size designs using convolutional neural network," in 2018 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2018, pp. 1-8.  
+[31] S. Zheng, L. Zou, S. Liu, Y. Lin, B. Yu, and M. Wong, "Mitigating distribution shift for congestion optimization in global placement," in 2023 60th ACM/IEEE Design Automation Conference (DAC), 2023, pp. 1-6.  
+[32] S. Yang, Z. Yang, D. Li, Y. Zhang, Z. Zhang, G. Song, and J. Hao, "Versatile multi-stage graph neural network for circuit representation," in Advances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35. Curran Associates, Inc., 2022, pp. 20313-20324.
+
+[33] R. Kirby, S. Godil, R. Roy, and B. Catanzaro, "Congestionnet: Routing congestion prediction using deep graph neural networks," in 2019 IFIP/IEEE 27th International Conference on Very Large Scale Integration (VLSI-SoC), 2019, pp. 217-222.  
+[34] B. Wang, G. Shen, D. Li, J. Hao, W. Liu, Y. Huang, H. Wu, Y. Lin, G. Chen, and P. A. Heng, “Lhnn: lattice hypergraph neural network for vlsi congestion prediction,” in Proceedings of the 59th ACM/IEEE Design Automation Conference, ser. DAC '22. New York, NY, USA: Association for Computing Machinery, 2022, p. 1297–1302. [Online]. Available: https://doi.org/10.1145/3489517.3530675  
+[35] A. Ghose, V. Zhang, Y. Zhang, D. Li, W. Liu, and M. Coates, "Generalizable cross-graph embedding for gnn-based congestion prediction," in 2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD). IEEE Press, 2021, p. 1-9. [Online]. Available: https://doi.org/10.1109/ICCAD51958.2021.9643446  
+[36] S. Zheng, L. Zou, P. Xu, S. Liu, B. Yu, and M. Wong, "Lay-net: Grafting netlist knowledge on layout-based congestion prediction," in 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD), 2023, pp. 1-9.  
+[37] K. Baek, H. Park, S. Kim, K. Choi, and T. Kim, "Pin accessibility and routing congestion aware DRC hotspot prediction using graph neural network and u-net," in IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2022.  
+[38] Z. Guo, M. Liu, J. Gu, S. Zhang, D. Z. Pan, and Y. Lin, “A timing engine inspired graph neural network model for pre-routing slack prediction,” in Proceedings of the 59th ACM/IEEE Design Automation Conference, ser. DAC '22. New York, NY, USA: Association for Computing Machinery, 2022, p. 1207–1212. [Online]. Available: https://doi.org/10.1145/3489517.3530597  
+[39] P. Cao, G. He, and T. Yang, "Tf-predictor: Transformer-based prerouting path delay prediction framework," IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 42, no. 7, pp. 2227-2237, 2023.  
+[40] M. Chen, J. Tworek, H. Jun, Q. Yuan, H. P. de Oliveira Pinto, J. Kaplan, H. Edwards, Y. Burda, N. Joseph, G. Brockman, A. Ray, R. Puri, G. Krueger, M. Petrov, H. Khlaaf, G. Sastry, P. Mishkin, B. Chan, S. Gray, N. Ryder, M. Pavlov, A. Power, L. Kaiser, M. Bavarian, C. Winter, P. Tillet, F. P. Such, D. Cummings, M. Plappert, F. Chantzis, E. Barnes, A. Herbert-Voss, W. H. Guss, A. Nichol, A. Paino, N. Tezak, J. Tang, I. Babuschkin, S. Balaji, S. Jain, W. Saunders, C. Hesse, A. N. Carr, J. Leike, J. Achiam, V. Misra, E. Morikawa, A. Radford, M. Knight, M. Brundage, M. Murati, K. Mayer, P. Welinder, B. McGrew, D. Amodei, S. McCandlish, I. Sutskever, and W. Zaremba, “Evaluating large language models trained on code,” CoRR, vol. abs/2107.03374, 2021. [Online]. Available: https://arxiv.org/abs/2107.03374  
+[41] Z. Feng, D. Guo, D. Tang, N. Duan, X. Feng, M. Gong, L. Shou, B. Qin, T. Liu, D. Jiang, and M. Zhou, "CodeBERT: A pre-trained model for programming and natural languages," in Findings of the Association for Computational Linguistics: EMNLP 2020, T. Cohn, Y. He, and Y. Liu, Eds. Online: Association for Computational Linguistics, Nov. 2020, pp. 1536-1547. [Online]. Available: https://aclanthology.org/2020-findings-emnlp.139/  
+[42] D. Guo, S. Ren, S. Lu, Z. Feng, D. Tang, S. LIU, L. Zhou, N. Duan, A. Svyatkovskiy, S. Fu, M. Tufano, S. K. Deng, C. Clement, D. Drain, N. Sundaresan, J. Yin, D. Jiang, and M. Zhou, "Graphcode{bert}: Pre-training code representations with data flow," in International Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/forum?id=jLoC4ez43PZ  
+[43] M. Jin, S. Shahriar, M. Tufano, X. Shi, S. Lu, N. Sundaresan, and A. Svyatkovskiy, “Inferfix: End-to-end program repair with llms,” in Proceedings of the 31st ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering, ser. ESEC/FSE 2023. New York, NY, USA: Association for Computing Machinery, 2023, p. 1646–1656. [Online]. Available: https://doi.org/10.1145/3611643.3613892  
+[44] H. Li, Y. Hao, Y. Zhai, and Z. Qian, "Enhancing static analysis for practical bug detection: An llm-integrated approach," Proceedings of the ACM on Programming Languages, vol. 8, no. OOPSLA1, pp. 474-499, 2024.
+
+[45] K. Liu, Y. Liu, Z. Chen, J. M. Zhang, Y. Han, Y. Ma, G. Li, and G. Huang, "Llm-powered test case generation for detecting tricky bugs," arXiv preprint arXiv:2404.10304, 2024.  
+[46] C. Wang, W. Zhang, Z. Su, X. Xu, and X. Zhang, “Sanitizing large language models in bug detection with data-flow,” in Findings of the Association for Computational Linguistics: EMNLP 2024, Y. Al-Onaizan, M. Bansal, and Y.-N. Chen, Eds. Miami, Florida, USA: Association for Computational Linguistics, Nov. 2024, pp. 3790–3805. [Online]. Available: https://aclanthology.org/2024-findings-emnlp.217/  
+[47] S. Thakur, B. Ahmad, H. Pearce, B. Tan, B. Dolan-Gavitt, R. Karri, and S. Garg, "Verigen: A large language model for verilog code generation," ACM Trans. Des. Autom. Electron. Syst., vol. 29, no. 3, Apr. 2024. [Online]. Available: https://doi.org/10.1145/3643681  
+[48] S. Liu, W. Fang, Y. Lu, Q. Zhang, H. Zhang, and Z. Xie, "Rtlcoder: Outperforming gpt-3.5 in design rtl generation with our open-source dataset and lightweight solution," in 2024 IEEE LLM Aided Design Workshop (LAD). IEEE, 2024, pp. 1-5.  
+[49] M. Liu, N. Pinckney, B. Khailany, and H. Ren, "Veriloggeval: Evaluating large language models for verilog code generation," 2023. [Online]. Available: https://arxiv.org/abs/2309.07544  
+[50] Y. Lu, S. Liu, Q. Zhang, and Z. Xie, "Rtllm: An open-source benchmark for designrtl generation with large language model," in Proceedings of the 29th Asia and South Pacific Design Automation Conference, ser. ASPDAC '24. IEEE Press, 2024, p. 722-727. [Online]. Available: https://doi.org/10.1109/ASP-DAC58780.2024.10473904  
+[51] Y. Zhao, D. Huang, C. Li, P. Jin, M. Song, Y. Xu, Z. Nan, M. Gao, T. Ma, L. Qi, Y. Pan, Z. Zhang, R. Zhang, X. Zhang, Z. Du, Q. Guo, and X. Hu, "Codev: Empowering llms withhdl generation through multi-level summarization," 2025. [Online]. Available: https://arxiv.org/abs/2407.10424  
+[52] Z. Pei, H. Zhen, M. Yuan, Y. Huang, and B. Yu, “BetterV: Controlled verilog generation with discriminative guidance,” in Proceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, R. Salakhutdinov, Z. Kolter, K. Heller, A. Weller, N. Oliver, J. Scarlett, and F. Berkenkamp, Eds., vol. 235. PMLR, 21-27 Jul 2024, pp. 40 145-40 153. [Online]. Available: https://proceedings.mlr.press/v235/pei24e.html  
+[53] X. Yao, Y. Wang, X. Li, Y. Lian, R. Chen, L. Chen, M. Yuan, H. Xu, and B. Yu, “Rtlewriter: Methodologies for large models aided rtl code optimization,” 2024. [Online]. Available: https://arxiv.org/abs/2409.11414  
+[54] Y.-D. Tsai, M. Liu, and H. Ren, "Rtlfixer: Automatically fixingRTL syntax errors with large language models," 2024. [Online]. Available: https://arxiv.org/abs/2311.16543  
+[55] G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y. Liu, "Lightgbm: A highly efficient gradient boosting decision tree," Advances in neural information processing systems, vol. 30, 2017.  
+[56] D. C. McElfresh, S. Khandagale, J. Valverde, V. P. C, G. Ramakrishnan, M. Goldblum, and C. White, "When do neural nets outperform boosted trees on tabular data?" in Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. [Online]. Available: https://openreview.net/forum?id=CjVdXey4zT  
+[57] OpenAI, "GPT-4," Mar. 2023. [Online]. Available: https://openai.com/research/gpt-4  
+[58] R. I. Bahar, A. K. Jones, S. Katkoori, P. H. Madden, D. Marculescu, and I. L. Markov, "Workshops on extreme scale design automation (esda) challenges and opportunities for 2025 and beyond," 2020. [Online]. Available: https://arxiv.org/abs/2005.01588  
+[59] X. Wang, Y. Jiang, N. Bach, T. Wang, Z. Huang, F. Huang, and K. Tu, "Automated concatenation of embeddings for structured prediction," in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), C. Zong, F. Xia, W. Li, and R. Navigli, Eds. Online: Association for Computational Linguistics, Aug. 2021, pp. 2643-2660. [Online]. Available: https://aclanthology.org/2021.acl-long.206/
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes]
+
+Justification: The central claims and motivation of the paper are articulated in the abstract and reiterated in the concluding paragraph of the introduction.
+
+Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: The limitations of this work are addressed in the final paragraph of the conclusion section.
+
+Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory assumptions and proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [NA]
+
+Justification: [NA]
+
+Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.  
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental result reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+Answer: [Yes]
+
+Justification: Comprehensive details of the experimental setup, including hyperparameters and dataset specifications, are provided in the results section. Additionally, the code has been made publicly available through an open-source release.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.  
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example
+
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+Answer: [Yes]
+
+Justification: The code required to reproduce the results has been open-sourced and is accessible at: https://github.com/ML4EDA/VeriLoC.git.
+
+Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.  
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental setting/details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+Answer: [Yes]
+
+Justification: All training and testing details—including data splits, hyperparameters, their selection criteria, optimizer type, and related configurations—are thoroughly discussed in the 'Experimental Setting' subsection.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment statistical significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+Answer: [No]
+
+Justification: Due to cost considerations we don't do multiple bootstrap runs.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.
+
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).  
+- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).  
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.  
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments compute resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [Yes]
+
+Justification: The necessary information regarding computational resources is provided in the Results section. This includes details on the type of compute infrastructure used, memory specifications, and execution time for the experiments, ensuring that the experiments can be reproduced reliably.
+
+# Guidelines:
+
+The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code of ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes]
+
+Justification: The research presented in the paper adheres to the NeurIPS Code of Ethics in all respects. It ensures transparency, reproducibility, and responsible use of data and models, with no ethical concerns related to data collection, experimental design, or potential downstream harm.
+
+Guidelines:
+
+- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [NA]
+
+Justification: The work is for Quality of Results prediction for chip design
+
+# Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.  
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.  
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.  
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.  
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [NA]
+
+Justification: No such risks arise for this work.
+
+# Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: See Section 4.1. For the Synopsys Tool, we have a valid license to use.
+
+# Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URL.
+
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.  
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.  
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.  
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer: [Yes]
+
+Justification: See https://github.com/ML4EDA/VeriLoC.git
+
+Guidelines:
+
+- The answer NA means that the paper does not release new assets.  
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and research with human subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [NA]
+
+Justification: [NA]
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA]
+
+Justification: [NA]
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.
+
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.  
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.
+
+# 16. Declaration of LLM usage
+
+Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required.
+
+Answer: [Yes]
+
+Justification: We use a domain-adapted LLM fine-tuned on Verilog to generate embedding, see Section 3.
+
+Guidelines:
+
+- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.  
+- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described.
+
+# A IC Design
+
+The design of modern ICs with billions of transistors and ever-more complex technology nodes is highly demanding. Thus, sophisticated tooling is essential in this domain known as electronic design automation (EDA). EDA tools are build upon strong notions of abstractions and hierarchical procedures. Next, we outline the industry-standard approach for these procedures. More details relevant for this work are also discussed in subsections in here.
+
+System Specification, Architectural Design. Objectives and requirements for functionality, performance, and physical implementation are formulated. Modeling languages like SystemC can be used for a formalized approach.
+
+Behavioral and Logic Design. The specification and architecture description are transformed into a behavioral model which describes inputs, outputs, timing behavior, etc. for the whole system. Toward that end, specific hardware description languages (HDLs) like Verilog are utilized. The abstraction level for this process is also often referred to as register-transfer level (RTL). To limit design time and efforts, third-party components, so-called IP modules, can be integrated at this stage.
+
+Logic Synthesis. The behavioral model is transformed into a low-level circuit description, the gate-level netlist (GLN). This step requires a technology library for mapping from a generic circuit to the technology-specific circuit that is to be manufactured in the end.
+
+Physical Design. The GLN is transformed into an actual physical layout of gates, memories, interconnects, etc. Given the high complexity of this stage, it is typically further divided into the following tasks: partitioning and/or floorplanning, power and ground delivery, placement, clock delivery, routing, and timing closure.
+
+Verification and Signoff. The physical layout must be verified against various design and manufacturing rules, to ensure correct functionality and electrical behaviour. Once all rules are met, the design can be signed-off and taped-out, i.e., send out for fabrication, packaging, and testing.
+
+Practical Challenge. Despite the general success of this compartmentalized approach, a key challenge remains: going through the full EDA stack end-to-end takes considerable time and efforts, in the range of months even for large teams. This challenge is known as the "productivity gap" and is expected to stay, especially for ever-more advanced technology nodes [58]. In an effort for best design quality, engineers often need to reiterate many times over key processes like placement and routing. Due to the very nature of the hierarchical tooling, the individual processes are often lacking detailed insights from prior stages and, more concerning, reasonable estimates for their impact on processes further down in the pipeline. In other words, there is a strong need to integrate well-informed quality assessment into early stages of the EDA pipeline—this reiterates the main motivation of this work at hand.
+
+# B Encoder-Decoder Architecture for Dimensionality Reduction
+
+This model is designed to effectively compress high-dimensional embeddings into a compact latent space while preserving their semantic content. Below, we analyze its structure and key features:
+
+# B.1 Network Architecture
+
+The encoder-decoder network consists of fully connected layers with batch normalization, dropout, and non-linear activation functions. The detailed architecture is shown in Table 6.
+
+# B.2 Regularization Techniques
+
+To ensure robust learning and prevent overfitting, the following techniques are incorporated:
+
+- Dropout: Applied with a  $30\%$  probability at multiple layers to prevent co-adaptation of neurons.  
+- Batch Normalization: Normalizes activations at each layer to stabilize training and improve convergence.  
+- LeakyReLU: Chosen over standard ReLU to avoid dead neurons and ensure a small gradient for negative values.
+
+Table 6: Detailed architecture of the encoder-decoder network.  
+
+<table><tr><td>Layer Type</td><td>Number of Units</td><td>Activation Function</td><td>Additional Features</td></tr><tr><td colspan="4">Encoder</td></tr><tr><td>Fully Connected</td><td>4096</td><td>LeakyReLU (α = 0.01)</td><td>BatchNorm, Dropout (30%)</td></tr><tr><td>Fully Connected</td><td>1024</td><td>LeakyReLU (α = 0.01)</td><td>BatchNorm, Dropout (30%)</td></tr><tr><td>Fully Connected</td><td>Latent Dim (d)</td><td>LeakyReLU (α = 0.01)</td><td>BatchNorm</td></tr><tr><td colspan="4">Decoder</td></tr><tr><td>Fully Connected</td><td>1024</td><td>LeakyReLU (α = 0.01)</td><td>BatchNorm, Dropout (30%)</td></tr><tr><td>Fully Connected</td><td>4096</td><td>LeakyReLU (α = 0.01)</td><td>BatchNorm, Dropout (30%)</td></tr><tr><td>Fully Connected</td><td>Input Dim</td><td>-</td><td>-</td></tr></table>
+
+# B.3 Optimization Strategy
+
+The model is trained using the AdamW optimizer, which combines adaptive learning rates with weight decay regularization. Additional details include:
+
+- Learning Rate: Set to a small value  $(10^{-4})$  to ensure gradual convergence.  
+- Reconstruction Loss: The mean squared error (MSE) between the input embeddings and their reconstructed versions is minimized.  
+- Weight Initialization: All linear layers are initialized using Xavier Uniform Initialization to ensure balanced gradients at the start of training.
+
+# B.4 Training Procedure
+
+The model is trained over 200 epochs using mini-batch gradient descent, with the following considerations:
+
+- Batch Size: A batch size of 128 is chosen to balance memory efficiency and gradient stability.  
+- Orthogonality Regularization: Although not implemented in this iteration, orthogonality constraints on the encoder weights can further enhance disentanglement in the latent space.  
+- Validation: Validation loss on a separate test dataset is monitored to ensure the model generalizes to unseen data.
+
+# B.5 Usage in Downstream Tasks
+
+The trained encoder produces latent embeddings that serve as input features for classification and regression tasks. These embeddings are compact, noise-robust, and retain essential semantic information from the original input.
+
+# B.6 Justification for Using the Last Hidden Layer in VeriLoC
+
+Although VeriLoC focuses on line-level quality prediction, we first conducted a complementary ablation study on module-level embeddings to evaluate the predictive quality of hidden states extracted from different layers of the CL-Verilog model. As shown in Table 7, we observed that embeddings derived from the last decoder layer consistently achieved the highest accuracy across all quality-of-result (QoR) metrics—yielding the best  $R^2$  scores and the lowest MAPE for both timing and congestion prediction. These results suggest that the final layer best captures high-level semantic and structural information necessary for downstream design-quality tasks.
+
+Based on this empirical evidence, VeriLoC exclusively uses the last hidden layer to extract both line-level and module-level embeddings. This design choice ensures that the model benefits from the richest representation available, without introducing ambiguity or requiring additional architectural tuning to select among intermediate layers. Furthermore, the consistently superior performance of the final layer justifies avoiding ensemble or multi-layer fusion strategies, which may add unnecessary complexity without proportional gains.
+
+Table 7: Prediction quality using embeddings derived from hidden states of the model, specifically examining the first (1st, 2nd, 3rd) and last (3rd-Last, 2nd-Last, Last) layers for CL-Verilog.  
+
+<table><tr><td rowspan="2">Hidden Layer</td><td colspan="2">Timing</td><td colspan="2">Congestion</td></tr><tr><td>R²</td><td>MAPE</td><td>R²</td><td>MAPE</td></tr><tr><td>1st</td><td>0.66</td><td>15.46</td><td>0.66</td><td>18.13</td></tr><tr><td>2nd</td><td>0.77</td><td>5.64</td><td>0.59</td><td>23.92</td></tr><tr><td>3rd</td><td>0.75</td><td>6.55</td><td>0.72</td><td>13.62</td></tr><tr><td>3rd-Last</td><td>0.85</td><td>5.88</td><td>0.53</td><td>22.34</td></tr><tr><td>2nd-Last</td><td>0.85</td><td>9.32</td><td>0.59</td><td>17.56</td></tr><tr><td>Last</td><td>0.89</td><td>5.57</td><td>0.74</td><td>11.66</td></tr></table>
+
+Table 8: Impact of Hidden Dimension and Classifier on Line-Level Detection Performance for Congestion and Timing. Precision (P), Recall (R).  
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Hidden Dim.</td><td colspan="2">Congestion</td><td colspan="2">Timing</td></tr><tr><td>P</td><td>R</td><td>P</td><td>R</td></tr><tr><td rowspan="4">XGB</td><td>32</td><td>0.76</td><td>0.64</td><td>0.77</td><td>0.71</td></tr><tr><td>64</td><td>0.88</td><td>0.71</td><td>0.86</td><td>0.83</td></tr><tr><td>128</td><td>0.94</td><td>0.78</td><td>0.94</td><td>0.94</td></tr><tr><td>256</td><td>0.94</td><td>0.78</td><td>0.94</td><td>0.94</td></tr><tr><td rowspan="4">LGBM</td><td>32</td><td>0.77</td><td>0.64</td><td>0.79</td><td>0.71</td></tr><tr><td>64</td><td>0.88</td><td>0.72</td><td>0.86</td><td>0.82</td></tr><tr><td>128</td><td>0.94</td><td>0.79</td><td>0.96</td><td>0.94</td></tr><tr><td>256</td><td>0.94</td><td>0.79</td><td>0.96</td><td>0.94</td></tr></table>
+
+# C Additional Ablation Studies
+
+# C.1 Role of Encoder and Choice of Hidden Dimensions
+
+As noted in Section 3.3, we use an encoder-decoder architecture to reduce the dimensionality of raw line- and module-level embeddings before classification/regression. Without the encoder, we obtained F1-scores of less than 0.5. To determine the optimal hidden dimension for the encoder, we study different embedding dimensions across XGBoost and LightGBM. The results are summarized in Table 8.
+
+We find that increasing the hidden dimension significantly improves both congestion and timing detection performance up to a dimension of 128. Beyond 128, however, there is no observable gain in performance, suggesting that further increasing the embedding size is not warranted and that 128 is the optimal hidden dimension, balancing performance and computational efficiency.
+
+# C.2 Impact of Comments in RTL
+
+As opposed to intermediate representations like AIGs, RTL code also contains comments that we hypothesized to be useful in design quality predictions. Here, we evaluate the role of comments in VeriLoC's accuracy. To this end, we conducted a study where we removed comments from both the train and test datasets and train VeriLoC classifiers on the resulting code. Table 9 compares line-level detection performance with (w) vs without (w/o) comments in the embeddings. The results show a modest but clear improvement in congestion and timing precision, recall, and F1-scores when comments are included. The benefits are starkest for FNN, where F1-scores increase from 0.71 to 0.77 for timing and 0.70 to 0.76 for congestion. XGBoost and LightGBM also benefit from comments, especially for congestion prediction with F1-scores increasing from 0.92 to 0.95 in the best case.
+
+Table 9: Effect of Comments (C) in Module Embedding Generation on Line Detection Performance. Precision (P), Recall (R).  
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Emb.</td><td colspan="3">Congestion</td><td colspan="3">Timing</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td rowspan="2">FNN</td><td>w/o C</td><td>0.80</td><td>0.64</td><td>0.71</td><td>0.63</td><td>0.78</td><td>0.70</td></tr><tr><td>w/ C</td><td>0.86</td><td>0.7</td><td>0.77</td><td>0.67</td><td>0.88</td><td>0.76</td></tr><tr><td rowspan="2">XGB</td><td>w/o C</td><td>0.93</td><td>0.76</td><td>0.84</td><td>0.91</td><td>0.92</td><td>0.91</td></tr><tr><td>w/ C</td><td>0.94</td><td>0.78</td><td>0.85</td><td>0.94</td><td>0.94</td><td>0.94</td></tr><tr><td rowspan="2">LGBM</td><td>w/o C</td><td>0.95</td><td>0.77</td><td>0.85</td><td>0.93</td><td>0.91</td><td>0.92</td></tr><tr><td>w/ C</td><td>0.94</td><td>0.79</td><td>0.86</td><td>0.96</td><td>0.94</td><td>0.95</td></tr></table>
+
+![](images/02829c786a5eb23d3fbab7aadcee2907bf44bf5abc7a492867cd213569821fe5.jpg)  
+Fig. 5 further supports these findings through saliency-based visualizations of batched vs individual line embeddings. Specifically, the saliency heatmaps reveal that, w/o comments, key structural elements like always @(posedge clk) receive a disproportionate amount of attention, whereas w/ comments, the model distributes attention more effectively across relevant components.  
+Figure 5: Saliency map comparing attention of VeriLoC model (bottom) vs. batched line embeddings (top).
+
+# C.3 Alternate Approaches
+
+Recall that including neighboring lines of code can significantly improve performance, especially for congestion prediction. An alternate method for including local context, however, would be to obtain a single embedding for a batch of consecutive lines  $B = \{l_{i - p},\ldots ,l_i,\ldots ,l_{i + p}\}$  by passing
+
+$B$  through CL-Verilog as a single input. The model computes batched embeddings  $z_{\mathrm{batch}}(B)$ , which natively capture inter-line dependencies via the LLM's attention mechanism.
+
+Although appealing, the batched approach results in lower F1-scores, achieving at best an F1-score of 0.8 for congestion prediction; recall that VeriLoC achieves an F1-score of 0.86. Fig. 5 depicts the attention patterns of the model for congestion detection where the line always @(posedge clk) is critical. For the batched approach, the model's attention is dispersed across multiple unrelated tokens. Conversely, line-wise embeddings allow the model to prioritize the relevant tokens more effectively, as shown by the darker-red shades. This also aligns with the findings in [59], which emphasize the importance of modular embedding strategies for structured tasks.
+
+# D Case Study Using GPT-4o and VeriLoC
+
+![](images/bd4be66e33cfb4522786860e71e213cb8861f66735c65db6b925a0edce9118af.jpg)  
+Figure 6: The lines of code highlighted by GPT-4o and VeriLoC for congestion metrics on aes_rcon design.
+
+![](images/0141cd760899247335195e6a92a263dcc2e8cf18249d0d5a6678925adb0abdd3.jpg)
+
+As a part of the case study, to understand the capability of generic LLMs to analyze RTL code, we took GPT-4o [57] and prompted it to report the line numbers in the code responsible for timing and congestion, respectively. We also showcase VeriLoC performance for the same codes. GPT-4o was unable to detect the correct line numbers, i.e., showed false positives, whereas VeriLoC has shown significant results when compared to the ground truth obtained using the EDA tool.
+
+Fig. 6-9 show such examples with highlighted text being the line number reported by GPT-4o and VeriLoC, respectively. Fig. 6 refers to the aes_rcon design, where GPT-4o predicts every line as critical for congestion issues, whereas VeriLoC selectively highlights the correct lines related to congestion. Fig. 7 again shows an inability of GPT-4o, now for the task of timing prediction.
+
+![](images/726a96d7c0af5a06b91f04f6595ee15e5a3891f88fe9d24fa5521bab185f4dd3.jpg)  
+Figure 7: The lines of code highlighted by GPT-4o and VeriLoC for timing metrics on ibex_csr design.
+
+![](images/39c3dfad5a445190d1ed6a38819cfebd556eda5d45e87992efcaa2e2791a07e3.jpg)
+
+Although it has highlighted the lines related to timing impact, namely the first always block, it also reports the next blocks as potential candidates. In Fig. 8, GPT-4o picks registers definitions as a potential reason for congestion and skips the actual critical line containing the additions operation. Finally, in Fig. 9 GPT-4o was unable to pick the correct always block responsible for timing issues.
+
+![](images/75e99ddcf603b828242c7096f6ea854496a8c8dff073239fa4ce9b04e8ae4888.jpg)  
+Figure 8: The lines of code highlighted by GPT-4o and VeriLoC for congestion metrics on dct_mac design.
+
+![](images/63a374d9e2397599bf26a845443d644633c95446e87be75398b07352d673cc84.jpg)  
+Figure 9: The lines of code highlighted by GPT-4o and VeriLoC for timing metrics on zigzag design.

2025/VeriLoC_ Line-of-Code Level Prediction of Hardware Design Quality from Verilog Code/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdad59f345fc46fa6084d0e54b48a06b29608a87d9346df1330c3f60829c124a
+size 1142795

2025/VeriThinker_ Learning to Verify Makes Reasoning Model Efficient/a0289d16-03a4-4105-9216-1417938420cf_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d93a7ada36afd388c10d52b8ac26f6953462c9c741402cd616a1967e1a72049f
+size 4875578

2025/VeriThinker_ Learning to Verify Makes Reasoning Model Efficient/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:400934df199d7ea1ffca7f37e6dbe07911b3b0cae91d0b5b60f1101c0fdec551
+size 626641

2025/Versatile Transferable Unlearnable Example Generator/0c230d75-f6ec-478b-87fa-3d67ddcc818b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:543bce14fa2bdc8e068d84e5a4d9ecdbd7bfc1e752b82ba99779f48e372851f6
+size 1463390

2025/Versatile Transferable Unlearnable Example Generator/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72fd47c2cbdc3b386cdfd9195036bfeb4a148d9fa96145b811d829b15925f65e
+size 1128991

2025/Vertical Federated Feature Screening/6990da2f-c640-414b-8dc1-c12a32a82387_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba8a33e909cfdc479177453cb12751edf46bd8036683a65ec319a17174e56ef2
+size 5099166

2025/Vertical Federated Feature Screening/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:674ef62bbfa819a6c8b13874b9b4a50feabe5e1f88ab989d1470099fe021a327
+size 1149311

2025/Vgent_ Graph-based Retrieval-Reasoning-Augmented Generation For Long Video Understanding/c5a4b368-9f46-42fb-9ad9-4fc0cdb91820_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13bfc9c6dab86d2f80fdee0525253a63f4d318608969b141f43cc34608351492
+size 2538906

2025/Vgent_ Graph-based Retrieval-Reasoning-Augmented Generation For Long Video Understanding/full.md

@@ -0,0 +1,714 @@
+# Vangent: Graph-based Retrieval-Reasoning-Augmented Generation For Long Video Understanding
+
+Xiaoqian Shen<sup>1</sup>, Wenxuan Zhang<sup>1</sup>, Jun Chen<sup>1,2</sup>, Mohamed Elhoseiny<sup>1</sup>  
+<sup>1</sup>King Abdullah University of Science and Technology, <sup>2</sup>Meta AI {xiaoqian.shen,wenxuan.zhang,jun.chen,mohamed.elhoseiny}@kaust.edu.sa
+
+# Abstract
+
+Understanding and reasoning over long videos pose significant challenges for large video language models (LVLMs) due to the difficulty in processing intensive video tokens beyond context window and retaining long-term sequential information. Retrieval-Augmented Generation (RAG) has demonstrated effectiveness in processing long context for Large Language Models (LLMs); however, applying RAG to long video faces challenges such as disrupted temporal dependencies and inclusion of irrelevant information that can hinder accurate reasoning. To address these limitations, we propose Vgent, a novel graph-based retrieval-reasoning-augmented generation framework to enhance LVLMs for long video understanding. Our approach introduces two key innovations: (i) It represents videos by structured graphs with semantic relationships across video clips preserved to improve retrieval effectiveness. (ii) It introduces an intermediate reasoning step to mitigate the reasoning limitation of LVLMs, which leverages structured verification to reduce retrieval noise and facilitate the explicit aggregation of relevant information across clips, resulting in more accurate and context-aware responses. We comprehensively evaluate our framework with various open-source LVLMs on three long-video understanding benchmarks. Our approach yielded an overall performance improvement of  $3.0\% \sim 5.4\%$  over base models on MLVU, and outperformed state-of-the-art video RAG methods by  $8.6\%$ . Our code is publicly available at https://xiaoqian-shen.github.io/Vgent.
+
+# 1 Introduction
+
+Multi-Modal Large Language Models (MLLMs) [6, 22, 30, 39, 40, 58] have demonstrated remarkable visual understanding and reasoning capabilities, paving the way for advancements in video tasks. Recently, numerous studies have showcased impressive progress in building Large Video Language Models (LVLMs) for video understanding [9, 7, 24, 35, 55]. Moreover, long-video understanding is particularly crucial for applications in web content, life-logging, and streaming media, where intricate narratives and evolving contexts span extended durations.
+
+However, processing and reasoning over long-context videos remain a formidable challenge for existing LVLMs, as representing video frames requires an extensive number of tokens—for example, a 30-minute video can exceed 200K tokens [4, 22], beyond most models' context limits. To handle longer videos, existing methods resort to sparse frame sampling [55, 22] or token compression [43], but these approaches inevitably lead to visual information loss, weakening fine-grained temporal understanding and coherent reasoning.
+
+Recent studies [1, 2, 12, 34, 52, 53] utilize Retrieval-Augmented Generation (RAG) [21] to enhance long-form video understanding by retrieving relevant information. However, they encounter certain limitations. First, some works segment lengthy videos into shorter clips, treating each as an individual document for retrieval [2], which disrupts the continuity of entities and temporal dependencies,
+
+![](images/28fa7d259b9e1f2d3119c378ad8f3eec98570555e1c43bb4f993e86ac2d97d00.jpg)  
+Figure 1: Overview of our graph-based retrieval-reasoning-augmented generation framework. Each video clip is represented as a node within a graph, interconnected through shared entities. This graph representation enables effective retrieval of relevant clips based on node connections, followed by an intermediate reasoning step to refine retrievals and aggregate over multimodal context for accurate generation.
+
+![](images/0f492b406289cf43d60befe11783aa528b26ad3ba08e8c99c1d5914d9385130d.jpg)
+
+leading to retrieval inaccuracies. Second, some methods [47, 34] rely on proprietary LLMs like GPT-4 [37] for multi-turn interactions, planning, and reasoning, making them costly and less flexible. Lastly, several approaches [47, 33] extract information from sparse key frames, neglecting temporal coherence in long videos.
+
+Graph representation for enhanced retrieval: To address the limitation of existing RAG methods, we propose a structured graph-based representation, where video clips are modeled as nodes interconnected by recurring subjects or scenes. This graph representation not only enables effectively retrieving nodes associated with specific entities, but also facilitates capturing temporal dependencies spanning over lengthy videos. Another advantage is that the graph construction is performed offline and is query-independent. Once the graph is built, it can be reused for multiple questions on the same video, allowing retrieval to operate directly on the graph without reprocessing the video. However, feeding all retrieved clips into LLMs can cause information overload, where key details are diluted by irrelevant content [14]. In videos, the issue is further amplified, as each frame consumes hundreds of tokens, with irrelevant information overshadowing the critical content.
+
+Structured post-retrieval reasoning: To address the aforementioned issue and fully harness the benefits from our GraphRAG, we introduce the structured reasoning step in the post-retrieval stage. Instead of generating answers directly from the retrieved clips, it decomposes the question and systematically verifies the relevance of each clip. As shown in Figure 1, this process refines the retrieved set by identifying clips that mention critical elements—such as the plate, sink, cabinet, and bread—and then aggregates information across them for temporal reasoning (e.g., the plate moving from the sink to the cabinet). This process mitigates noise, facilitates information aggregation across refined clips, and thus creates a more reliable pathway for producing accurate responses.
+
+We evaluate our framework upon seven different LVLMs with sizes ranging from 2B to 7B across three long-video benchmarks: MLVU [57], VideoMME [13] and LongVideoBench [50]. Experimental results demonstrate that our framework consistently improves the performance of existing LVLMs by  $3.0\% - 5.4\%$ . We further show that our framework surpasses existing RAG-based video understanding works by  $8.6\%$ .
+
+Contribution. Our contribution is summarized as follows:
+
+- We developed a novel graph-based RAG framework for long-video understanding, where video clips are represented as nodes within a graph, interconnected through shared entities, thereby preserving semantic relationships and temporal dependencies across clips, facilitating more effective retrieval.  
+- We propose structured reasoning to tackle the limited reasoning ability of LVLMs, which can be distracted by hard negative retrieved samples. Our approach introduces an intermediate reasoning step for retrieval verification and aggregates information across verified clips to enhance generation accuracy.
+
+- Our graph-based retrieval-reasoning-augmented framework demonstrates  $3.0\% - 5.4\%$  improvements over various LVLMs ranging from 2B to 7B and surpasses existing RAG-based video understanding works by  $8.6\%$  in long-video understanding tasks.
+
+# 2 Related Work
+
+# 2.1 Large Video Language Models
+
+Multimodal large language models (MLLMs) [6, 58, 31, 30, 45] have demonstrated remarkable progress in vision-language tasks. Recent advancements have further extended their capabilities to video understanding tasks [2, 9, 24, 25, 28, 32]. Large Video Language Models (LVLMs) process videos by extracting and encoding frames, and then rearranging them into final video representations. Some approaches [24, 25, 9] leverage the Q-Former module from BLIP-2 [23] to integrate visual and textual features, while others [2, 28, 32] directly concatenate frame features. However, these models struggle with processing hour-long videos in a single pass as the number of video tokens exceeds their training context size. To address these limitations, most existing works train on sparsely sampled frames no matter how long the video is [24, 2, 9, 55, 22], while others try to handle long videos by token pooling [35, 27, 44], token compression [43], or memory aggregation [19]. However, they struggle to effectively capture and reason about temporal dependencies spanning hour-long videos.
+
+# 2.2 Agent-based Video Understanding
+
+A dominant trend in long-context video question-answering involves equipping Large Language Models (LLMs) with tools that heavily rely on proprietary models to process queries and handle video clips. MM-VID [29] uses a video-to-script generation with GPT-4V [37] to transcribe multimodal elements into a long textual script. VideoAgent [47] integrates diverse foundation models through a unified memory architecture. DrVideo [34], VideoTree [49], VideoAgent [12] and OmAgent [53] dynamically invoke tools to enhance query processing and accuracy. Such methods consequently suffer from high operational costs and a critical reliance on external, closed-source systems, limiting their adaptability. In contrast, our work targets the development of a self-contained pipeline designed for flexible deployment with open-source LVLMs.
+
+# 2.3 Video Retrieval-Augmented Generation
+
+Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by retrieving relevant information to improve long context memory, factual accuracy, and reduce hallucinations. The process involves three stages: (i) indexing, which organizes raw data into a knowledge base; (ii) retrieval, which searches for relevant information based on user queries; and (iii) generation, where the model takes the retrieved context to generate the final response. Recent advancements of RAG in LLM mainly follow two directions, i.e. chunk-based methods [15, 5] and graph based methods [11, 17, 26, 18], both of which have been applied in video understanding tasks. Goldfish [3] chunks long videos into shorter clips, processes each clip independently, and retrieves the most relevant clip in response to user queries. Wang et al. [48] applied graph structures for action recognition in short clips and Hussein et al. [20], Luo et al. [33] employ scene graphs for video understanding. However, constructing graphs for long videos and effectively retrieving information from the noisy and complex graph remains a challenge. Only recently, a concurrent work [42] constructs graphs for long-context video understanding, but they heavily relies on external proprietary LLM for graph construction, while graph-based video RAG with open-sourced LVLM itself remains unexplored, which our work aims to address.
+
+# 3 Method
+
+We introduce a novel, training-free framework, Vgent, for long-context video understanding. Unlike conventional Retrieval-Augmented Generation (RAG), our pipeline proposes a graph-based retrieval-reasoning-augmented generation paradigm, specifically designed to address complex video scenarios with improved contextual comprehension and structured reasoning. As illustrated in Figure 2, our proposed pipeline contains four stages: (1) Offline video graph construction (Section 3.1): Builds a video graph offline by extracting knowledge from long videos. (2) Graph-based
+
+![](images/81732fc7a798cbf40b95f7b0cf8b7a38b3dbdff5526bf76d4ceeb37f79230396.jpg)  
+Figure 2: Pipeline of Vgent, a novel framework for long-context video understanding in the proposed graph-based retrieval-reasoning-augmented generation paradigm. It consists of four key stages: (1) Offline video graph construction (Section 3.1): Builds a video graph by extracting knowledge from long videos. (2) Graph-based retrieval (Section 3.2): Retrieves relevant clips based on keywords extracted from the user query. (3) Structured reasoning (Section 3.3): Refines clips using structured queries and aggregates information. (4) Multimodal augmented generation (Section 3.4): Combines refined clips and reasoning results to generate the final response.
+
+retrieval (Section 3.2): Retrieves relevant video clips from graph based on the user query. (3) Structured Reasoning (Section 3.3): Refines the retrieved clips using structured queries and aggregates information across the filtered clips. (4) Multimodal Augmented Generation (Section 3.4): Combines refined clips and intermediate reasoning results to generate the final response.
+
+# 3.1 Video Graph Construction
+
+To better capture the complex relationships and dependencies in long-context videos, we propose a graph-based representation to store video content and enhance semantic connections. Specifically, given a video  $V$  with  $F$  frames, we first partition it into a sequence of short video clips  $\{V_1, V_2, \ldots, V_{\lceil \frac{F}{K} \rceil}\}$ , where each video clip  $V_i$  consists of  $K$  frames. We then dynamically construct the graph by a series of structured steps, as detailed below.
+
+Visual Entity Extraction. For each video clip, we leverage the LVLM to extract the key semantic entities (i.e., the primary subjects, actions, or scenes) from both the spoken content (subtitles)  $C_i$  and video clip  $V_i$ .
+
+$$
+\left\{\left(e _ {1} ^ {i}, t _ {1} ^ {i}\right), \left(e _ {2} ^ {i}, t _ {2} ^ {i}\right), \dots \right\} \leftarrow \text {L V L M} \left(C _ {i}, V _ {i}\right), \tag {1}
+$$
+
+where the set of entity is denoted as  $E_{i} = \{e_{i}^{1}, e_{i}^{2}, \ldots \}$  and its corresponding description set is denoted as  $T_{i} = \{t_{i}^{1}, t_{i}^{2}, \ldots \}$ . In this step, the LVLM captures subjects, actions, and scene dynamics, seamlessly linking visual entities with spoken content to extract meaningful knowledge. Please refer to Appendix B.1 for illustrative examples.
+
+Graph Construction. Based on extracted information, we construct a video knowledge graph  $\mathcal{G} = (\mathcal{V},\mathcal{E})$ , where  $\mathcal{V}$  denotes the nodes set representing video clips, and edges in  $\mathcal{E}$  represents the connectivity between nodes. Additionally, we define a global set of unique prototype entities  $\mathcal{U} = \{u\in E_i,i = 1,\dots ,\lceil \frac{F}{K}\rceil \}$  that spans across all nodes. As more video clips are processed, we will dynamically add newly extracted unique entity  $u$  to the set or link it to an existing entity. We define  $t^u$  as the description of each entity  $u\in \mathcal{U}$ .
+
+Entity Merging and Node Connection. Since LVLMs process video clips independently, it is essential to identify and unify semantically equivalent entities across clips. Given a newly extracted entity-description pair  $(e_i^j,t_i^j)$  from video clip  $V_{i}$ , we determine whether it belongs to an existing entity in the global entity set  $\mathcal{U}$ . Specifically, we compute the similarity score between the textual descriptions  $t_j^i$  and descriptions of entities in  $\mathcal{U}$  based on their respective text embeddings. If the similarity score  $> \tau$ , the entity  $e_j^i$  is considered semantically equivalent to an existing entity and these two are merged into a single entity representation. Otherwise,  $e_j^i$  is treated as a distinct entity and added to  $\mathcal{U}$ . This process is formulated as follows:
+
+$$
+s ^ {*} = \max  _ {u \in \mathcal {U}} \operatorname {s i m} \left(t _ {j} ^ {i}, t\right), u ^ {*} = \arg \max  _ {u \in \mathcal {U}} \operatorname {s i m} \left(t _ {j} ^ {i}, t ^ {u}\right), e _ {j} ^ {i} \rightarrow \left\{\begin{array}{l l}u ^ {*},&\text {i f} s ^ {*} \geq \tau\\\mathcal {U} \leftarrow \mathcal {U} \cup \left\{e _ {j} ^ {i} \right\},&\text {o t h e r w i s e}\end{array}\right. \tag {2}
+$$
+
+Once entity is merged, we then build edges from the node  $v_{i}$  associated with the video clip  $V_{i}$  to all the nodes that have the same entity  $u^{*}$ , denoted as  $\mathcal{V}(u^{*})$ .
+
+$$
+\mathcal {E} \leftarrow \mathcal {E} \cup \left\{\left(v _ {i}, v\right) \mid v \in \mathcal {V} ^ {\left(u ^ {*}\right)} \right\} \tag {3}
+$$
+
+As new video clips are processed, the graph is dynamically updated such that nodes containing the same entity are connected, which preserves semantic relationships and contextual dependencies. This forms a structured representation that facilitates effective video retrieval in subsequent processing stages.
+
+# 3.2 Graph-based Retrieval
+
+Keywords Extraction. Direct retrieval based on the original query may not provide sufficient context, especially when reasoning across multiple temporal clips is required. To address this, we extract keywords from the query for effective retrieval. Specifically, we prompt the LVLM to identify key semantic elements, denoted as  $\mathcal{K}$ , from the query  $Q$ . The detailed prompt is provided in Appendix B.2.
+
+Graph-based Clip Retrieval. Next, we leverage these extracted keywords for graph-based retrieval. Specifically, for each keyword  $k \in \mathcal{K}$  and each entity  $u \in \mathcal{U}$ , we compute a similarity score  $sim(k, t^u)$  to determine whether the entity matches the keyword. If  $sim(k, t^u) > \theta$ , we include all nodes associated with entity  $u$  as the target retrieval node set  $\mathcal{R}$ :
+
+$$
+\mathcal {R} = \bigcup_ {u \in \mathcal {U}, k \in \mathcal {K}} \left\{v \in \mathcal {V} \mid u \in \mathcal {U} (v), \operatorname {s i m} \left(k, t ^ {u}\right) > \theta \right\} \tag {4}
+$$
+
+After obtaining the retrieval node set  $\mathcal{R}$ , we refine the results by re-ranking the nodes based on the similarity between the query's keywords and the extracted information of each node, including entities, corresponding textual descriptions, and subtitles if available. Finally, we select the Top- $N$  nodes with the highest average similarity scores across all associated information of each video clip.
+
+# 3.3 Structured Reasoning
+
+Feeding all relevant clips directly into LLMs can lead to information overload, diluting the focus on key details with irrelevant content [14]. Our empirical analysis also reveals that in roughly  $40\%$  of failure cases, the correct clip is successfully retrieved, yet the model still generates incorrect responses—even though it can answer correctly when provided with that clip alone. We then introduce structured reasoning in the post-retrieval stage that refines the retrieved clips and aggregates useful information towards final generation.
+
+Structured Query Refinement. We introduce the divide-and-conquer strategy to refine the retrieval through structured query verification. Specifically, we prompt the LVLM to generate structured subqueries, denoted as  $\mathcal{Q}$ , based on the original query  $Q$  and extracted keywords  $\mathcal{K}$ . These subqueries focus on verifying the presence of relevant entities or quantifying their occurrences, whose answers are expected to be binary (yes/no) or numerical value. Please refer to Appendix B.3 for the detailed prompt and Figure 3 for an example of generated subqueries.
+
+After generating the subqueries, we process the Top-  $N$  retrieved video clips using the LVLM, producing either binary (yes/no) or numerical responses for each subquery. As shown in Figure 2, this structured verification systematically assesses the relevance of each clip to the original query, filtering out irrelevant clips that were wrongly retrieved based on semantic embedding similarity. Denoting 1 to yes and 0 to no in binary questions, this refined clip set  $\mathcal{R}'$  can be formulated as:
+
+$$
+\mathcal {R} ^ {\prime} = \left\{v _ {i} \in \mathcal {R} \mid \exists q _ {j} \in \mathcal {Q}, f \left(v _ {i}, q _ {j}\right) > 0 \right\} \tag {5}
+$$
+
+where  $f(v_{i},q_{j})$  denotes the response of retrieved clip  $v_{i}$  to subquery  $q_{j}$ . We keep at most  $r$  clips after refinement. This refinement step ensures that only video clips satisfying the structured queries are retained, effectively eliminating hard negatives from the initial retrieval.
+
+Information Aggregation. As shown in Figure 2, we then let LVLM aggregate and summarize all useful information from structured queries and their corresponding results for each video clip, providing an enriched auxiliary context that enhances the final inference.
+
+# 3.4 Multimodal Augmented Generation.
+
+We incorporate both the intermediate reasoning results and the filtered video clips as multimodal context inputs to the LVLM for the final response. This enriched input allows the model to leverage both structured reasoning and relevant visual information, enabling it to generate a more accurate and contextually grounded final response to the original question.
+
+# 4 Experiments
+
+# 4.1 Experimental Setups
+
+Baselines. We apply our framework Vgent on open-sourced LVLMs including InternVL2.5 [8], Qwen2 [46], Qwen2.5-VL [4], LongVU [43] and LLaVA-Video [56] as base video understanding model. We further compare Vgent against state-of-the-art RAG baselines as follows: NaiveRAG [3], Video-RAG [33], and proprietary LLM-based methods including VideoAgent [47], LLoVi [52], DrVideo [34] and VideoTree [49]. More details can be found in Appendix A.3.
+
+**Benchmarks.** We evaluate the performances of each model across three long-video benchmarks. Video-MME [13] is a widely used benchmark designed to evaluate LVLMs' capability to process detailed, real-world videos. It comprises three subsets categorized by video length, ranging from 11 seconds to 1 hour. MLVU [57] is a long-video understanding benchmark with videos ranging from 3 minutes to 2 hours, with an average length of about 12 minutes. LongVideoBench (LVB) [50] focuses on referred reasoning tasks that require models to analyze long frame sequences. These questions depend on extensive temporal context and cannot be effectively addressed using a single frame or a small set of sparsely sampled frames.
+
+Implementation Details. During the offline video graph construction, we sample videos at 1.0 FPS, segmenting the long video into clips, each containing  $K = 64$  frames. We use the BAAI/bge-large-en-v1.5 [51] embedding for similarity calculation. The entity merging threshold is set to  $\tau = 0.7$ . In the online retrieval stage, we use BAAI/bge-large-en-v1.5 to retrieve the top  $N = 20$  clips based on extracted keywords (maximum to 20 to discard low-relevance, with a similarity threshold  $\theta = 0.5$ ). After structured query refinement, we retain a maximum of  $r = 5$  clips. Thresholds are set as the same for all three benchmarks, with hyper-parameter selection details provided in the supplementary. For MLVU [57], we extract spoken content using openai/whisper-large, while for VideoMME [13] and LongVideoBench [50], we use the provided subtitles from benchmark. All experiments are conducted on A100 80G GPUs.
+
+# 4.2 Main Results
+
+Comparison with LVLMs. In Table 1 and 6, we present the performance of our proposed Vgent framework on the MLVU [57] benchmark, where we consistently observe substantial improvements across all models. Specifically, our framework enhances LongVU [43] by  $5.4\%$  and boosts Qwen2.5VL [4] (7B) by  $3.3\%$ . Notably, when applied to Qwen2.5VL (3B), Vgent achieves
+
+Table 1: Performance comparison with LVLMs. Vgent consistently improves all models on MLVU [57], enhancing LongVU by  $5.4\%$  and Qwen2.5VL (7B) by  $3.3\%$ . Notably, Vgent achieves  $70.4\%$  accuracy on Qwen2.5VL (3B), surpassing its 7B counterpart and improving the base model by  $4.2\%$ . Vgent outperforms base models across all video lengths on VideoMME [13] achieving improvement of  $3.2\%$  overall.  
+
+<table><tr><td rowspan="2">Models</td><td rowspan="2">Size</td><td rowspan="2">MLVU</td><td colspan="2">VideoMME</td><td rowspan="2">LVB</td></tr><tr><td>w/o sub.</td><td>w/ sub.</td></tr><tr><td colspan="6">Proprietary LVLMs</td></tr><tr><td>Gemini 1.5 Pro [16]</td><td>-</td><td>-</td><td>75.0</td><td>81.3</td><td>64.0</td></tr><tr><td>GPT-4o [38]</td><td>-</td><td>64.6</td><td>71.9</td><td>77.2</td><td>66.7</td></tr><tr><td colspan="6">Open-Source LVLMs</td></tr><tr><td>InternVL2.5 [8]</td><td>2B</td><td>56.7</td><td>49.5</td><td>55.2</td><td>52.0</td></tr><tr><td>InternVL2.5 + Vgent (Ours)</td><td>2B</td><td>61.1+4.4</td><td>50.9+1.4</td><td>56.8+1.6</td><td>54.8+2.8</td></tr><tr><td>Qwen2.5-VL [4]</td><td>3B</td><td>66.2</td><td>61.4</td><td>67.6</td><td>54.2</td></tr><tr><td>Qwen2.5-VL + Vgent (Ours)</td><td>3B</td><td>70.4+4.2</td><td>63.0+1.6</td><td>69.6+2.0</td><td>57.8+3.6</td></tr><tr><td>LongVU [54]</td><td>7B</td><td>65.4</td><td>55.2</td><td>60.9</td><td>50.2</td></tr><tr><td>LongVU + Vgent (Ours)</td><td>7B</td><td>70.8+5.4</td><td>57.3+2.1</td><td>63.7+2.8</td><td>52.7+2.5</td></tr><tr><td>Qwen2-VL [46]</td><td>7B</td><td>65.7</td><td>62.7</td><td>68.1</td><td>55.6</td></tr><tr><td>Qwen2-VL + Vgent (Ours)</td><td>7B</td><td>70.3+4.6</td><td>63.5+0.8</td><td>70.1+2.0</td><td>58.4+2.8</td></tr><tr><td>LLaVA-Video [56]</td><td>7B</td><td>69.5</td><td>64.3</td><td>69.2</td><td>59.5</td></tr><tr><td>LLaVA-Video + Vgent (Ours)</td><td>7B</td><td>72.5+3.0</td><td>66.7+2.4</td><td>71.1+1.9</td><td>62.4+2.9</td></tr><tr><td>Qwen2.5-VL [4]</td><td>7B</td><td>68.8</td><td>65.1</td><td>71.1</td><td>56.0</td></tr><tr><td>Qwen2.5-VL + Vgent (Ours)</td><td>7B</td><td>72.1+3.3</td><td>68.9+3.8</td><td>74.3+3.2</td><td>59.7+3.7</td></tr></table>
+
+an accuracy of  $70.4\%$ , surpassing its larger 7B counterpart and improving the base model by  $4.2\%$ . This result underscores the effectiveness of our approach in bridging the performance gap between small models and their larger counterparts. At the category level (Table 6), our framework notably improves Count and Order tasks, which demand event-level understanding and multi-clips reasoning.
+
+In Table 1, we showcase the results of Vgent on the VideoMME [13] benchmark, where it consistently outperforms base models across all video lengths, achieving an average performance gain of  $4.2\%$ . Notably, our framework excels in long-video scenarios, surpassing the best baseline by  $5.4\%$ . These findings highlight the strength of our structured graph-based retrieval and reasoning approach, demonstrating its ability to enhance long-video comprehension by effectively capturing cross-segment dependencies and refining information retrieval for improved reasoning and final response generation.
+
+Comparison with SoTA RAG Methods. In Table 2, we provide a comprehensive comparison of Vgent against state-of-the-art RAG methods on MLVU [57] and VideoMME [13] benchmarks.
+
+(1) Our framework consistently outperforms the RAG baseline, Video-RAG [33], across three different LVLM base models. Unlike Video-RAG [33], which relies on CLIP [41]-based keyframe selection and external tools such as object detection and OCR for frame-level information extraction, Vgent eliminates these dependencies by leveraging LVLMs themselves for graph construction, verification, and intermediate reasoning. This structured approach significantly enhances retrieval precision and reasoning accuracy, leading to more reliable final responses.  
+(2) Our framework also surpasses proprietary RAG-based methods for long-video understanding. Compared to closed-source API-dependent methods which heavily rely on closed-source APIs, our framework is more flexible and effective solution for long-video understanding.
+
+Table 2: RAG methods comparison. † denotes results are sourced from [34]. Proprietary LVLMs refer to methods that rely on closed-source APIs. We include them here for reference only, as our primary focus is on building a self-contained pipeline to improve open-source LVLMs.  
+
+<table><tr><td rowspan="2">Models</td><td rowspan="2">Size</td><td rowspan="2">MLVU</td><td colspan="2">VideoMME</td></tr><tr><td>w/o sub.</td><td>w/ sub.</td></tr><tr><td colspan="5">Proprietary LVLMs</td></tr><tr><td>\( VideoAgent^† \)[47]</td><td>-</td><td>-</td><td>-</td><td>44.4</td></tr><tr><td>\( LLoVi^† \)[52]</td><td>-</td><td>-</td><td>-</td><td>67.7</td></tr><tr><td>\( DrVideo^† \)[34]</td><td>-</td><td>-</td><td>-</td><td>71.7</td></tr><tr><td colspan="5">Open-Source LVLMs</td></tr><tr><td>Qwen2.5-VL + Video-RAG [33]</td><td>3B</td><td>62.2</td><td>60.3</td><td>65.1</td></tr><tr><td>Qwen2.5-VL + Vgent (Ours)</td><td>3B</td><td>70.4</td><td>63.0</td><td>69.6</td></tr><tr><td>LLaVA-Video + Video-RAG [33]</td><td>7B</td><td>71.3</td><td>64.8</td><td>70.0</td></tr><tr><td>LLaVA-Video + Vgent (Ours)</td><td>7B</td><td>72.5</td><td>66.7</td><td>71.1</td></tr><tr><td>Qwen2.5-VL + Video-RAG [33]</td><td>7B</td><td>63.4</td><td>60.5</td><td>65.7</td></tr><tr><td>Qwen2.5-VL + Vgent (Ours)</td><td>7B</td><td>72.1</td><td>68.9</td><td>74.3</td></tr></table>
+
+Table 3: Ablation study results of the performance improvement contributed by each component of our proposed pipeline. SR denotes our proposed structured reasoning.  
+
+<table><tr><td>Models</td><td>MLVU</td><td>VideoMME</td><td>LongVideoBench</td></tr><tr><td>Qwen2.5-VL [4]</td><td>68.8</td><td>71.1</td><td>56.0</td></tr><tr><td>Qwen2.5-VL + NaïveRAG</td><td>65.4</td><td>68.3</td><td>56.2</td></tr><tr><td>Qwen2.5-VL + GraphRAG</td><td>69.5</td><td>72.7</td><td>57.1</td></tr><tr><td>Qwen2.5-VL + NaïveRAG + SR</td><td>68.6</td><td>69.8</td><td>57.3</td></tr><tr><td>Qwen2.5-VL + GraphRAG + SR (default)</td><td>72.1</td><td>74.3</td><td>59.7</td></tr></table>
+
+# 4.3 Ablation Studies
+
+NaïveRAG vs GraphRAG. As shown in Table 3, integrating GraphRAG yields an average improvement of  $2.9\%$  over NaïveRAG, with a particularly notable  $4.1\%$  gain on MLVU [57]. This is because NaïveRAG's difficulty in handling complex queries that requires temporal reasoning across multiple clips, as it treats each video clip as an independent document. In contrast, our GraphRAG effectively preserves semantic relationships between clips, enabling more accurate retrieval and reasoning. By structuring video content into a graph representation, our approach addresses retrieval inconsistencies inherent in traditional RAG methods.
+
+However, the improvement remains marginal compared to the base models. Upon checking failure cases in MLVU, we observe that in  $44\%$  of the failures, the correct clip is actually present within the model's retrieved set, which indicates that while the retrieval was successful, irrelevant retrievals still distract the model, hindering accurate responses. Consequently, a post-retrieval stage is necessary to amplify the potential of our GraphRAG by refining the retrieved nodes and improving reasoning towards more precise answers.
+
+Structured Reasoning (SR). By refining retrieved nodes through intermediate reasoning with structured queries, we achieve an additional  $2.6\%$  improvement on MLVU [57] and  $1.6\%$  on VideoMME [13], resulting in an overall  $3.4\%$  average gain over the base model. This interme
+
+diate reasoning step decomposes complex queries into targeted sub-questions and generates binary or numerical answers. These structured response are then used to systematically filter out irrelevant clips and aggregate relevant information across clips, guiding the model toward the correct final answer. Our findings also indicate that the final improvement is contingent upon Graph-based RAG. Specifically, if SR is applied to NaiveRAG, the inherent inaccuracy of NaiveRAG's retrievals restricts the potential for significant improvement through refinement alone.
+
+Number of retrieval  $r$  We conduct an ablation study to examine the impact of the number of retrieved clips after structured query refinement. Table 4 presents both the overall performance and results across several MLVU [57] subcategories. Among these, Count and Order are two tasks that heavily require reasoning across multiple video clips. Count involves identifying the number of events or actions throughout an entire video, while Order requires the model to arrange multiple events in chronological sequence.  $r$  represents the maximum number of video clips retained after refinement. Our findings indicate that increasing the number of retrieved clips consistently improves performance, particularly for tasks demanding multi-clip reasoning, with the highest performance observed at  $r = 5$ .
+
+Table 4: The number of retrieved clips impacts performance on MLVU [57].  
+
+<table><tr><td>#Retrieval</td><td>Count</td><td>Ego</td><td>Needle</td><td>Order</td><td>PlotQA</td><td>Anomaly</td><td>Topic</td><td>Overall</td></tr><tr><td>r=1</td><td>25.7</td><td>54.2</td><td>75.7</td><td>51.7</td><td>67.4</td><td>71.0</td><td>84.3</td><td>63.2</td></tr><tr><td>r=2</td><td>40.2</td><td>55.6</td><td>78.0</td><td>57.1</td><td>69.1</td><td>73.5</td><td>87.0</td><td>66.9</td></tr><tr><td>r=3</td><td>47.5</td><td>57.1</td><td>78.0</td><td>61.0</td><td>70.0</td><td>71.5</td><td>87.2</td><td>68.4</td></tr><tr><td>r=4</td><td>58.7</td><td>56.6</td><td>78.8</td><td>65.2</td><td>73.6</td><td>72.5</td><td>87.6</td><td>71.0</td></tr><tr><td>r=5 (default)</td><td>58.7</td><td>59.5</td><td>79.7</td><td>67.1</td><td>74.6</td><td>74.0</td><td>88.0</td><td>72.1</td></tr><tr><td>r=6</td><td>58.7</td><td>58.4</td><td>78.8</td><td>67.2</td><td>73.9</td><td>73.5</td><td>87.4</td><td>71.9</td></tr></table>
+
+Further details, including category-level performance on MLVU (A.1), limitations (D), ablation studies on the number of retrievals  $N$  (A.5), confidence-based refinement (A.2), retrieval threshold  $\tau$  (A.6) are provided in the Appendix.
+
+Table 5: Inference time analysis. Since processing time depends on the video duration, we report the normalized time required to process each minute of video.  
+
+<table><tr><td>Model / Time (sec)</td><td>Query Independent (offline)</td><td>Query Dependent (online)</td></tr><tr><td>#Proprietary LVLMs</td><td></td><td></td></tr><tr><td>VideoAgent [47]</td><td>N/A</td><td>67.25</td></tr><tr><td>#Open-Source LVLMs</td><td></td><td></td></tr><tr><td>Qwen2.5VL-7B [4]</td><td>N/A</td><td>2.95</td></tr><tr><td>+Video-RAG [33]</td><td>N/A</td><td>20.81</td></tr><tr><td>+ Vgent (Ours)</td><td>20.13</td><td>3.93</td></tr></table>
+
+# 4.4 Inference Time Analysis
+
+We analyze the computational trade-offs and report the processing times in Table 5 for the API-based method VideoAgent [47], Video-RAG [33] as well as our framework built on Qwen2.5VL [4]. VideoAgent [47] leverages a proprietary LLM (GPT-4 [36]) to iteratively perform self-reflection for frame selection and aggregating key information from the video. Video-RAG [33] relies on query-dependent key frame selection and per-frame object detection, introducing online computational overhead. In contrast, our framework can offline constructs a query-independent graph from the video, which takes 20.13 seconds. Once the graph is built, the online retrieval, reasoning and generation process requires only 3.93 seconds per minute-video.
+
+Our offline graph construction further improves efficiency in multi-question scenarios (e.g., three questions per video in VideoMME [13]). Unlike query-dependent methods that reprocess the entire
+
+![](images/9afa4a918defb62ec267b412bca745d985351f2eff432e674667a7364efc868e.jpg)  
+Figure 3: A qualitative example illustrates our graph-based retrieval-reasoning approach, which connects relevant video clips via shared entities. While the model initially fails to correctly identify the action of opening the laptop, misled by hallucinations from hard negatives. However, adding an intermediate reasoning step that validates each retrieved node through structured subqueries enables the model to generate the accurate response.
+
+video for each question, our approach constructs the graph once, allowing the model to retrieve relevant clips based on entity descriptions—without the need to rewatch the entire video. As a result, our method achieves a  $1.73 \times$  speedup over Video-RAG [33] when performing inference on VideoMME [13].
+
+# 4.5 Qualitative Examples
+
+We show a qualitative example in Figure 3, 5 and 6. Our graph construction effectively connects relevant video clips through shared entities. In Figure 3 the graph-based retrieval system can identify relevant nodes that contains a laptop, with Clip 6 providing crucial evidence to answer the query. However, the model incorrectly responded "No" to the question "Did I open the laptop?", presumably due to hard negatives from multiple clips featuring a opened laptop, hallucinating the model to overlook the closed laptop and the action of opening it.
+
+In contrast, with an intermediate reasoning step, we validate each retrieved node with structured subqueries (e.g., "Is there a laptop open?" "Is someone interacting with the laptop?"). This verified information is aggregated to form an enhanced reasoning chain, allowing the model to correctly infer that the laptop was opened, overcoming the distraction from hard negatives.
+
+# 5 Conclusion
+
+In this work, we introduced a novel graph-based Retrieval-Augmented Generation (RAG) framework designed for long-video understanding. Our approach represents video clips as nodes in a graph and leverages entities to maintain semantic relationships, thereby enhancing retrieval effectiveness. To address retrieval noise, we proposed a structured query refinement strategy that systematically filters out irrelevant clips, ensuring a more precise selection of relevant video content. Additionally, we introduced an intermediate reasoning step that summarizes the response to the structured query, using the filtered retrieved clips as multimodal context to significantly improve the accuracy of the final answer generation. Our framework outperforms state-of-the-art video RAG methods by  $8.6\%$ , demonstrating its effectiveness in enhancing long-video understanding tasks. This work paves the way for more accurate and context-aware long-form video retrieval and reasoning systems.
+
+# References
+
+[1] Md Adnan Arefeen, Biplob Debnath, Md Yusuf Sarwar Uddin, and Srimat Chakradhar. irag: Advancing rag for videos with an incremental approach. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 4341-4348, 2024.  
+[2] Kirolos Ataallah, Xiaogian Shen, Eslam Abdelrahman, Essam Sleiman, Deyao Zhu, Jian Ding, and Mohamed Elhoseiny. Minigpt4-video: Advancing multimodal llms for video understanding with interleaved visual-textual tokens. arXiv preprint arXiv:2404.03413, 2024.  
+[3] Kirolos Ataallah, Xiaogian Shen, Eslam Abdelrahman, Essam Sleiman, Mingchen Zhuge, Jian Ding, Deyao Zhu, Jürgen Schmidhuber, and Mohamed Elhoseiny. Goldfish: Vision-language understanding of arbitrarily long videos. arXiv preprint arXiv:2407.12679, 2024.  
+[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025.  
+[5] Chi-Min Chan, Chunpu Xu, Ruibin Yuan, Hongyin Luo, Wei Xue, Yike Guo, and Jie Fu. Rq-rag: Learning to refine queries for retrieval augmented generation. arXiv preprint arXiv:2404.00610, 2024.  
+[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023.  
+[7] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024.  
+[8] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024.  
+[9] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024.  
+[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019.  
+[11] Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130, 2024.  
+[12] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024.  
+[13] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024.  
+[14] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, W. Zhang, Pan Lu, Conghui He, Xiangyu Yue, Hongsheng Li, and Yu Jiao Qiao. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023.  
+[15] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2, 2023.  
+[16] Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024.  
+[17] Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. Lightrag: Simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779, 2024.  
+[18] Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan Ding, Yongjia Lei, Mahantesh Halappanavar, Ryan A Rossi, Subhabrata Mukherjee, Xianfeng Tang, et al. Retrieval-augmented generation with graphs (graphrag). arXiv preprint arXiv:2501.00309, 2024.
+
+[19] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024.  
+[20] Noureldien Hussein, Efstratios Gavves, and Arnold WM Smeulders. Videograph: Recognizing minutes-long human activities in videos. arXiv preprint arXiv:1905.05143, 2019.  
+[21] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459-9474, 2020.  
+[22] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024.  
+[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023.  
+[24] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023.  
+[25] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024.  
+[26] Mufei Li, Siqi Miao, and Pan Li. Simple is effective: The roles of graphs and large language models in knowledge-graph-based retrieval-augmented generation. arXiv preprint arXiv:2410.20724, 2024.  
+[27] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023.  
+[28] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023.  
+[29] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023.  
+[30] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava- next: Improved reasoning,ocr,and world knowledge,2024.  
+[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024.  
+[32] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Minghui Qiu, Pengcheng Lu, Tao Wang, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.  
+[33] Yongdong Luo, Xiawu Zheng, Xiao Yang, Guilin Li, Haojia Lin, Jina Huang, Jiayi Ji, Fei Chao, Jiebo Luo, and Rongrong Ji. Video-rag: Visually-aligned retrieval-augmented long video comprehension. arXiv preprint arXiv:2411.13093, 2024.  
+[34] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding. arXiv preprint arXiv:2406.12846, 2024.  
+[35] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023.  
+[36] OpenAI. Gpt-4 technical report, 2023.  
+[37] OpenAI. Gpt-4v(ision) system card, 2023.  
+[38] OpenAI. Gpt-4o system card, 2024.  
+[39] Team OpenGVLab. Internvl2: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024.
+
+[40] Team Qwen. Qwen2 technical report, 2024.  
+[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021.  
+[42] Xubin Ren, Lingrui Xu, Long Xia, Shuaiqiang Wang, Dawei Yin, and Chao Huang. Videorag: Retrievalaugmented generation with extreme long-context videos. arXiv preprint arXiv:2502.01549, 2025.  
+[43] Xiaoqian Shen, Yunyang Xiong, Changsheng Zhao, Lemeng Wu, Jun Chen, Chenchen Zhu, Zechun Liu, Fanyi Xiao, Balakrishnan Varadarajan, Florian Bordes, et al. Longvu: Spatiotemporal adaptive compression for long video-language understanding. arXiv preprint arXiv:2410.17434, 2024.  
+[44] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023.  
+[45] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024.  
+[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024.  
+[47] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024.  
+[48] Yang Wang, Gedas Bertasius, Tae-Hyun Oh, Abhinav Gupta, Minh Hoai, and Lorenzo Torresani. Supervoxel attention graphs for long-range video modeling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 155-166, 2021.  
+[49] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for ltm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024.  
+[50] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37: 28828-28857, 2025.  
+[51] Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. C-pack: Packaged resources to advance general chinese embedding, 2023.  
+[52] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple ltm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023.  
+[53] Lu Zhang, Tiancheng Zhao, Heting Ying, Yibo Ma, and Kyusong Lee. Omagent: A multi-modal agent framework for complex video understanding with task divide-and-conquer. arXiv preprint arXiv:2406.16620, 2024.  
+[54] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024.  
+[55] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model, 2024.  
+[56] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024.  
+[57] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024.  
+[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023.
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes]
+
+Justification: The main claims made in the abstract and introduction accurately reflect the paper's contributions and scope.
+
+Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: The paper discussed the limitations of the work in Appendix D.
+
+Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory assumptions and proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [NA]
+
+Justification: The paper does not contain theoretical results.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.  
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental result reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+# Answer: [Yes]
+
+Justification: The paper disclosed all the information needed to reproduce the main experimental results.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.  
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example  
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+# Answer: [Yes]
+
+Justification: The paper provided open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results.
+
+# Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.  
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental setting/details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+# Answer: [Yes]
+
+Justification: The paper specified all the training and test details.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment statistical significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+# Answer: [No]
+
+Justification: While we fixed the random seed to ensure experimental stability and reproducibility, we did not report error bars or conduct statistical significance testing.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.  
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).  
+- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).
+
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.  
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments compute resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [Yes]
+
+Justification: The paper provided sufficient information on the computer resources needed to reproduce the experiments.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code of ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes]
+
+Justification: The research is conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics.
+
+Guidelines:
+
+- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [NA]
+
+Justification: There is no societal impact of the work performed
+
+Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.  
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.  
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.
+
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.  
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [NA]
+
+Justification: The paper poses no such risks.
+
+Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: The models used in the paper are properly credited.
+
+Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URL.  
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.  
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.  
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.
+
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer: [NA]
+
+Justification: The paper does not release new assets.
+
+Guidelines:
+
+- The answer NA means that the paper does not release new assets.  
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and research with human subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [NA]
+
+Justification: The paper does not involve crowdsourcing nor research with human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA]
+
+Justification: The paper does not involve crowdsourcing nor research with human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.  
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.
+
+# 16. Declaration of LLM usage
+
+Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required.
+
+# Answer: [NA]
+
+Justification: The core method development in this research does not involve LLMs.
+
+# Guidelines:
+
+- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.  
+- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described.
+
+# A Experiments
+
+# A.1 Category-level performance on MLVU
+
+Table 6 shows performance on the multiple-choice task of MLVU [57]. Our framework consistently improves all models, enhancing LongVU by  $5.4\%$  and Qwen2.5VL (7B) by  $3.3\%$ . Notably, Vgent achieves  $70.4\%$  accuracy on Qwen2.5VL (3B), surpassing its 7B counterpart and improving the base model by  $4.2\%$ . Significant gains are observed in Count and Order tasks, highlighting the effectiveness of our approach in cross-segment reasoning and long-video understanding.
+
+Table 6: Category-level performance on MLVU [57]. Our framework consistently improves all models, enhancing LongVU by  $5.4\%$  and Qwen2.5VL (7B) by  $3.3\%$ . Notably, Vgent achieves  $70.4\%$  accuracy on Qwen2.5VL (3B), surpassing its 7B counterpart and improving the base model by  $4.2\%$ . Significant gains are observed in Count and Order tasks, highlighting the effectiveness of our approach in cross-segment reasoning and long-video understanding.  
+
+<table><tr><td>Model</td><td>Size</td><td>Count</td><td>Ego</td><td>Needle</td><td>Order</td><td>PlotQA</td><td>Anomaly</td><td>Topic</td><td>Overall</td></tr><tr><td colspan="10">Proprietary LVLMs</td></tr><tr><td>GPT-4o</td><td>-</td><td>46.3</td><td>57.1</td><td>64.8</td><td>56.7</td><td>65.1</td><td>74.5</td><td>87.4</td><td>64.6</td></tr><tr><td colspan="10">Open-Source LVLMs</td></tr><tr><td>InternVL2.5 [8]</td><td>2B</td><td>34.9</td><td>50.4</td><td>61.6</td><td>34.7</td><td>62.8</td><td>61.5</td><td>81.5</td><td>56.7</td></tr><tr><td>InternVL2.5 + Vgent (Ours)</td><td>2B</td><td>59.2</td><td>53.1</td><td>66.7</td><td>38.2</td><td>63.9</td><td>62.0</td><td>81.1</td><td>61.1+4.4</td></tr><tr><td>Qwen2-VL [46]</td><td>2B</td><td>30.1</td><td>56.0</td><td>72.3</td><td>32.8</td><td>65.3</td><td>55.5</td><td>80.4</td><td>58.6</td></tr><tr><td>Qwen2-VL + Vgent (Ours)</td><td>2B</td><td>58.7</td><td>57.6</td><td>76.9</td><td>34.3</td><td>63.8</td><td>59.5</td><td>80.3</td><td>62.5+3.9</td></tr><tr><td>Qwen2.5-VL [4]</td><td>3B</td><td>36.4</td><td>53.0</td><td>77.7</td><td>55.5</td><td>70.1</td><td>75.5</td><td>86.4</td><td>66.2</td></tr><tr><td>Qwen2.5-VL + Vgent (Ours)</td><td>3B</td><td>60.1</td><td>58.1</td><td>78.5</td><td>61.7</td><td>70.5</td><td>74.5</td><td>87.3</td><td>70.4+4.2</td></tr><tr><td>LongVU [54]</td><td>7B</td><td>28.9</td><td>59.3</td><td>76.3</td><td>58.3</td><td>71.6</td><td>76.0</td><td>87.5</td><td>65.4</td></tr><tr><td>LongVU + Vgent (Ours)</td><td>7B</td><td>60.0</td><td>62.3</td><td>76.5</td><td>60.1</td><td>71.6</td><td>76.4</td><td>87.8</td><td>70.8+5.4</td></tr><tr><td>Qwen2-VL [46]</td><td>7B</td><td>32.5</td><td>62.0</td><td>79.1</td><td>53.2</td><td>69.6</td><td>63.0</td><td>85.3</td><td>65.7</td></tr><tr><td>Qwen2-VL + Vgent (Ours)</td><td>7B</td><td>60.2</td><td>65.8</td><td>80.2</td><td>60.2</td><td>70.6</td><td>63.5</td><td>86.1</td><td>70.3+4.6</td></tr><tr><td>LLaVA-Video [56]</td><td>7B</td><td>42.2</td><td>61.5</td><td>76.3</td><td>61.0</td><td>75.8</td><td>72.0</td><td>85.3</td><td>69.5</td></tr><tr><td>LLaVA-Video + Vgent (Ours)</td><td>7B</td><td>58.7</td><td>63.0</td><td>76.9</td><td>67.1</td><td>76.4</td><td>72.5</td><td>86.9</td><td>72.5+3.0</td></tr><tr><td>Qwen2.5-VL [4]</td><td>7B</td><td>41.7</td><td>58.1</td><td>78.0</td><td>61.0</td><td>73.6</td><td>72.5</td><td>87.4</td><td>68.8</td></tr><tr><td>Qwen2.5-VL + Vgent (Ours)</td><td>7B</td><td>58.7</td><td>59.5</td><td>79.7</td><td>67.1</td><td>74.6</td><td>74.0</td><td>88.1</td><td>72.1+3.3</td></tr></table>
+
+# A.2 Confidence-based Refinement
+
+A straightforward solution is to filter out the hard negative retrievals by their relevance scores. Initially, we experimented with confidence-based refinement, as used in VideoAgent [47], where the model self-reflect the relevance of retrieved nodes. However, this approach proved ineffective in our case, as the confidence score failed to reliably reflect video clip relevance, leading to an average improvement of only  $0.2\%$ , as shown in Table 7.
+
+Table 7: Ablation study results of the performance improvement contributed by each component of our proposed pipeline. CR denotes confidence-based reasoning and SR is our proposed structured reasoning.  
+
+<table><tr><td>Models</td><td>MLVU</td><td>VideoMME</td><td>LongVideoBench</td></tr><tr><td>Qwen2.5-VL [4]</td><td>68.8</td><td>71.1</td><td>56.0</td></tr><tr><td>Qwen2.5-VL + GraphRAG + CR</td><td>69.5</td><td>72.9</td><td>57.5</td></tr><tr><td>Qwen2.5-VL + GraphRAG + SR (default)</td><td>72.1</td><td>74.3</td><td>59.7</td></tr></table>
+
+# A.3 Baseline Details
+
+NaiveRAG: Following GoldFish [3], we construct a NaiveRAG baseline for video understanding by representing each video clip as plain text and retrieving relevant clips based on similarity to the query.
+
+Video-RAG: [33]: This method selects keyframes by evaluating CLIP similarity between each frame's features and the text embeddings of keywords extracted from the question. Additionally, an object detection model and an Optical Character Recognition (OCR) model are applied to each keyframe to extract detailed information.
+
+Proprietary LLM-based: VideoAgent [47], LLoVi [52], DrVideo [34] and VideoTree [49] utilizes interactive reasoning and planning of proprietary LLM APIs to enhance long-video understanding.
+
+# A.4 Retrieval Embedding
+
+We explore different types of retrieval embeddings, i.e., CLIP [41], BERT [10] and BGE [51] on VideoMME [13] benchmark with Qwen2.5-VL [4] backbone, as shown in Figure 4 (left).
+
+# A.5 Number of Retrieval  $N$
+
+We conduct ablation on the number of retrieval  $N$  before structured reasoning (SR) on VideoMME [13] benchmark with Qwen2.5-VL [4] backbone, as shown in Figure 4 (middle). We set  $N = 20$  by default.
+
+# A.6 Retrieval Threshold  $\tau$
+
+We investigate retrieval threshold  $\tau$  on VideoMME [13] benchmark with Qwen2.5-VL [4] backbone, as shown in Figure 4 (right). As the value of  $\tau$  increases, less video clips are retrieved based on similarity scores, potentially leading to the loss of relevant information. We set  $\tau = 0.5$  by default.
+
+# A.7 Qualitative Results
+
+We show a qualitative example in Figure 3, 5 and 6. Our graph construction effectively connects relevant video clips through shared entities. In Figure 3 the graph-based retrieval system can identify relevant nodes that contains a laptop, with Clip 6 providing crucial evidence to answer the query. However, the model incorrectly responded "No" to the question "Did I open the laptop?", presumably due to hard negatives from multiple clips featuring a opened laptop, hallucinating the model to overlook the closed laptop and the action of opening it.
+
+In contrast, with an intermediate reasoning step, we validate each retrieved node with structured subqueries (e.g., "Is there a laptop open?" "Is someone interacting with the laptop?"). This verified information is aggregated to form an enhanced reasoning chain, allowing the model to correctly infer that the laptop was opened, overcoming the distraction from hard negatives.
+
+![](images/9757e7df83977198a23d65a5af889eca8e4b4323bdb66881f9b70645fedab7d7.jpg)  
+Figure 4: Ablation studies. Left: retrieval embedding. Middle: number of retrieval  $N$  before SR. Right: ablation on retrieval threshold  $\tau$ .
+
+![](images/489aa0327283760c0760f9b089fd21646fb8c79581508555cb357d7e30857468.jpg)
+
+![](images/1c28ee00c39aea8343e55736cd54145ca28ca5d862e3bc25e3a90046f156d4cc.jpg)
+
+![](images/898984f79dcfd42fa03a00b7a9e44883ab2e11009730081e6105943f6b9d2021.jpg)  
+Figure 5: A qualitative example illustrates our graph-based retrieval-reasoning approach.
+
+![](images/94d4c461149cb8e252ecc21601bbcc7ce98daeeefd64a4ad383cf978e3d0181d.jpg)  
+Figure 6: A qualitative example illustrates our graph-based retrieval-reasoning approach.
+
+# B Prompts
+
+# B.1 Visual Entity Extraction
+
+Figure 7 illustrates the prompts used to describe entities, actions, and scenes given a video segment for the LVLM.
+
+# B.2 Keyword Extraction
+
+Figure 8 presents the prompt designed for the LVLM to perform task identification and extract keywords from the original question to facilitate retrieval.
+
+# B.3 Subqueries Generation
+
+Figure 9 presents the prompt designed for the LVLM to generate structured subqueries for retrieved nodes refinement.
+
+# Prompt: Describe entities, actions, scenes.
+
+Please analyze the given video and extract key information in a structured JSON format in English. Identify and describe:
+
+Entities: List all distinct objects, people, animals, or other significant elements present in the video.
+
+Actions: If the entities are interacting, describe their actions and relationships in a structured manner.
+
+Scenes: Identify and describe the locations, environments, or contexts where the events occur. If the video is filmed from a first-person point of view, please also describe "subject" as "me" and actions or interactions from this person.
+
+Ensure the output strictly follows the JSON format below:  
+{ "entities": ["entity name": "", "description": "", "actions": ["entity name": "action description"], "scenes": ["location": ""]}
+
+The "entity name" in actions should belong to "entity name" in entities.
+
+Each section should be detailed but concise, capturing all relevant interactions and contextual elements from the video. Avoid unnecessary text outside the JSON output.
+
+Figure 7: Prompt for video segment description.
+
+# Prompt: Task Identification and Keyword Extraction.
+
+Given a question of a long video and potential candidates:
+
+Question: {query}
+
+Candidates: {candidates}
+
+You need to retrieve the relevant video segments to answer the question. Note that you do not need to see the video. But based on the question please think step by step what are the important things for retrieval.
+
+[keywords] Please identify the information, like entities, scene, action from the question that is important to retrieve the segments for further answer the question. Do not include the candidates in the keywords.
+
+[candidates_necessary] Do you think the information in the candidates is necessary for retrieval? Answer yes or no.
+
+[multiple] Do you think it needs to aggregate the information from multiple segments to answer the question? ONLY answer yes or no.
+
+[time] Please identify if it can tell the question is asking which part of the video. Answer begin, end or none.
+
+[tool] Do you think it needs additional step for answering the question, please select from [object counting, action counting, order, none].
+
+[global] Can this question be answered based on the overall understanding of the whole video? (e.g., "What is the main topic of the video?" or "What is the main content of the video?")
+
+Please output the final answer in json format, for example:
+
+{"multiple": "no", "keywords": ["man in black"], "time": "begin", "tool": none, "candidates_necessary": "yes", "global": "yes"}
+
+Figure 8: Prompt for task identification and keywords extraction.
+
+# Prompt: Subqueries Generation.
+
+Given a question of a long video and potential candidates:
+
+Question: {query}
+
+Candidates: {candidates}
+
+Given a multiple-choice question about a video, break it down into several sub-questions that analyze the key elements required to answer it step by step.
+
+First, identify the key subject or event in the question (e.g., an object, an animal, an action, or a location). Form yes/no or counting questions to verify the presence of the subject or event in the video (e.g., "Does the video show [subject/event]?"). Ensure the sub-questions cover all necessary aspects to reach the correct answer.
+
+$= =$  important  $= =$  Please give me the answer in JSON format. Do not include references to specific time positions in the video when generating questions (e.g., "at the beginning," "in the middle," or "at the end") Do not go through all the numbers in the candidates for counting quotients.
+
+Figure 9: Prompt for subqueries generation.
+
+# C Model Output Examples
+
+# C.1 Visual Entity Extraction
+
+```jsonl
+{
+"idx": 0,
+"info": {
+"entities": [
+{
+"entity name": "sailboat",
+"description": "A classic sailboat with white sails
+and wooden rigging"
+},
+{
+"entity name": "man",
+"description": "A man wearing a dark sweater"
+},
+{
+"entity name": "ocean",
+"description": "A calm ocean under a partly cloudy sky"
+}
+],
+"actions": [
+{
+"entity name": "sailboat",
+"description": "sailing smoothly on the water"
+},
+{
+"entity name": "man",
+"description": "steering the sailboat"
+},
+{
+"entity name": "man",
+"description": "looking around"
+},
+{
+"entity name": "man",
+"description": "adjusting his hair"
+}
+],
+"scenes": [
+```
+
+```txt
+"location": "open sea" } 1
+```
+
+# D Limitation
+
+In this work, we represent video content using textual descriptions—such as entities and their associated details—as a lightweight and efficient alternative to raw visual features. However, we do not incorporate visual embeddings or frame-level features into our graph. While computing similarity across frames can be computationally intensive, it remains a promising direction for future improvement.
+
+Additionally, our framework is model-agnostic and compatible with any LVLM, meaning its performance is inherently bounded by the capabilities of the base LVLM. As more powerful LVLMs emerge, our pipeline can be readily adapted to take advantage of their enhanced video understanding and reasoning abilities.

2025/Vgent_ Graph-based Retrieval-Reasoning-Augmented Generation For Long Video Understanding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fdc218e143be3c6f9a234dd66e4e36154e7442887d8deeb81ca56a4d708047cd
+size 839925

2025/ViCrit_ A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs/9f5de5f5-6130-4a1f-933d-16e900744798_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a51a25d6d51c22d58bd65f862bff649de25fd309035b455098e179239e4c1cf
+size 4361152

2025/ViCrit_ A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77b4a13e8916b69b4fdd9ff4c3c2d551d4b8a21cdc8fb2e3cc6623d9691beac3
+size 647832