Daily Papers

Deriving compact and temporally aware visual representations from dynamic scenes is essential for successful execution of sequential scene understanding tasks such as visual tracking and robotic manipulation. In this paper, we introduce Token Bottleneck (ToBo), a simple yet intuitive self-supervised learning pipeline that squeezes a scene into a bottleneck token and predicts the subsequent scene using minimal patches as hints. The ToBo pipeline facilitates the learning of sequential scene representations by conservatively encoding the reference scene into a compact bottleneck token during the squeeze step. In the expansion step, we guide the model to capture temporal dynamics by predicting the target scene using the bottleneck token along with few target patches as hints. This design encourages the vision backbone to embed temporal dependencies, thereby enabling understanding of dynamic transitions across scenes. Extensive experiments in diverse sequential tasks, including video label propagation and robot manipulation in simulated environments demonstrate the superiority of ToBo over baselines. Moreover, deploying our pre-trained model on physical robots confirms its robustness and effectiveness in real-world environments. We further validate the scalability of ToBo across different model scales.

Autoregressive language models generate text one token at a time, yet natural language is inherently structured in multi-token units, including phrases, n-grams, and collocations that carry meaning jointly. This one-token bottleneck limits both the expressiveness of the model during pre-training and its throughput at inference time. Existing remedies such as speculative decoding or diffusion-based language models either leave the underlying bottleneck intact or sacrifice the causal structure essential to language modeling. We propose BitLM, a language model that represents each token as a fixed-length binary code and employs a lightweight diffusion head to denoise multiple tokens in parallel within each block. Crucially, BitLM preserves left-to-right causal attention across blocks while making joint lexical decisions within each block, combining the reliability of autoregressive modeling with the parallelism of iterative refinement. By replacing the large-vocabulary softmax with bitwise denoising, BitLM reframes token generation as iterative commitment in a compact binary space, enabling more efficient pre-training and substantially faster inference without altering the causal foundation that makes language models effective. Our results demonstrate that the one-token-at-a-time paradigm is not a fundamental requirement but an interface choice, and that changing it can yield a stronger and faster language model. We hope BitLM points toward a promising direction for next-generation language model architectures.

LLM reasoning transparency is a critical affordance for understanding model decisions, mitigating misuse and misalignment, and debugging surprising model behaviors. However, DiffusionGemma performs a larger fraction of its computation in a continuous latent space; does this make its reasoning less transparent? We study this question by decomposing transparency into two components: variable transparency, whether we understand intermediate snapshots of a model's computational state; and algorithmic transparency, whether we can use these snapshots to reconstruct the process by which the model arrived at its outputs. Naively, DiffusionGemma has poor variable transparency: its opaque serial depth, the amount of serial computation that occurs in between interpretable model states, seems at first 28.6X higher than the corresponding autoregressive Gemma 4 model. However, we show that we can map the information flowing between denoising steps through an interpretable token bottleneck with no decrease in downstream performance. Treating these intermediate states as interpretable reduces the opaque serial depth to just 1.1X that of Gemma 4. Algorithmic transparency is harder for diffusion models than for autoregressive models because all token predictions in the canvas can change at every denoising step, giving the model the power to implement complicated distributed algorithms during the denoising process. To begin bridging this gap, we conduct a suite of interpretability case studies, uncovering initial evidence of novel diffusion-specific phenomena such as non-chronological reasoning, token and sequence smearing, and intermediate-context reasoning. Finally, we test monitorability, a key application of transparency that measures whether model outputs are useful for downstream tasks. We find that DiffusionGemma is similarly monitorable to Gemma 4.

Soft attention is a critical mechanism powering LLMs to locate relevant parts within a given context. However, individual attention weights are determined by the similarity of only a single query and key token vector. This "single token attention" bottlenecks the amount of information used in distinguishing a relevant part from the rest of the context. To address this issue, we propose a new attention method, Multi-Token Attention (MTA), which allows LLMs to condition their attention weights on multiple query and key vectors simultaneously. This is achieved by applying convolution operations over queries, keys and heads, allowing nearby queries and keys to affect each other's attention weights for more precise attention. As a result, our method can locate relevant context using richer, more nuanced information that can exceed a single vector's capacity. Through extensive evaluations, we demonstrate that MTA achieves enhanced performance on a range of popular benchmarks. Notably, it outperforms Transformer baseline models on standard language modeling tasks, and on tasks that require searching for information within long contexts, where our method's ability to leverage richer information proves particularly beneficial.

Camouflaged object detection (COD) segments objects that intentionally blend with the background, so predictions depend on subtle texture and boundary cues. COD is often needed under tight on-device memory and latency budgets, making low-bit inference highly desirable. However, COD is unusually hard to quantify aggressively. We study post-training W4A4 quantization of Transformer-based COD and find a task-specific cliff: heavy-tailed background tokens dominate a shared activation range, inflating the step size and pushing weak-but-structured boundary cues into the zero bin. This exposes a token-local bottleneck -- remove cross-token range domination and bound the zero-bin mass under 4-bit activations. To address this, we introduce COD-TDQ, a COD-aware Token-group Dual-constraint activation Quantization method. COD-TDQ addresses this token-local bottleneck with two coupled steps: Direct-Sum Token-Group (DSTG) assigns token-group scales to suppress cross-token range domination, and Dual-Constraint Range Projection (DCRP) projects each token-group clip range to keep the step-to-dispersion ratio and the zero-bin mass bounded. Across four COD benchmarks and two baseline models (CFRN and ESCNet), COD-TDQ consistently achieves an Sαscore more than 0.12 higher than that of the state-of-the-art quantization method without retraining. The code will be released.

For robotic agents operating in dynamic environments, learning visual state representations from streaming video observations is essential for sequential decision making. Recent self-supervised learning methods have shown strong transferability across vision tasks, but they do not explicitly address what a good visual state should encode. We argue that effective visual states must capture what-is-where by jointly encoding the semantic identities of scene elements and their spatial locations, enabling reliable detection of subtle dynamics across observations. To this end, we propose CroBo, a visual state representation learning framework based on a global-to-local reconstruction objective. Given a reference observation compressed into a compact bottleneck token, CroBo learns to reconstruct heavily masked patches in a local target crop from sparse visible cues, using the global bottleneck token as context. This learning objective encourages the bottleneck token to encode a fine-grained representation of scene-wide semantic entities, including their identities, spatial locations, and configurations. As a result, the learned visual states reveal how scene elements move and interact over time, supporting sequential decision making. We evaluate CroBo on diverse vision-based robot policy learning benchmarks, where it achieves state-of-the-art performance. Reconstruction analyses and perceptual straightness experiments further show that the learned representations preserve pixel-level scene composition and encode what-moves-where across observations. Project page available at: https://seokminlee-chris.github.io/CroBo-ProjectPage.

As text generation has become a core capability of modern Large Language Models (LLMs), it underpins a wide range of downstream applications. However, most existing LLMs rely on autoregressive (AR) generation, producing one token at a time based on previously generated context-resulting in limited generation speed due to the inherently sequential nature of the process. To address this challenge, an increasing number of researchers have begun exploring parallel text generation-a broad class of techniques aimed at breaking the token-by-token generation bottleneck and improving inference efficiency. Despite growing interest, there remains a lack of comprehensive analysis on what specific techniques constitute parallel text generation and how they improve inference performance. To bridge this gap, we present a systematic survey of parallel text generation methods. We categorize existing approaches into AR-based and Non-AR-based paradigms, and provide a detailed examination of the core techniques within each category. Following this taxonomy, we assess their theoretical trade-offs in terms of speed, quality, and efficiency, and examine their potential for combination and comparison with alternative acceleration strategies. Finally, based on our findings, we highlight recent advancements, identify open challenges, and outline promising directions for future research in parallel text generation. We have also created a GitHub repository for indexing relevant papers and open resources available at https://github.com/zhanglingzhe0820/Awesome-Parallel-Text-Generation.

Video large language models (Vid-LLMs), which excel in diverse video-language tasks, can be effectively constructed by adapting image-pretrained vision-language models (VLMs). However, this adaptation remains challenging, as it requires processing dense and temporally extended visual inputs that exceed the capacity of image-based models. This paper identifies the perception bottleneck and token overload as key challenges in extending image-based VLMs to the video domain. To address these issues, we propose D-CoDe, a training-free adaptation framework that incorporates dynamic compression and question decomposition. Specifically, dynamic compression alleviates the perception bottleneck through adaptive selection of representative frames and content-aware aggregation of spatial tokens, thereby reducing redundancy while preserving informative content. In parallel, question decomposition mitigates token overload by reformulating the original query into sub-questions, guiding the model to focus on distinct aspects of the video and enabling more comprehensive understanding. Experiments demonstrate that D-CoDe effectively improves video understanding across various benchmarks. Furthermore, strong performance on the challenging long-video benchmark highlights the potential of D-CoDe in handling complex video-language tasks. Code is available at https://github.com/hukcc/D-CoDe.

Diffusion Language Models (DLMs) offer attractive advantages over Auto-Regressive (AR) models, such as full-attention parallel decoding and flexible generation. However, they suffer from a notable train-inference mismatch: DLMs are trained with a static, single-step masked prediction objective, but deployed through a multi-step progressive denoising trajectory. We propose MemDLM (Memory-Enhanced DLM), which narrows this gap by embedding a simulated denoising process into training via Bi-level Optimization. An inner loop updates a set of fast weights, forming a Parametric Memory that captures the local trajectory experience of each sample, while an outer loop updates the base model conditioned on this memory. By offloading memorization pressure from token representations to parameters, MemDLM yields faster convergence and lower training loss. Moreover, the inner loop can be re-enabled at inference time as an adaptation step, yielding additional gains on long-context understanding. We find that, when activated at inference time, this Parametric Memory acts as an emergent in-weight retrieval mechanism, helping MemDLM further reduce token-level attention bottlenecks on challenging Needle-in-a-Haystack retrieval tasks. Code: https://github.com/JarvisPei/MemDLM.

Advanced Large Language Models (LLMs) have achieved impressive performance across a wide range of complex and long-context natural language tasks. However, performing long-context LLM inference locally on a commodity GPU (a PC) with privacy concerns remains challenging due to the increasing memory demands of the key-value (KV) cache. Existing systems typically identify important tokens and selectively offload their KV data to GPU and CPU memory. The KV data needs to be offloaded to disk due to the limited memory on a commodity GPU, but the process is bottlenecked by token importance evaluation overhead and the disk's low bandwidth. In this paper, we present LeoAM, the first efficient importance-aware long-context LLM inference system for a single commodity GPU with adaptive hierarchical GPU-CPU-Disk KV management. Our system employs an adaptive KV management strategy that partitions KV data into variable-sized chunks based on the skewed distribution of attention weights across different layers to reduce computational and additional transmission overheads. Moreover, we propose a lightweight KV abstract method, which minimizes transmission latency by storing and extracting the KV abstract of each chunk on disk instead of the full KV data. LeoAM also leverages the dynamic compression and pipeline techniques to further accelerate inference. Experimental results demonstrate that LongInfer achieves an average inference latency speedup of 3.46x, while maintaining comparable LLM response quality. In scenarios with larger batch sizes, it achieves up to a 5.47x speedup.

The quadratic complexity of attention remains the central bottleneck in long-context inference for large language models. Prior acceleration methods either sparsify the attention map with structured patterns or permanently evict tokens at specific layers, which can retain irrelevant tokens or rely on irreversible early decisions despite the layer-/head-wise dynamics of token importance. In this paper, we propose Token Sparse Attention, a lightweight and dynamic token-level sparsification mechanism that compresses per-head Q, K, V to a reduced token set during attention and then decompresses the output back to the original sequence, enabling token information to be reconsidered in subsequent layers. Furthermore, Token Sparse Attention exposes a new design point at the intersection of token selection and sparse attention. Our approach is fully compatible with dense attention implementations, including Flash Attention, and can be seamlessly composed with existing sparse attention kernels. Experimental results show that Token Sparse Attention consistently improves accuracy-latency trade-off, achieving up to times3.23 attention speedup at 128K context with less than 1% accuracy degradation. These results demonstrate that dynamic and interleaved token-level sparsification is a complementary and effective strategy for scalable long-context inference.

Motivated by the success of Transformers when applied to sequences of discrete symbols, token-based world models (TBWMs) were recently proposed as sample-efficient methods. In TBWMs, the world model consumes agent experience as a language-like sequence of tokens, where each observation constitutes a sub-sequence. However, during imagination, the sequential token-by-token generation of next observations results in a severe bottleneck, leading to long training times, poor GPU utilization, and limited representations. To resolve this bottleneck, we devise a novel Parallel Observation Prediction (POP) mechanism. POP augments a Retentive Network (RetNet) with a novel forward mode tailored to our reinforcement learning setting. We incorporate POP in a novel TBWM agent named REM (Retentive Environment Model), showcasing a 15.4x faster imagination compared to prior TBWMs. REM attains superhuman performance on 12 out of 26 games of the Atari 100K benchmark, while training in less than 12 hours. Our code is available at https://github.com/leor-c/REM.

Large Language Models (LLMs) have achieved remarkable success but face significant computational and memory challenges, particularly due to their extensive output vocabularies. The final linear projection layer, mapping hidden states to vocabulary-sized logits, often constitutes a substantial portion of the model's parameters and computational cost during inference. Existing methods like adaptive softmax or hierarchical softmax introduce structural complexities. In this paper, we propose VQ-Logits, a novel approach that leverages Vector Quantization (VQ) to drastically reduce the parameter count and computational load of the LLM output layer. VQ-Logits replaces the large V * dmodel output embedding matrix with a small, shared codebook of K embedding vectors (K << V ). Each token in the vocabulary is mapped to one of these K codebook vectors. The LLM predicts logits over this compact codebook, which are then efficiently "scattered" to the full vocabulary space using the learned or preassigned mapping. We demonstrate through extensive experiments on standard language modeling benchmarks (e.g., WikiText-103, C4) that VQ-Logits can achieve up to 99% parameter reduction in the output layer and 6x speedup in logit computation, with only a marginal 4% increase in perplexity compared to full softmax baselines. We further provide detailed ablation studies on codebook size, initialization, and learning strategies, showcasing the robustness and effectiveness of our approach.

Large language models (LLMs) have recently demonstrated remarkable progress in reasoning capabilities through reinforcement learning with verifiable rewards (RLVR). By leveraging simple rule-based rewards, RL effectively incentivizes LLMs to produce extended chain-of-thought (CoT) reasoning trajectories, progressively guiding them toward correct answers. However, existing approaches remain largely heuristic and intuition-driven, limiting the development of principled methodologies. In this paper, we present a theoretical characterization of LLM reasoning grounded in information bottleneck (IB) principle, introducing IB-aware reasoning optimization (IBRO), a framework that encourages reasoning trajectories to be both informative about the final correct answer and generalizable across diverse prompts. We derive a practical token-level surrogate objective and propose an efficient approximation, resulting in the lightweight IB regularization method. This technique integrates seamlessly into existing RL-based post-training frameworks without additional computational overhead, requiring only a one-line code modification. Empirically, we validate IB regularization across multiple mathematical reasoning benchmarks and RL algorithms, demonstrating consistent improvements in LLM reasoning performance.

Recent studies reveal that large language models (LLMs) exhibit limited logical reasoning abilities in mathematical problem-solving, instead often relying on pattern-matching and memorization. We systematically analyze this limitation, focusing on logical relationship understanding, which is a core capability underlying genuine logical reasoning, and reveal that errors related to this capability account for over 90\% of incorrect predictions, with Chain-of-Thought Supervised Fine-Tuning (CoT-SFT) failing to substantially reduce these errors. To address this bottleneck, we propose First-Step Logical Reasoning (FSLR), a lightweight training framework targeting logical relationship understanding. Our key insight is that the first planning step-identifying which variables to use and which operation to apply-encourages the model to derive logical relationships directly from the problem statement. By training models on this isolated step, FSLR provides explicit supervision for logical relationship understanding, unlike CoT-SFT which implicitly embeds such relationships within complete solution trajectories. Extensive experiments across multiple models and datasets demonstrate that FSLR consistently outperforms CoT-SFT under both in-distribution and out-of-distribution settings, with average improvements of 3.2\% and 4.6\%, respectively. Moreover, FSLR achieves 4-6x faster training and reduces training token consumption by over 80\%.

Memory and computation remain core bottlenecks in long-horizon LLM inference due to the quadratic cost of self-attention and the ever-growing key-value (KV) cache. Existing strategies for memory-bounded inference, such as quantization, offloading, or heuristic KV eviction, either incur high orchestration costs or rely on unreliable attention-based proxies of importance. We propose TRIM-KV, a novel approach that learns each token's intrinsic importance at creation time via a lightweight retention gate. Each gate predicts a scalar retention score that decays over time, reflecting the long-term utility of the token for a specific layer and head. Tokens with low scores are evicted when the memory budget is exceeded, ensuring that the cache always contains the most critical tokens. TRIM-KV is trained efficiently through distillation from a frozen LLM combined with a capacity loss, requiring only gate fine-tuning and adding negligible inference overhead. Across mathematical reasoning (GSM8K, MATH-500, AIME24), procedural generation (LongProc), conversational long-memory benchmarks (LongMemEval), and long-context understanding (LongBench and SCBench), TRIM-KV consistently outperforms strong eviction and learnable retrieval baselines, especially in low-memory regimes. Remarkably, it even surpasses full-cache models in some settings, showing that selective retention can serve as a form of regularization, suppressing noise from uninformative tokens. Qualitative analyses further reveal that learned retention scores align with human intuition, naturally recovering heuristics such as sink tokens, sliding windows, and gist compression without explicit design. Beyond efficiency, retention scores provide insights into layer- and head-specific roles, suggesting a new path toward LLM interpretability.

Current skeleton representation learning paradigms face distinct limitations: Contrastive Learning (CL) often overlooks fine-grained motion details, while Masked Auto-Encoders (MAE) rely on coordinate-level reconstruction. This reconstruction inherently demands dense token sequences and heavy decoders, wasting pre-training computation on discarded components and forcing downstream inference to process dense token grids. To resolve these bottlenecks, we propose SLiM (Skeleton Less is More), a compact-token framework that unifies masked feature prediction and contrastive learning via a shared encoder. By shifting the objective from raw coordinate reconstruction to decoder-free, teacher-guided feature prediction, SLiM breaks the reliance on dense tokenization and enables effective learning with a highly compact token grid. Crucially, to prevent trivial shortcut learning arising from strong inter-joint dependencies of human, we introduce Semantic Tube Masking together with Skeleton-Aware Augmentations to enforce deep skeletal-temporal reasoning and anatomical consistency. Extensive experiments across multiple downstream protocols demonstrate that SLiM achieves state-of-the-art performance while structurally reducing inference computation by 7.89times compared to dense-token MAE baselines.

The key-value (KV) cache is a major bottleneck in long-context inference, where memory and computation grow with sequence length. Existing KV eviction methods reduce this cost but typically degrade performance relative to full-cache inference. Our key insight is that full-cache attention is not always optimal: in long contexts, irrelevant tokens can dilute attention away from useful evidence, so selective, learnable eviction can improve generation rather than merely approximate the full cache. We introduce a global retention-based KV eviction method that learns each token's future utility under a unified memory budget. Lightweight retention gates assign utility scores to cached KV entries, and a shared final scoring projection calibrates these scores across all layers and heads. This enables a single global eviction policy in which tokens from different layers, heads, and modalities compete directly for cache capacity. We further provide theoretical analysis showing that preferentially retaining useful tokens reduces attention dilution, and we justify geometric retention as a query-agnostic proxy for future utility. Across diverse long-context language and vision-language reasoning, and multi-turn dialogue benchmarks, our method substantially reduces KV memory while matching or surpassing full-cache inference. These results suggest that learned, globally calibrated KV eviction is not only a compression technique, but also a mechanism for improving long-context reasoning.

Latent actions provide a compact interface between action-free video and downstream decision-making, yet existing Latent Action Models (LAMs) force every transition through a fixed-capacity bottleneck. We identify a bottleneck trade-off: overly tight codes can discard transition cues needed for action alignment, while overly loose codes preserve additional transition variation that must be resolved when alignment labels are scarce or narrowly distributed. FlexLAM replaces this fixed capacity with variable-length latent actions trained by nested dropout, yielding prefix-valid codes that capture compact transition structure first and add detail only when needed, without new architectures or losses. A single FlexLAM matches or surpasses separately trained fixed-capacity LAMs at every evaluated token budget under standard scarce-label supervision and under a low-return single-task alignment stress test, indicating that FlexLAM is not merely adjustable at inference time but learns a better latent-action interface at the same token budgets. The same model supports inference-time token-budget adjustment without retraining, and FlexLAM improves Ego4D transition reconstruction. These results suggest that variable-length latent actions are an architecture-free, drop-in upgrade to the fixed-capacity bottleneck in latent action models, latent-action world models, and video-pretrained action interfaces.

Recent advancements in Deep Neural Network (DNN) models have significantly improved performance across computer vision tasks. However, achieving highly generalizable and high-performing vision models requires expansive datasets, resulting in significant storage requirements. This storage challenge is a critical bottleneck for scaling up models. A recent breakthrough by SeiT proposed the use of Vector-Quantized (VQ) feature vectors (i.e., tokens) as network inputs for vision classification. This approach achieved 90% of the performance of a model trained on full-pixel images with only 1% of the storage. While SeiT needs labeled data, its potential in scenarios beyond fully supervised learning remains largely untapped. In this paper, we extend SeiT by integrating Masked Token Modeling (MTM) for self-supervised pre-training. Recognizing that self-supervised approaches often demand more data due to the lack of labels, we introduce TokenAdapt and ColorAdapt. These methods facilitate comprehensive token-friendly data augmentation, effectively addressing the increased data requirements of self-supervised learning. We evaluate our approach across various scenarios, including storage-efficient ImageNet-1k classification, fine-grained classification, ADE-20k semantic segmentation, and robustness benchmarks. Experimental results demonstrate consistent performance improvement in diverse experiments, validating the effectiveness of our method. Code is available at https://github.com/naver-ai/seit.

Transformer LLMs have been shown to exhibit strong reasoning ability that scales with inference-time compute, most prominently through token-space "thinking" chains of thought. A growing line of work pushes extra computation into the model's latent space, which we term Auxiliary Latent-Space Computation (ALSC). Existing ALSC methods largely fall into three buckets: (i) token-mediated latent rollouts, (ii) residual/activation steering, and (iii) memory (KV) compression. An underexplored alternative is memory consolidation/reconsolidation, two processes in the brain that are responsible for stabilising newly formed memory traces, and, upon recall, transiently rendering established traces plastic such they can integrate new contextual information before restabilising. In Transformer LLMs, this can be seen as analogous to performing in-place rewrites of new KV segments, and rewrites of recalled past segments. In this work, we give a theoretical justification as to why memory (re)consolidation via KV cache rewrites is beneficial for improved reasoning. We do this through the lens of Information Bottleneck (IB) theory, which posits that model generalisation emerges from an optimal balance between input information compression and retention of predictive information in latent representations. We then introduce the Bottlenecked Transformer, which augments a backbone LLM with a Cache Processor, an auxiliary Transformer that performs periodic, non-causal, in-place KV rewrites at newline-delimited reasoning step boundaries. The Processor consolidates recently written KV entries and reconsolidates a small, top-k attention-selected set of prior entries. We evaluate our Bottlenecked Transformer architecture on math reasoning benchmarks. Our model sees consistent performance gains over vanilla Transformers and pause-token augmented baselines, with gains of up to +6.6pp for selected tasks/backbones.

As supervised fine-tuning (SFT) evolves from a lightweight post-training step into a compute-intensive phase rivaling mid-training in scale, data efficiency has become critical for aligning large language models (LLMs) under tight budgets. Existing data pruning methods suffer from a fragmented design: they operate either at the sample level or the token level in isolation, failing to jointly optimize both dimensions. This disconnect leads to significant inefficiencies--high-value samples may still contain redundant tokens, while token-level pruning often discards crucial instructional or corrective signals embedded in individual examples. To address this bottleneck, we introduce the Error-Uncertainty (EU) Plane, a diagnostic framework that jointly characterizes the heterogeneous utility of training data across samples and tokens. Guided by this insight, we propose Quadrant-based Tuning (Q-Tuning), a unified framework that strategically coordinates sample pruning and token pruning. Q-Tuning employs a two-stage strategy: first, it performs sample-level triage to retain examples rich in informative misconceptions or calibration signals; second, it applies an asymmetric token-pruning policy, using a context-aware scoring mechanism to trim less salient tokens exclusively from misconception samples while preserving calibration samples in their entirety. Our method sets a new state of the art across five diverse benchmarks. Remarkably, on SmolLM2-1.7B, Q-Tuning achieves a +38\% average improvement over the full-data SFT baseline using only 12.5\% of the original training data. As the first dynamic pruning approach to consistently outperform full-data training, Q-Tuning provides a practical and scalable blueprint for maximizing data utilization in budget-constrained LLM SFT.

Chain-of-Thought (CoT) prompting improves LLM accuracy on complex tasks but often increases token usage and inference cost. Existing "Budget Forcing" methods reducing cost via fine-tuning with heuristic length penalties, suppress both essential reasoning and redundant filler. We recast efficient reasoning as a lossy compression problem under the Information Bottleneck (IB) principle, and identify a key theoretical gap when applying naive IB to transformers: attention violates the Markov property between prompt, reasoning trace, and response. To resolve this issue, we model CoT generation under the Conditional Information Bottleneck (CIB) principle, where the reasoning trace Z acts as a computational bridge that contains only the information about the response Y that is not directly accessible from the prompt X. This yields a general Reinforcement Learning objective: maximize task reward while compressing completions under a prior over reasoning traces, subsuming common heuristics (e.g., length penalties) as special cases (e.g., uniform priors). In contrast to naive token-counting-based approaches, we introduce a semantic prior that measures token cost by surprisal under a language model prior. Empirically, our CIB objective prunes cognitive bloat while preserving fluency and logic, improving accuracy at moderate compression and enabling aggressive compression with minimal accuracy drop.

We introduce Orthrus, a simple and efficient dual-architecture framework that unifies the exact generation fidelity of autoregressive Large Language Models (LLMs) with the high-speed parallel token generation of diffusion models. The sequential nature of standard autoregressive decoding represents a fundamental bottleneck for high-throughput inference. While diffusion language models attempt to break this barrier via parallel generation, they suffer from significant performance degradation, high training costs, and a lack of rigorous convergence guarantees. Orthrus resolves this dichotomy natively. Designed to seamlessly integrate into existing Transformers, the framework augments a frozen LLM with a lightweight, trainable module to create a parallel diffusion view alongside the standard autoregressive view. In this unified system, both views attend to the exact same high-fidelity Key-Value (KV) cache; the autoregressive head executes context pre-filling to construct accurate KV representations, while the diffusion head executes parallel generation. By employing an exact consensus mechanism between the two views, Orthrus guarantees lossless inference, delivering up to a 7.8x speedup with only an O(1) memory cache overhead and minimal parameter additions.

The efficiency of large vision-language models (LVLMs) is constrained by the computational bottleneck of the attention mechanism during the prefill phase and the memory bottleneck of fetching the key-value (KV) cache in the decoding phase, particularly in scenarios involving high-resolution images or videos. Visual content often exhibits substantial redundancy, resulting in highly sparse attention maps within LVLMs. This sparsity can be leveraged to accelerate attention computation or compress the KV cache through various approaches. However, most studies focus on addressing only one of these bottlenecks and do not adequately support dynamic adjustment of sparsity concerning distinct layers or tasks. In this paper, we present ZipVL, an efficient inference framework designed for LVLMs that resolves both computation and memory bottlenecks through a dynamic ratio allocation strategy of important tokens. This ratio is adaptively determined based on the layer-specific distribution of attention scores, rather than fixed hyper-parameters, thereby improving efficiency for less complex tasks while maintaining high performance for more challenging ones. Then we select important tokens based on their normalized attention scores and perform attention mechanism solely on those important tokens to accelerate the prefill phase. To mitigate the memory bottleneck in the decoding phase, we employ mixed-precision quantization to the KV cache, where high-bit quantization is used for caches of important tokens, while low-bit quantization is applied to those of less importance. Our experiments demonstrate that ZipVL can accelerate the prefill phase by 2.6times and reduce GPU memory usage by 50.0%, with a minimal accuracy reduction of only 0.2% on Video-MME benchmark over LongVA-7B model, effectively enhancing the generation efficiency of LVLMs.

Attention-based transformers have become the standard architecture in many deep learning fields, primarily due to their ability to model long-range dependencies and handle variable-length input sequences. However, the attention mechanism with its quadratic complexity is a significant bottleneck in the transformer architecture. This algorithm is only uni-directional in the decoder and converges to a static pattern in over-parametrized decoder-only models. I address this issue by developing a generative function as attention or activation replacement. It still has the auto-regressive character by comparing each token with the previous one. In my test setting with nanoGPT this yields a smaller loss while having a smaller model. The loss further drops by incorporating an average context vector. This concept of attention replacement is distributed under the GNU AGPL v3 license at https://gitlab.com/Bachstelze/causal_generation.

Transformer-based models have driven significant advancements in Multimodal Large Language Models (MLLMs), yet their computational costs surge drastically when scaling resolution, training data, and model parameters. A key bottleneck stems from the proliferation of visual tokens required for fine-grained image understanding. We propose Skip-Vision, a unified framework addressing both training and inference inefficiencies in vision-language models. On top of conventional token compression approaches, our method introduces two complementary acceleration strategies. For training acceleration, we observe that Feed-Forward Network (FFN) computations on visual tokens induce marginal feature updates. This motivates our Skip-FFN strategy, which bypasses FFN layers for redundant visual tokens. For inference acceleration, we design a selective KV-cache removal mechanism that prunes the skipped key-value pairs during decoding while preserving model performance. Experimental results demonstrate that Skip-Vision reduces training time by up to 35\%, inference FLOPs by 75\%, and latency by 45\%, while achieving comparable or superior performance to existing methods. Our work provides a practical solution for scaling high-performance MLLMs with enhanced efficiency.

Vision-Language Models (VLMs) face a bottleneck of prohibitive computational costs arising from massive visual token sequences during inference. Existing vision token reduction methods alleviate this burden, but they unintentionally preserve the isolated visual subject strictly aligned with the user's query, which fails to substantially explore salient subjects and their contextual relationships. In this paper, we propose SPpruner, a subject-centric progressive reduction paradigm that emulates the Focus-then-Context mechanism of the human visual perception system. Specifically, we first construct a focus identification module to explicitly model the interplay between visual saliency and semantic relevance. Herein, it can excavate the comprehensive visual subject spectrum to ensure a high-fidelity representation of visual input. Subsequently, a context-aware structural scanning module is developed to aggregate contextual cues from neighboring regions. As such, it can effectively restore global relational dependencies to uphold the structural integrity of the preserved subjects. Extensive experiments demonstrate that our paradigm consistently outperforms SOTA methods, achieving up to 2.53 times speedup with only 22.2% of visual tokens retained in Qwen2.5-VL and a 67% FLOPs reduction on LLaVA with a negligible 0.6% accuracy drop.

Multimodal large language models (MLLMs) have made remarkable strides, largely driven by their ability to process increasingly long and complex contexts, such as high-resolution images, extended video sequences, and lengthy audio input. While this ability significantly enhances MLLM capabilities, it introduces substantial computational challenges, primarily due to the quadratic complexity of self-attention mechanisms with numerous input tokens. To mitigate these bottlenecks, token compression has emerged as an auspicious and critical approach, efficiently reducing the number of tokens during both training and inference. In this paper, we present the first systematic survey and synthesis of the burgeoning field of multimodal long context token compression. Recognizing that effective compression strategies are deeply tied to the unique characteristics and redundancies of each modality, we categorize existing approaches by their primary data focus, enabling researchers to quickly access and learn methods tailored to their specific area of interest: (1) image-centric compression, which addresses spatial redundancy in visual data; (2) video-centric compression, which tackles spatio-temporal redundancy in dynamic sequences; and (3) audio-centric compression, which handles temporal and spectral redundancy in acoustic signals. Beyond this modality-driven categorization, we further dissect methods based on their underlying mechanisms, including transformation-based, similarity-based, attention-based, and query-based approaches. By providing a comprehensive and structured overview, this survey aims to consolidate current progress, identify key challenges, and inspire future research directions in this rapidly evolving domain. We also maintain a public repository to continuously track and update the latest advances in this promising area.

While long context support of large language models has extended their abilities, it also incurs challenges in memory and compute which becomes crucial bottlenecks in resource-restricted devices. Token eviction, a widely adopted post-training methodology designed to alleviate the bottlenecks by evicting less important tokens from the cache, typically uses attention scores as proxy metrics for token importance. However, one major limitation of attention score as a token-wise importance metrics is that it lacks the information about contribution of tokens to the attention output. In this paper, we propose a simple eviction criterion based on the contribution of cached tokens to attention outputs. Our method, CAOTE, optimizes for eviction error due to token eviction, by seamlessly integrating attention scores and value vectors. This is the first method which uses value vector information on top of attention-based eviction scores. Additionally, CAOTE can act as a meta-heuristic method with flexible usage with any token eviction method. We show that CAOTE, when combined with the state-of-the-art attention score-based methods, always improves accuracies on the downstream task, indicating the importance of leveraging information from values during token eviction process.

The inference of transformer-based large language models consists of two sequential stages: 1) a prefilling stage to compute the KV cache of prompts and generate the first token, and 2) a decoding stage to generate subsequent tokens. For long prompts, the KV cache must be computed for all tokens during the prefilling stage, which can significantly increase the time needed to generate the first token. Consequently, the prefilling stage may become a bottleneck in the generation process. An open question remains whether all prompt tokens are essential for generating the first token. To answer this, we introduce a novel method, LazyLLM, that selectively computes the KV for tokens important for the next token prediction in both the prefilling and decoding stages. Contrary to static pruning approaches that prune the prompt at once, LazyLLM allows language models to dynamically select different subsets of tokens from the context in different generation steps, even though they might be pruned in previous steps. Extensive experiments on standard datasets across various tasks demonstrate that LazyLLM is a generic method that can be seamlessly integrated with existing language models to significantly accelerate the generation without fine-tuning. For instance, in the multi-document question-answering task, LazyLLM accelerates the prefilling stage of the LLama 2 7B model by 2.34x while maintaining accuracy.

Large language models (LLMs) are typically trained via next-word prediction (NWP), which provides strong surface-level fluency but often lacks support for robust reasoning. We propose BOttlenecked next Word exploration (BOW), a novel RL framework that rethinks NWP by introducing a reasoning bottleneck where a policy model first generates a reasoning path rather than predicting the next token directly, after which a frozen judge model predicts the next token distribution based solely on this reasoning path. We train the policy model using GRPO with rewards that quantify how effectively the reasoning path facilitates next-word recovery. Compared with other continual pretraining baselines, we show that BOW improves both the general and next-word reasoning capabilities of the base model, evaluated on various benchmarks. Our findings show that BOW can serve as an effective and scalable alternative to vanilla NWP.

Large Language Models (LLMs) rely on the Key-Value (KV) Cache to store token history, enabling efficient decoding of tokens. As the KV-Cache grows, it becomes a major memory and computation bottleneck, however, there is an opportunity to alleviate this bottleneck, especially because prior research has shown that only a small subset of tokens contribute meaningfully to each decoding step. A key challenge in finding these critical tokens is that they are dynamic, and heavily input query-dependent. Existing methods either risk quality by evicting tokens permanently, or retain the full KV-Cache but rely on retrieving chunks (pages) of tokens at generation, failing at dense, context-rich tasks. Additionally, many existing KV-Cache sparsity methods rely on inaccurate proxies for token importance. To address these limitations, we introduce TokenButler, a high-granularity, query-aware predictor that learns to identify these critical tokens. By training a light-weight predictor with less than 1.2% parameter overhead, TokenButler prioritizes tokens based on their contextual, predicted importance. This improves perplexity & downstream accuracy by over 8% relative to SoTA methods for estimating token importance. We evaluate TokenButler on a novel synthetic small-context co-referential retrieval task, demonstrating near-oracle accuracy. Code, models and benchmarks: https://github.com/abdelfattah-lab/TokenButler

Key-Value (KV) Caching has become an essential technique for accelerating the inference speed and throughput of generative Large Language Models~(LLMs). However, the memory footprint of the KV cache poses a critical bottleneck in LLM deployment as the cache size grows with batch size and sequence length, often surpassing even the size of the model itself. Although recent methods were proposed to select and evict unimportant KV pairs from the cache to reduce memory consumption, the potential ramifications of eviction on the generative process are yet to be thoroughly examined. In this paper, we examine the detrimental impact of cache eviction and observe that unforeseen risks arise as the information contained in the KV pairs is exhaustively discarded, resulting in safety breaches, hallucinations, and context loss. Surprisingly, we find that preserving even a small amount of information contained in the evicted KV pairs via reduced precision quantization substantially recovers the incurred degradation. On the other hand, we observe that the important KV pairs must be kept at a relatively higher precision to safeguard the generation quality. Motivated by these observations, we propose Mixed-precision KV cache~(MiKV), a reliable cache compression method that simultaneously preserves the context details by retaining the evicted KV pairs in low-precision and ensure generation quality by keeping the important KV pairs in high-precision. Experiments on diverse benchmarks and LLM backbones show that our proposed method offers a state-of-the-art trade-off between compression ratio and performance, compared to other baselines.

Adapting decoder-only multimodal large language models (MLLMs) for unified multimodal retrieval faces two structural gaps. First, existing methods rely on implicit pooling, which overloads the hidden state of a standard vocabulary token (e.g., <EOS>) as the sequence-level representation, a mechanism never designed for information aggregation. Second, contrastive fine-tuning specifies what the embedding should match but provides no token-level guidance on how information should be compressed into it. We address both gaps with two complementary components. Architecturally, we introduce Bottleneck Tokens (BToks), a small set of learnable tokens that serve as a fixed-capacity explicit pooling mechanism. For training, we propose Generative Information Condensation: a next-token prediction objective coupled with a Condensation Mask that severs the direct attention path from target tokens to query tokens. All predictive signals are thereby forced through the BToks, converting the generative loss into dense, token-level supervision for semantic compression. At inference time, only the input and BToks are processed in a single forward pass with negligible overhead over conventional last-token pooling. On MMEB-V2 (78 datasets, 3 modalities, 9 meta-tasks), our approach achieves state-of-the-art among 2B-scale methods under comparable data conditions, attaining an Overall score of 59.0 (+3.6 over VLM2Vec-V2) with substantial gains on semantically demanding tasks (e.g., +12.6 on Video-QA).

The quadratic computational complexity of softmax transformers has become a bottleneck in long-context scenarios. In contrast, linear attention model families provide a promising direction towards a more efficient sequential model. These linear attention models compress past KV values into a single hidden state, thereby efficiently reducing complexity during both training and inference. However, their expressivity remains limited by the size of their hidden state. Previous work proposed interleaving softmax and linear attention layers to reduce computational complexity while preserving expressivity. Nevertheless, the efficiency of these models remains bottlenecked by their softmax attention layers. In this paper, we propose Neural Attention Search Linear (NAtS-L), a framework that applies both linear attention and softmax attention operations within the same layer on different tokens. NAtS-L automatically determines whether a token can be handled by a linear attention model, i.e., tokens that have only short-term impact and can be encoded into fixed-size hidden states, or require softmax attention, i.e., tokens that contain information related to long-term retrieval and need to be preserved for future queries. By searching for optimal Gated DeltaNet and softmax attention combinations across tokens, we show that NAtS-L provides a strong yet efficient token-level hybrid architecture.

Self-attention and transformers have been widely used in deep learning. Recent efforts have been devoted to incorporating transformer blocks into different neural architectures, including those with convolutions, leading to various visual transformers for computer vision tasks. In this paper, we propose a novel and compact transformer block, Transformer with Learnable Token Merging (LTM), or LTM-Transformer. LTM-Transformer performs token merging in a learnable scheme. LTM-Transformer is compatible with many popular and compact transformer networks, and it reduces the FLOPs and the inference time of the visual transformers while maintaining or even improving the prediction accuracy. In the experiments, we replace all the transformer blocks in popular visual transformers, including MobileViT, EfficientViT, ViT, and Swin, with LTM-Transformer blocks, leading to LTM-Transformer networks with different backbones. The LTM-Transformer is motivated by reduction of Information Bottleneck, and a novel and separable variational upper bound for the IB loss is derived. The architecture of the mask module in our LTM blocks, which generates the token merging mask, is designed to reduce the derived upper bound for the IB loss. Extensive results on computer vision tasks evidence that LTM-Transformer renders compact and efficient visual transformers with comparable or much better prediction accuracy than the original visual transformers. The code of the LTM-Transformer is available at https://github.com/Statistical-Deep-Learning/LTM}

Despite the recent success in many applications, the high computational requirements of vision transformers limit their use in resource-constrained settings. While many existing methods improve the quadratic complexity of attention, in most vision transformers, self-attention is not the major computation bottleneck, e.g., more than 80% of the computation is spent on fully-connected layers. To improve the computational complexity of all layers, we propose a novel token downsampling method, called Token Pooling, efficiently exploiting redundancies in the images and intermediate token representations. We show that, under mild assumptions, softmax-attention acts as a high-dimensional low-pass (smoothing) filter. Thus, its output contains redundancy that can be pruned to achieve a better trade-off between the computational cost and accuracy. Our new technique accurately approximates a set of tokens by minimizing the reconstruction error caused by downsampling. We solve this optimization problem via cost-efficient clustering. We rigorously analyze and compare to prior downsampling methods. Our experiments show that Token Pooling significantly improves the cost-accuracy trade-off over the state-of-the-art downsampling. Token Pooling is a simple and effective operator that can benefit many architectures. Applied to DeiT, it achieves the same ImageNet top-1 accuracy using 42% fewer computations.

The attention mechanism is becoming increasingly popular in Natural Language Processing (NLP) applications, showing superior performance than convolutional and recurrent architectures. However, attention becomes the compution bottleneck because of its quadratic computational complexity to input length, complicated data movement and low arithmetic intensity. Moreover, existing NN accelerators mainly focus on optimizing convolutional or recurrent models, and cannot efficiently support attention. In this paper, we present SpAtten, an efficient algorithm-architecture co-design that leverages token sparsity, head sparsity, and quantization opportunities to reduce the attention computation and memory access. Inspired by the high redundancy of human languages, we propose the novel cascade token pruning to prune away unimportant tokens in the sentence. We also propose cascade head pruning to remove unessential heads. Cascade pruning is fundamentally different from weight pruning since there is no trainable weight in the attention mechanism, and the pruned tokens and heads are selected on the fly. To efficiently support them on hardware, we design a novel top-k engine to rank token and head importance scores with high throughput. Furthermore, we propose progressive quantization that first fetches MSBs only and performs the computation; if the confidence is low, it fetches LSBs and recomputes the attention outputs, trading computation for memory reduction. Extensive experiments on 30 benchmarks show that, on average, SpAtten reduces DRAM access by 10.0x with no accuracy loss, and achieves 1.6x, 3.0x, 162x, 347x speedup, and 1,4x, 3.2x, 1193x, 4059x energy savings over A3 accelerator, MNNFast accelerator, TITAN Xp GPU, Xeon CPU, respectively.

As LLM agents evolve, tokens have emerged as the core economic primitives of Agentic AI. However, their exponential consumption introduces severe computational, collaborative, and security bottlenecks. Current surveys remain fragmented across system optimization, architecture design, and trust, lacking a unified framework to evaluate the fundamental trade-off between output quality and economic cost. To bridge this gap, this survey presents the first comprehensive survey of Token Economics. By unifying computer science and economics, we conceptualize tokens as production factors, exchange mediums, and units of account. We synthesize existing literature across a four-dimensional taxonomy: (1) Micro-level (Single Agent): Optimizing budget-constrained factor substitution via neoclassical firm theory. (2) Meso-level (Multi-Agent Systems): Minimizing collaboration friction using transaction cost and principal-agent theories. (3) Macro-level (Agent Ecosystems): Addressing congestion externalities and pricing via mechanism design. (4) Security: Internalizing adversarial threats as endogenous economic constraints. Finally, we outline frontier directions, including differentiable token budgets and dynamic markets, to lay the theoretical foundation for scalable next-generation agent systems.

Audio-Visual Large Language Models (AV-LLMs) face prohibitive computational overhead from massive audio and video tokens. Token reduction, while extensively explored for video-only LLMs, is insufficient for the audio-visual domain, as these unimodal methods cannot leverage audio-visual cross-modal synergies. Furthermore, the distinct and dynamic information densities of audio and video render static budgets per modality suboptimal. How to perform token reduction on a joint audio-visual stream thus remains an unaddressed bottleneck. To fill this gap, we introduce EchoingPixels, a framework inspired by the coexistence and interaction of visuals and sound in real-world scenes. The core of our framework is the Cross-Modal Semantic Sieve (CS2), a module enabling early audio-visual interaction. Instead of compressing modalities independently, CS2 co-attends to the joint multimodal stream and reduces tokens from an entire combined pool of audio-visual tokens rather than using fixed budgets per modality. This single-pool approach allows it to adaptively allocate the token budget across both modalities and dynamically identify salient tokens in concert. To ensure this aggressive reduction preserves the vital temporal modeling capability, we co-design a Synchronization-Augmented RoPE (Sync-RoPE) to maintain critical temporal relationships for the sparsely selected tokens. Extensive experiments demonstrate that EchoingPixels achieves performance comparable to strong baselines using only 5-20% of the original tokens, with a 2-3x speedup and memory reduction.

High-quality main content extraction from web pages is a critical prerequisite for constructing large-scale training corpora. While traditional heuristic extractors are efficient, they lack the semantic reasoning required to handle the structural heterogeneity of the modern web. Conversely, well-pretrained generative Large Language Models (LLMs) offer superior document comprehension but are prohibited by excessive computational costs, limited context windows, and hallucination risks when applied at web scale. We present Dripper, a lightweight framework that resolves these bottlenecks through four contributions: (1) We reformulate extraction as a constrained sequence labeling task using SLMs (Small Language Models). This paradigm eliminates generative hallucinations and achieves exceptional efficiency, reaching a throughput of 3.08 pages per second on a single A100 GPU. (2) We construct WebMainBench, a rigorous benchmark of 7,809 human-annotated pages covering 5,434 unique domains and multiple languages. Evaluations show our Dripper-0.6B model outperforms heuristics like Trafilatura and rivals massive models like DeepSeek-V3.2(685B), GPT-5 and Gemini-2.5-Pro, offering an optimal efficiency-accuracy trade-off. (3) We demonstrate infrastructural value by pre-training a 1B model on a Dripper-curated corpus (63B tokens). This model significantly outperforms baselines in downstream tasks, proving the critical role of extraction quality and the effectiveness of our framework. (4) We open-source the Dripper-0.6B weights and codebase to facilitate the construction of high-quality datasets.

Large Vision-Language Models (LVLMs) enable sophisticated reasoning over images and videos, yet their inference is hindered by a systemic efficiency barrier known as visual token dominance. This overhead is driven by a multi-regime interplay between high-resolution feature extraction, quadratic attention scaling, and memory bandwidth constraints. We present a systematic taxonomy of efficiency techniques structured around the inference lifecycle, consisting of encoding, prefilling, and decoding. Unlike prior reviews focused on isolated optimizations, we analyze the end-to-end pipeline to reveal how upstream decisions dictate downstream bottlenecks, covering compute-bound visual encoding, the intensive prefilling of massive contexts, and the ''visual memory wall'' in bandwidth-bound decoding. By decoupling the efficiency landscape into the axes of shaping information density, managing long-context attention, and overcoming memory limits, this work provides a structured analysis of how isolated optimizations compose to navigate the trade-off between visual fidelity and system efficiency. The survey concludes by outlining four future frontiers supported by pilot empirical insights, including hybrid compression based on functional unit sensitivity, modality-aware decoding with relaxed verification, progressive state management for streaming continuity, and stage-disaggregated serving through hardware-algorithm co-design. Our literature repository is at https://github.com/SuDIS-ZJU/Efficient-LVLMs-Inference.

Multimodal Large Language Models (MLLMs) incur significant computational cost from processing numerous vision tokens through all LLM layers. Prior pruning methods operate either before the LLM, limiting generality due to diverse encoder-projector designs or within the LLM using heuristics that are incompatible with FlashAttention. We take a different approach: rather than identifying unimportant tokens, we treat the LLM itself as the optimal guide for compression. Observing that deeper layers naturally transmit vision-to-text information, we introduce Attention-Driven Self-Compression (ADSC), a simple, broadly applicable method that progressively reduces vision tokens using only the LLM's attention mechanism. Our method applies uniform token downsampling at selected layers, forming bottlenecks that encourage the model to reorganize and compress information into the remaining tokens. It requires no score computation, auxiliary modules, or attention modification, and remains fully compatible with FlashAttention. Applied to LLaVA-1.5, ADSC reduces FLOPs by 53.7% and peak KV-cache memory by 56.7%, while preserving 98.2% of the original model performance. Across multiple benchmarks, it outperforms prior pruning approaches in both efficiency and accuracy. Crucially, under high compression ratios, our method remains robust while heuristic-based techniques degrade sharply.

Training vision-language models on cognitively-plausible amounts of data requires rethinking how models integrate multimodal information. Within the constraints of the Vision track for the BabyLM Challenge 2025, we propose a lightweight decoder-based architecture with (1) token-wise dynamic gating for adaptive fusion of linguistic and visual cues, (2) feature modulation and channel attention to maximise the utility of limited visual information and (3) auxiliary contrastive objectives for visual grounding. Evaluation on five benchmarks (BLiMP, BLiMP Supplement, EWoK, Winoground and VQA) shows competitive or superior performance to multimodal baselines. More notably, our dynamic gate discovers interpretable patterns without explicit supervision, favouring visual cues for content words and linguistic cues for function words. While we identify limitations in the Challenge constraints, such as the information bottleneck created by global image embeddings and training instability from the dataset split, our findings establish dynamic gating as a powerful tool for efficient multimodal learning, offering both interpretability and performance even under severe constraints.

The application of large vision-language models to computational pathology holds great promise for diagnostic assistants but faces a critical computational bottleneck: the gigapixel scale of Whole Slide Images (WSIs). A single WSI typically contains over 105 patches, creating sequence lengths that exceed the constraints of standard Transformer architectures. Existing solutions often resort to spatial sampling, which risks discarding diagnostically critical evidence. To address this, we propose TC-SSA (Token Compression via Semantic Slot Aggregation), a learnable token compression framework that aggregates patch features into a fixed number of semantic slots. A gated routing module assigns patches to slots using sparse Top-2 routing, followed by weighted aggregation, enabling global slide coverage under a strict token budget. The resulting representation retains diagnostically relevant information while reducing the number of visual tokens to 1.7% of the original sequence. On the SlideBench(TCGA), our model achieves 78.34% overall accuracy and 77.14% on the diagnosis subset, outperforming sampling-based baselines under comparable token budgets. The method also generalizes to MIL classification, reaching AUC of 95.83% on TCGA-BRCA, 98.27% on TCGA-NSCLC and 79.80% on PANDA. These results suggest that learnable semantic aggregation provides an effective trade-off between efficiency and diagnostic performance for gigapixel pathology reasoning.

Large-scale video-language pre-training has made remarkable strides in advancing video-language understanding tasks. However, the heavy computational burden of video encoding remains a formidable efficiency bottleneck, particularly for long-form videos. These videos contain massive visual tokens due to their inherent 3D properties and spatiotemporal redundancy, making it challenging to capture complex temporal and spatial relationships. To tackle this issue, we propose an efficient method called TEmporal-Spatial Token Aggregation (TESTA). TESTA condenses video semantics by adaptively aggregating similar frames, as well as similar patches within each frame. TESTA can reduce the number of visual tokens by 75% and thus accelerate video encoding. Building upon TESTA, we introduce a pre-trained video-language model equipped with a divided space-time token aggregation module in each video encoder block. We evaluate our model on five datasets for paragraph-to-video retrieval and long-form VideoQA tasks. Experimental results show that TESTA improves computing efficiency by 1.7 times, and achieves significant performance gains from its scalability in processing longer input frames, e.g., +13.7 R@1 on QuerYD and +6.5 R@1 on Condensed Movie.

Inference-time scaling via chain-of-thought (CoT) reasoning is a major driver of state-of-the-art LLM performance, but it comes with substantial latency and compute costs. We address a fundamental theoretical question: how many reasoning tokens are required to solve a problem as input size grows? By extending the bounded attention prefix oracle (BAPO) model--an abstraction of LLMs that quantifies the information flow required to solve a task--we prove lower bounds on the CoT tokens required for three canonical BAPO-hard tasks: binary majority, triplet matching, and graph reachability. We show that each requires Ω(n) reasoning tokens when the input size is n. We complement these results with matching or near-matching upper bounds via explicit constructions. Finally, our experiments with frontier reasoning models show approximately linear reasoning token scaling on these tasks and failures when constrained to smaller reasoning budgets, consistent with our theoretical lower bounds. Together, our results identify fundamental bottlenecks in inference-time compute through CoT and offer a principled tool for analyzing optimal reasoning length.

Medical image segmentation is vital for clinical diagnosis and quantitative analysis, yet remains challenging due to the heterogeneity of imaging modalities and the high cost of pixel-level annotations. Although general interactive segmentation models like SAM have achieved remarkable progress, their transfer to medical imaging still faces two key bottlenecks: (i) the lack of adaptive mechanisms for modality- and anatomy-specific tasks, which limits generalization in out-of-distribution medical scenarios; and (ii) current medical adaptation methods fine-tune on large, heterogeneous datasets without selection, leading to noisy supervision, higher cost, and negative transfer. To address these issues, we propose SegMoTE, an efficient and adaptive framework for medical image segmentation. SegMoTE preserves SAM's original prompt interface, efficient inference, and zero-shot generalization while introducing only a small number of learnable parameters to dynamically adapt across modalities and tasks. In addition, we design a progressive prompt tokenization mechanism that enables fully automatic segmentation, significantly reducing annotation dependence. Trained on MedSeg-HQ, a curated dataset less than 1% of existing large-scale datasets, SegMoTE achieves SOTA performance across diverse imaging modalities and anatomical tasks. It represents the first efficient, robust, and scalable adaptation of general segmentation models to the medical domain under extremely low annotation cost, advancing the practical deployment of foundation vision models in clinical applications.

The Visual Geometry Grounded Transformer (VGGT) marks a significant leap forward in 3D scene reconstruction, as it is the first model that directly infers all key 3D attributes (camera poses, depths, and dense geometry) jointly in one pass. However, this joint inference mechanism requires global attention layers that perform all-to-all attention computation on tokens from all views. For reconstruction of large scenes with long-sequence inputs, this causes a significant latency bottleneck. In this paper, we propose head-wise temporal merging (HTTM), a training-free 3D token merging method for accelerating VGGT. Existing merging techniques merge tokens uniformly across different attention heads, resulting in identical tokens in the layers' output, which hinders the model's representational ability. HTTM tackles this problem by merging tokens in multi-head granularity, which preserves the uniqueness of feature tokens after head concatenation. Additionally, this enables HTTM to leverage the spatial locality and temporal correspondence observed at the head level to achieve higher merging ratios with lower merging costs compared to existing methods. Thus, HTTM achieves up to 7x acceleration with negligible performance drops in a GPU-based inference.

Large Language Models (LLMs) have demonstrated remarkable capabilities in handling long context inputs, but this comes at the cost of increased computational resources and latency. Our research introduces a novel approach for the long context bottleneck to accelerate LLM inference and reduce GPU memory consumption. Our research demonstrates that LLMs can identify relevant tokens in the early layers before generating answers to a query. Leveraging this insight, we propose an algorithm that uses early layers of an LLM as filters to select and compress input tokens, significantly reducing the context length for subsequent processing. Our method, GemFilter, demonstrates substantial improvements in both speed and memory efficiency compared to existing techniques, such as standard attention and SnapKV/H2O. Notably, it achieves a 2.4times speedup and 30\% reduction in GPU memory usage compared to SOTA methods. Evaluation on the Needle in a Haystack task shows that GemFilter significantly outperforms standard attention, SnapKV and demonstrates comparable performance on the LongBench challenge. GemFilter is simple, training-free, and broadly applicable across different LLMs. Crucially, it provides interpretability by allowing humans to inspect the selected input sequence. These findings not only offer practical benefits for LLM deployment, but also enhance our understanding of LLM internal mechanisms, paving the way for further optimizations in LLM design and inference. Our code is available at https://github.com/SalesforceAIResearch/GemFilter.

Omnimodal large language models (OmniLLMs) have attracted increasing research attention of late towards unified audio-video understanding, wherein processing audio-video token sequences creates a significant computational bottleneck, however. Existing token compression methods have yet to accommodate this emerging need of jointly compressing multimodal tokens. To bridge this gap, we present OmniZip, a training-free, audio-guided audio-visual token-compression framework that optimizes multimodal token representation and accelerates inference. Specifically, OmniZip first identifies salient audio tokens, then computes an audio retention score for each time group to capture information density, thereby dynamically guiding video token pruning and preserving cues from audio anchors enhanced by cross-modal similarity. For each time window, OmniZip compresses the video tokens using an interleaved spatio-temporal scheme. Extensive empirical results demonstrate the merits of OmniZip - it achieves 3.42X inference speedup and 1.4X memory reduction over other top-performing counterparts, while maintaining performance with no training.

Streaming Video Large Language Models (VideoLLMs) have demonstrated impressive performance across various video understanding tasks, but they face significant challenges in real-time deployment due to the high computational cost of processing dense visual tokens from continuous video streams. In streaming video scenarios, the primary bottleneck lies in the Vision Transformer (ViT) encoding stage, where redundant processing of temporally similar frames leads to inefficiency. Additionally, inflated token sequences during LLM pre-filling further exacerbate latency and memory overhead. To address these challenges, we propose Streaming Token Compression (STC), a plug-and-play hierarchical framework that seamlessly integrates into existing streaming VideoLLMs, optimizing both ViT encoding and LLM pre-filling stages to accelerate processing. STC introduces two token-level accelerators: STC-Cacher, which reduces ViT encoding overhead by caching and reusing features from temporally similar frames, and STC-Pruner, which compresses the visual token sequence before it enters the LLM, preserving only the most salient tokens based on both spatial and temporal relevance. Extensive experiments on four baseline streaming VideoLLMs across five benchmarks demonstrate that STC outperforms other compression methods. Notably, STC retains up to 99\% of accuracy on the ReKV framework while reducing ViT encoding latency and LLM pre-filling latency by 24.5\% and 45.3\%.

Document parsing, as a fundamental yet crucial vision task, is being revolutionized by vision-language models (VLMs). However, the autoregressive (AR) decoding inherent to VLMs creates a significant bottleneck, severely limiting parsing speed. In this paper, we propose Parallel-Token Prediction (PTP), a plugable, model-agnostic and simple-yet-effective method that enables VLMs to generate multiple future tokens in parallel with improved sample efficiency. Specifically, we insert some learnable tokens into the input sequence and design corresponding training objectives to equip the model with parallel decoding capabilities for document parsing. Furthermore, to support effective training, we develop a comprehensive data generation pipeline that efficiently produces large-scale, high-quality document parsing training data for VLMs. Extensive experiments on OmniDocBench and olmOCR-bench demonstrate that our method not only significantly improves decoding speed (1.6x-2.2x) but also reduces model hallucinations and exhibits strong generalization abilities.

Modern BPE tokenizers often split calendar dates into meaningless fragments, e.g., 20250312 rightarrow 202, 503, 12, inflating token counts and obscuring the inherent structure needed for robust temporal reasoning. In this work, we (1) introduce a simple yet interpretable metric, termed date fragmentation ratio, that measures how faithfully a tokenizer preserves multi-digit date components; (2) release DateAugBench, a suite of 6500 examples spanning three temporal reasoning tasks: context-based date resolution, format-invariance puzzles, and date arithmetic across historical, contemporary, and future regimes; and (3) through layer-wise probing and causal attention-hop analyses, uncover an emergent date-abstraction mechanism whereby large language models stitch together the fragments of month, day, and year components for temporal reasoning. Our experiments show that excessive fragmentation correlates with accuracy drops of up to 10 points on uncommon dates like historical and futuristic dates. Further, we find that the larger the model, the faster the emergent date abstraction that heals date fragments is accomplished. Lastly, we observe a reasoning path that LLMs follow to assemble date fragments, typically differing from human interpretation (year rightarrow month rightarrow day).

Large language models (LLMs) have achieved remarkable progress across diverse tasks, yet their internal mechanisms remain largely opaque. In this work, we address a fundamental question: to what extent can the original input text be recovered from a single last-token representation within an LLM? We propose Rep2Text, a novel framework for decoding full text from last-token representations. Rep2Text employs a trainable adapter that projects a target model's internal representations into the embedding space of a decoding language model, which then autoregressively reconstructs the input text. Experiments on various model combinations (Llama-3.1-8B, Gemma-7B, Mistral-7B-v0.1, Llama-3.2-3B) demonstrate that, on average, over half of the information in 16-token sequences can be recovered from this compressed representation while maintaining strong semantic integrity and coherence. Furthermore, our analysis reveals an information bottleneck effect: longer sequences exhibit decreased token-level recovery while preserving strong semantic integrity. Besides, our framework also demonstrates robust generalization to out-of-distribution medical data.

Inference accounts for the majority of latency and energy consumption in large language model (LLM) deployments, often exceeding 90% of total cost. While training-time efficiency has seen extensive progress, runtime optimization remains a key bottleneck, particularly under autoregressive decoding. Existing approaches -- such as pruning, quantization, early exits, and speculative decoding -- often require retraining, architectural changes, or disrupt decoding compatibility. We introduce QuickSilver, a modular, token-level framework that enables semantic adaptivity at inference time without altering model weights or structure. QuickSilver integrates four synergistic mechanisms: (i) Dynamic Token Halting, which halts computation for tokens with converged representations; (ii) KV Cache Skipping, which selectively suppresses memory writes to reduce attention overhead; and (iii) Contextual Token Fusion, which collapses redundant tokens into shared paths to shrink sequence length. Unlike speculative decoding or MoE routing, QuickSilver operates entirely on frozen, dense models and requires no auxiliary networks. Applied to GPT-2 and Llama-2 across WikiText-103 and C4, QuickSilver achieves up to 39.6% FLOP reduction with negligible perplexity degradation (<=0.2).

Large multilingual language models typically rely on a single vocabulary shared across 100+ languages. As these models have increased in parameter count and depth, vocabulary size has remained largely unchanged. This vocabulary bottleneck limits the representational capabilities of multilingual models like XLM-R. In this paper, we introduce a new approach for scaling to very large multilingual vocabularies by de-emphasizing token sharing between languages with little lexical overlap and assigning vocabulary capacity to achieve sufficient coverage for each individual language. Tokenizations using our vocabulary are typically more semantically meaningful and shorter compared to XLM-R. Leveraging this improved vocabulary, we train XLM-V, a multilingual language model with a one million token vocabulary. XLM-V outperforms XLM-R on every task we tested on ranging from natural language inference (XNLI), question answering (MLQA, XQuAD, TyDiQA), and named entity recognition (WikiAnn) to low-resource tasks (Americas NLI, MasakhaNER).

Large language models (LLMs) excel at generation but dominant autoregressive (AR) decoding is inherently sequential, creating a throughput bottleneck. Diffusion Language Models (DLMs)--especially block-wise variants--enable parallel generation and intra-block bidirectional reasoning, yet training large DLMs from scratch is costly and wastes the knowledge in mature AR checkpoints. Prior "adaptation" attempts either modify logits or randomly grow attention masks to full-sequence diffusion, or simply transplant AR weights into a block-diffusion recipe, leaving a fundamental mismatch between AR causality and block-wise bidirectionality unaddressed. We reframe adaptation as a intra-paradigm path from AR to Block-Diffusion by viewing AR as Block-Diffusion with blocksize=1. Concretely, we design the pathway of adaptation as follows: we use a context-causal attention mask (causal in context, bidirectional only within the active block), an efficient parallel adaptation procedure, an auxiliary AR loss to maximize data utilization and retain pretrained knowledge, and gradual increment of the generation block size. The recipe integrates cleanly with masked block-diffusion and maintains train-inference consistency. Built on these components, NBDiff-7B (Base and Instruct) could inherit the long-context modeling and reasoning capabilities, and achieve state-of-the-art performance among the 7B-class DLMs, delivering strong gains on general-knowledge, math, and code benchmarks over strong baselines. These results demonstrate that principled AR-to-block-diffusion adaptation is an effective and compute-efficient alternative to training DLMs from scratch. Codes: https://github.com/YuchuanTian/NBDiff.

Structured State Space Models (SSMs) have emerged as alternatives to transformers. While SSMs are often regarded as effective in capturing long-sequence dependencies, we rigorously demonstrate that they are inherently limited by strong recency bias. Our empirical studies also reveal that this bias impairs the models' ability to recall distant information and introduces robustness issues. Our scaling experiments then discovered that deeper structures in SSMs can facilitate the learning of long contexts. However, subsequent theoretical analysis reveals that as SSMs increase in depth, they exhibit another inevitable tendency toward over-smoothing, e.g., token representations becoming increasingly indistinguishable. This fundamental dilemma between recency and over-smoothing hinders the scalability of existing SSMs. Inspired by our theoretical findings, we propose to polarize two channels of the state transition matrices in SSMs, setting them to zero and one, respectively, simultaneously addressing recency bias and over-smoothing. Experiments demonstrate that our polarization technique consistently enhances the associative recall accuracy of long-range tokens and unlocks SSMs to benefit further from deeper architectures. All source codes are released at https://github.com/VITA-Group/SSM-Bottleneck.

Token attribution methods provide intuitive explanations for language model outputs by identifying causally important input tokens. However, as modern LLMs increasingly rely on extended reasoning chains, existing schemes face two critical challenges: (1) efficiency bottleneck, where attributing a target span of M tokens within a context of length N requires O(M*N) operations, making long-context attribution prohibitively slow; and (2) faithfulness drop, where intermediate reasoning tokens absorb attribution mass, preventing importance from propagating back to the original input. To address these, we introduce FlashTrace, an efficient multi-token attribution method that employs span-wise aggregation to compute attribution over multi-token targets in a single pass, while maintaining faithfulness. Moreover, we design a recursive attribution mechanism that traces importance through intermediate reasoning chains back to source inputs. Extensive experiments on long-context retrieval (RULER) and multi-step reasoning (MATH, MorehopQA) tasks demonstrate that FlashTrace achieves over 130x speedup over existing baselines while maintaining superior faithfulness. We further analyze the dynamics of recursive attribution, showing that even a single recursive hop improves faithfulness by tracing importance through the reasoning chain.

Multimodal Large Language Models (MLLMs) have demonstrated impressive capabilities, yet they encounter significant computational bottlenecks due to the massive volume of visual tokens. Consequently, visual token pruning, which substantially reduces the token count, has emerged as a critical technique for accelerating MLLM inference. Existing approaches focus on token importance, diversity, or an intuitive combination of both, without a principled framework for their optimal integration. To address this issue, we first conduct a systematic analysis to characterize the trade-off between token importance and semantic diversity. Guided by this analysis, we propose the Importance and Diversity Pruner (IDPruner), which leverages the Maximal Marginal Relevance (MMR) algorithm to achieve a Pareto-optimal balance between these two objectives. Crucially, our method operates without requiring attention maps, ensuring full compatibility with FlashAttention and efficient deployment via one-shot pruning. We conduct extensive experiments across various model architectures and multimodal benchmarks, demonstrating that IDPruner achieves state-of-the-art performance and superior generalization across diverse architectures and tasks. Notably, on Qwen2.5-VL-7B-Instruct, IDPruner retains 95.18\% of baseline performance when pruning 75\% of the tokens, and still maintains 86.40\% even under an extreme 90\% pruning ratio. Our code is available at https://github.com/Tencent/AngelSlim.

Multimodal Large Language Models (MLLMs) combine visual and textual representations to enable rich reasoning capabilities. However, the high computational cost of processing dense visual tokens remains a major bottleneck. A critical component in this pipeline is the visual projector, which bridges the vision encoder and the language model. Standard designs often employ a simple multi-layer perceptron for direct token mapping, but this approach scales poorly with high-resolution inputs, introducing significant redundancy. We present Delta-LLaVA, a token-efficient projector that employs a low-rank DeltaProjection to align multi-level vision features into a compact subspace before further interaction. On top of this base alignment, lightweight Transformer blocks act as specialization layers, capturing both global and local structure under constrained token budgets. Extensive experiments and ablations demonstrate that this base-then-specialize design yields consistent gains across multiple benchmarks with only 144 tokens, highlighting the importance of token formation prior to scaling interaction capacity. With Delta-LLaVA, inference throughput improves by up to 55%, while end-to-end training accelerates by nearly 4-5x in pretraining and over 1.5x in finetuning, highlighting the dual benefits of our design in both efficiency and scalability.

As large language models (LLMs) tackle increasingly complex tasks and longer documents, their computational and memory costs during inference become a major bottleneck. To address this, we propose PromptDistill, a novel, training-free method that improves inference efficiency while preserving generation quality. PromptDistill identifies and retains the most informative tokens by leveraging attention interactions in early layers, preserving their hidden states while reducing the computational burden in later layers. This allows the model to focus on essential contextual information without fully processing all tokens. Unlike previous methods such as H2O and SnapKV, which perform compression only after processing the entire input, or GemFilter, which selects a fixed portion of the initial prompt without considering contextual dependencies, PromptDistill dynamically allocates computational resources to the most relevant tokens while maintaining a global awareness of the input. Experiments using our method and baseline approaches with base models such as LLaMA 3.1 8B Instruct, Phi 3.5 Mini Instruct, and Qwen2 7B Instruct on benchmarks including LongBench, InfBench, and Needle in a Haystack demonstrate that PromptDistill significantly improves efficiency while having minimal impact on output quality compared to the original models. With a single-stage selection strategy, PromptDistill effectively balances performance and efficiency, outperforming prior methods like GemFilter, H2O, and SnapKV due to its superior ability to retain essential information. Specifically, compared to GemFilter, PromptDistill achieves an overall 1% to 5% performance improvement while also offering better time efficiency. Additionally, we explore multi-stage selection, which further improves efficiency while maintaining strong generation performance.

Large Language Models (LLMs) based autonomous agents demonstrate multifaceted capabilities to contribute substantially to economic production. However, existing benchmarks remain focused on single agentic capability, failing to capture long-horizon real-world scenarios. Moreover, the reliance on human-in-the-loop feedback for realistic tasks creates a scalability bottleneck, hindering automated rollout collection and evaluation. To bridge this gap, we introduce AgencyBench, a comprehensive benchmark derived from daily AI usage, evaluating 6 core agentic capabilities across 32 real-world scenarios, comprising 138 tasks with specific queries, deliverables, and rubrics. These scenarios require an average of 90 tool calls, 1 million tokens, and hours of execution time to resolve. To enable automated evaluation, we employ a user simulation agent to provide iterative feedback, and a Docker sandbox to conduct visual and functional rubric-based assessment. Experiments reveal that closed-source models significantly outperform open-source models (48.4% vs 32.1%). Further analysis reveals significant disparities across models in resource efficiency, feedback-driven self-correction, and specific tool-use preferences. Finally, we investigate the impact of agentic scaffolds, observing that proprietary models demonstrate superior performance within their native ecosystems (e.g., Claude-4.5-Opus via Claude-Agent-SDK), while open-source models exhibit distinct performance peaks, suggesting potential optimization for specific execution frameworks. AgencyBench serves as a critical testbed for next-generation agents, highlighting the necessity of co-optimizing model architecture with agentic frameworks. We believe this work sheds light on the future direction of autonomous agents, and we release the full benchmark and evaluation toolkit at https://github.com/GAIR-NLP/AgencyBench.

Scaling the context length of large language models (LLMs) offers significant benefits but is computationally expensive. This expense stems primarily from the self-attention mechanism, whose O(N^2) complexity with respect to sequence length presents a major bottleneck for both memory and latency. Fortunately, the attention matrix is often sparse, particularly for long sequences, suggesting an opportunity for optimization. Block-sparse attention has emerged as a promising solution that partitions sequences into blocks and skips computation for a subset of these blocks. However, the effectiveness of this method is highly dependent on the underlying attention patterns, which can lead to sub-optimal block-level sparsity. For instance, important key tokens for queries within a single block may be scattered across numerous other blocks, leading to computational redundancy. In this work, we propose Permuted Block-Sparse Attention (PBS-Attn), a plug-and-play method that leverages the permutation properties of attention to increase block-level sparsity and enhance the computational efficiency of LLM prefilling. We conduct comprehensive experiments on challenging real-world long-context datasets, demonstrating that PBS-Attn consistently outperforms existing block-sparse attention methods in model accuracy and closely matches the full attention baseline. Powered by our custom permuted-FlashAttention kernels, PBS-Attn achieves an end-to-end speedup of up to 2.75times in long-context prefilling, confirming its practical viability. Code available at https://github.com/xinghaow99/pbs-attn

Recent advances in online reinforcement learning (RL) for large language models (LLMs) have demonstrated promising performance in complex reasoning tasks. However, they often exhibit an imbalanced exploration-exploitation trade-off, resulting in unstable optimization and sub-optimal performance. We introduce IB-Score, a novel metric grounded in Information Bottleneck theory that evaluates policy's exploration-exploitation balance by quantifying the trade-off between step-level reasoning diversity and mutual information shared with the correct answer. Analysis based on IB-Score shows that popular online RL approaches (e.g., GRPO) with common regularizers fail to consistently maintain balance during training with suboptimal results. To address this, we propose Information Bottleneck-driven Tree-based Policy Optimization (IB-TPO), a principled framework that formulates IB-Score as a fine-grained optimization objective and utilizes a novel IB-guided tree sampling strategy that not only improves the efficiency of online sampling with 50% more trajectories under the same token budget, but also reuses the tree structure for effective IB-Score Monte Carlo estimation. Extensive experiments across standard benchmarks show that our method significantly outperforms GRPO baseline by 2.9% to 3.6% and also outperforms other state-of-the-art online RL approaches. Our code is available at https://github.com/alibaba/EfficientRL.

With the growing requirement for natural human-computer interaction, speech-based systems receive increasing attention as speech is one of the most common forms of daily communication. However, the existing speech models still experience high latency when generating the first audio token during streaming, which poses a significant bottleneck for deployment. To address this issue, we propose VITA-Audio, an end-to-end large speech model with fast audio-text token generation. Specifically, we introduce a lightweight Multiple Cross-modal Token Prediction (MCTP) module that efficiently generates multiple audio tokens within a single model forward pass, which not only accelerates the inference but also significantly reduces the latency for generating the first audio in streaming scenarios. In addition, a four-stage progressive training strategy is explored to achieve model acceleration with minimal loss of speech quality. To our knowledge, VITA-Audio is the first multi-modal large language model capable of generating audio output during the first forward pass, enabling real-time conversational capabilities with minimal latency. VITA-Audio is fully reproducible and is trained on open-source data only. Experimental results demonstrate that our model achieves an inference speedup of 3~5x at the 7B parameter scale, but also significantly outperforms open-source models of similar model size on multiple benchmarks for automatic speech recognition (ASR), text-to-speech (TTS), and spoken question answering (SQA) tasks.

Compressed short-text generators can fail in two different places: the codec may discard information before generation starts, or the latent generator may produce weak codes. Without separating these failure modes, researchers can spend compute improving the wrong component. We study this problem in a controlled 64-to-16 TinyStories case study built from a hierarchical VQ-VAE-2 codec and a masked discrete diffusion generator (MDLM). We use a staged validation protocol that separates codec reconstruction fidelity, latent generation quality, and auxiliary latent diagnostics under one shared external GPT-2 scorer, while reporting complementary semantic metrics for the geometry study. In the tested configuration, codec reconstruction alone raises median external perplexity from 15.17 to 27.36 (+80.4%) and p95 from 25.10 to 98.91 (+294.1%), showing that the dominant quality loss appears before latent generation begins. Under the same scorer, code-space MDLM remains materially stronger than token-space diffusion, reducing mean, median, and p95 by 32.9%, 30.9%, and 36.6%, respectively. Geometry-aware regularization improves local latent proxies but does not improve decoded-text metrics in the available runs. The contribution is methodological rather than algorithmic: the paper presents a reusable staged diagnosis for one concrete pipeline and shows that, in this setting, codec fidelity rather than latent denoising sets the practical quality ceiling.

Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) for Large Language Model (LLM) reasoning have been hindered by a persistent challenge: exploration collapse. The semantic homogeneity of random rollouts often traps models in narrow, over-optimized behaviors. While existing methods leverage policy entropy to encourage exploration, they face inherent limitations. Global entropy regularization is susceptible to reward hacking, which can induce meaningless verbosity, whereas local token-selective updates struggle with the strong inductive bias of pre-trained models. To address this, we propose Latent Policy Optimization via Iterative Information Bottleneck (IIB-LPO), a novel approach that shifts exploration from statistical perturbation of token distributions to topological branching of reasoning trajectories. IIB-LPO triggers latent branching at high-entropy states to diversify reasoning paths and employs the Information Bottleneck principle both as a trajectory filter and a self-reward mechanism, ensuring concise and informative exploration. Empirical results across four mathematical reasoning benchmarks demonstrate that IIB-LPO achieves state-of-the-art performance, surpassing prior methods by margins of up to 5.3% in accuracy and 7.4% in diversity metrics.

Vision-Language Models (VLMs) have revolutionized document parsing by enabling end-to-end mapping from images to structured text, imposing a significant latency bottleneck, particularly for token-dense documents. While Multi-Token Prediction (MTP) has emerged as a promising approach for accelerating inference, its potential is constrained by optimization instability when scaling to deeper look-ahead depth. In this paper, we propose P-MTP, a framework that leverages Progressive Multi-Token Prediction with a lightweight MTP module to scale the look-ahead depth for high-throughput document parsing. Specifically, we introduce Progressive Curriculum Loss that adaptively re-weights different look-ahead depths using cumulative path reliability and retrospective target consistency. By effectively suppressing gradient noise in long-range predictions, P-MTP, facilitates an automated easy-to-hard optimization transition, enabling the model to master increasingly distant look-ahead depths. Furthermore, we propose Confidence-Gated Dynamic Drafting to maximize the effective look-ahead depth and acceptance rate by adaptively calibrating speculative length during inference, thereby minimizing computational waste and further pushing the boundaries of inference speedup. Experimental results across multiple benchmarks and architectures demonstrate that P-MTP, achieves up to a 5times speedup with negligible loss in accuracy, providing the first successful validation of extensive look-ahead MTP in the document parsing domain.

Improving time-to-first-token (TTFT) is an essentially important objective in modern large language model (LLM) inference engines. Optimizing TTFT directly results in higher maximal QPS and meets the requirements of many critical applications. However, boosting TTFT is notoriously challenging since it is compute-bounded and the performance bottleneck shifts from the self-attention that many prior works focus on to the MLP part. In this work, we present SpecPrefill, a training free framework that accelerates the inference TTFT for both long and medium context queries based on the following insight: LLMs are generalized enough to preserve the quality given only a carefully chosen subset of prompt tokens. At its core, SpecPrefill leverages a lightweight model to speculate locally important tokens based on the context. These tokens, along with the necessary positional information, are then sent to the main model for processing. We evaluate SpecPrefill with a diverse set of tasks, followed by a comprehensive benchmarking of performance improvement both in a real end-to-end setting and ablation studies. SpecPrefill manages to serve Llama-3.1-405B-Instruct-FP8 with up to 7times maximal end-to-end QPS on real downstream tasks and 7.66times TTFT improvement.

Fusion-in-Decoder (FiD) is an effective retrieval-augmented language model applied across a variety of open-domain tasks, such as question answering, fact checking, etc. In FiD, supporting passages are first retrieved and then processed using a generative model (Reader), which can cause a significant bottleneck in decoding time, particularly with long outputs. In this work, we analyze the contribution and necessity of all the retrieved passages to the performance of reader models, and propose eliminating some of the retrieved information, at the token level, that might not contribute essential information to the answer generation process. We demonstrate that our method can reduce run-time by up to 62.2%, with only a 2% reduction in performance, and in some cases, even improve the performance results.

As large language models (LLMs) become increasingly powerful, the sequential nature of autoregressive generation creates a fundamental throughput bottleneck that limits the practical deployment. While Multi-Token Prediction (MTP) has demonstrated remarkable benefits for model training efficiency and performance, its inherent potential for inference acceleration remains largely unexplored. This paper introduces FastMTP, a simple yet effective method that improves multi-step draft quality by aligning MTP training with its inference pattern, significantly enhancing speculative decoding performance. Our approach fine-tunes a single MTP head with position-shared weights on self-distilled data, enabling it to capture dependencies among consecutive future tokens and maintain high acceptance rates across multiple recursive draft steps. By integrating language-aware dynamic vocabulary compression into the MTP head, we further reduce computational overhead in the drafting process. Experimental results across seven diverse benchmarks demonstrate that FastMTP achieves an average of 2.03x speedup compared to standard next token prediction with lossless output quality, outperforming vanilla MTP by 82%. FastMTP requires only lightweight training and seamlessly integrates with existing inference frameworks, offering a practical and rapidly deployable solution for accelerating LLM inference.

The rapid growth in the parameters of large language models (LLMs) has made inference latency a fundamental bottleneck, limiting broader application of LLMs. Speculative decoding represents a lossless approach to accelerate inference through a guess-and-verify paradigm, leveraging the parallel capabilities of modern hardware. Some speculative decoding methods rely on additional structures to guess draft tokens, such as small models or parameter-efficient architectures, which need extra training before use. Alternatively, retrieval-based train-free techniques build libraries from pre-existing corpora or by n-gram generation. However, they face challenges like large storage requirements, time-consuming retrieval, and limited adaptability. Observing that candidate tokens generated during the decoding process are likely to reoccur in future sequences, we propose Token Recycling. This approach stores candidate tokens in an adjacency matrix and employs a breadth-first search (BFS)-like algorithm on the matrix to construct a draft tree. The tree is then validated through tree attention. New candidate tokens from the decoding process are then used to update the matrix. Token Recycling requires \textless2MB of additional storage and achieves approximately 2x speedup across all sizes of LLMs. It significantly outperforms existing train-free methods by 30\% and even a training method by 25\%. It can be directly applied to any existing LLMs and tasks without the need for adaptation.

Recently, large language models (LLM) based on transformers are facing memory bottleneck issues due to KV cache, especially in long sequence handling. Previous researches proposed KV cache compression techniques that identify insignificant tokens based on Accumulative Attention Scores and removes their items from KV cache, noting that only few tokens play an important role in attention operations. However, we have observed that the existing Accumulative Attention Score is not suitable for the transformer decoder structure. In the decoder model, the number of times the Attention Score accumulates varies depending on the order of token appearance due to the effect of masking, causing an uneven comparison between tokens. To solve this, we propose Accumulative Attention Score with Forgetting Factor (A2SF) technique, which introduces a Forgetting Factor in the Attention Score accumulation process. A2SF applies a penalty to the past Attention Score generated from old tokens by repeatedly multiplying the Forgetting Factor to the Attention Score over time. Therefore, older tokens receive a larger penalty, providing fairness among different ages of tokens. Through the fair comparison among tokens, we can more effectively select important tokens. We have verified the accuracy improvement through A2SF in the OPT and LLaMA models and A2SF improves the accuracy of LLaMA 2 by up to 7.8% and 5.1% on 1-shot and 0-shot.

MoE (Mixture of Experts) prevails as a neural architecture that can scale modern transformer-based LLMs (Large Language Models) to unprecedented scales. Nevertheless, large MoEs' great demands of computing power, memory capacity and memory bandwidth make scalable serving a fundamental challenge and efficient parallel inference has become a requisite to attain adequate throughput under latency constraints. DeepSpeed-MoE, one state-of-the-art MoE inference framework, adopts a 3D-parallel paradigm including EP (Expert Parallelism), TP (Tensor Parallel) and DP (Data Parallelism). However, our analysis shows DeepSpeed-MoE's inference efficiency is largely bottlenecked by EP, which is implemented with costly all-to-all collectives to route token activation. Our work aims to boost DeepSpeed-MoE by strategically reducing EP's communication overhead with a technique named Speculative MoE. Speculative MoE has two speculative parallelization schemes, speculative token shuffling and speculative expert grouping, which predict outstanding tokens' expert routing paths and pre-schedule tokens and experts across devices to losslessly trim EP's communication volume. Besides DeepSpeed-MoE, we also build Speculative MoE into a prevailing MoE inference engine SGLang. Experiments show Speculative MoE can significantly boost state-of-the-art MoE inference frameworks on fast homogeneous and slow heterogeneous interconnects.

Diffusion-based world models have shown strong potential for unified world simulation, but the iterative denoising remains too costly for interactive use and long-horizon rollouts. While feature caching can accelerate inference without training, we find that policies designed for single-modal diffusion transfer poorly to world models due to two world-model-specific obstacles: token heterogeneity from multi-modal coupling and spatial variation, and non-uniform temporal dynamics where a small set of hard tokens drives error growth, making uniform skipping either unstable or overly conservative. We propose WorldCache, a caching framework tailored to diffusion world models. We introduce Curvature-guided Heterogeneous Token Prediction, which uses a physics-grounded curvature score to estimate token predictability and applies a Hermite-guided damped predictor for chaotic tokens with abrupt direction changes. We also design Chaotic-prioritized Adaptive Skipping, which accumulates a curvature-normalized, dimensionless drift signal and recomputes only when bottleneck tokens begin to drift. Experiments on diffusion world models show that WorldCache delivers up to 3.7times end-to-end speedups while maintaining 98\% rollout quality, demonstrating the vast advantages and practicality of WorldCache in resource-constrained scenarios. Our code is released in https://github.com/FofGofx/WorldCache.

On-policy distillation (OPD) trains a student on its own trajectories with token-level teacher feedback and often outperforms off-policy distillation and standard reinforcement learning. However, we find that its standard advantage weighted policy gradient suffers from three structural weaknesses, including high variance updates, vanishing gradients in zero-advantage regions, and exploration bottlenecks when corrective signals are insufficient. We therefore propose Asymmetric On-Policy Distillation (AOPD), which replaces ineffective negative reinforcement with localized divergence minimization in non-positive advantage regions while preserving positive reinforcement learning. Experiments on mathematical reasoning benchmarks show that AOPD consistently outperforms standard OPD, with average gains of 4.09 / 8.34 under strong / weak initialization, respectively. AOPD also maintains higher policy entropy during training and better capability retention during sequential tool-use adaptation.

As large language models (LLMs) continue to scale, the memory footprint of key-value (KV) caches during inference has become a significant bottleneck. Existing approaches primarily focus on compressing KV caches within a single prompt or reusing shared prefixes or frequently ocurred text segments across prompts. However, such strategies are limited in scenarios where prompts are semantically similar but lexically different, which frequently occurs in tasks such as multi-document summarization and conversational agents. We propose SemShareKV, a KV cache sharing and compression framework that accelerates LLM inference by reusing KVCache in semantically similar prompts. Instead of relying on exact token matches, SemShareKV applies fuzzy token matching using locality-sensitive hashing (LSH) on token embeddings and incorporates Rotary Position Embedding (RoPE) to better preserve positional information. By selectively reusing relevant key-value pairs from a reference prompt's cache, SemShareKV reduces redundant computation while maintaining output quality. Experiments on diverse summarization datasets show up to 6.25times speedup and 42\% lower GPU memory usage with 5k tokens input, with negligible quality degradation. These results highlight the potential of semantic-aware cache sharing for efficient LLM inference.

Mixture-of-Experts (MoE) model architecture has emerged as a promising solution for scaling transformer models efficiently, offering sparse activation that reduces computational costs while increasing model capacity. However, as MoE models scale, they need to be distributed across GPU devices, thus face critical performance bottlenecks due to their large memory footprint. Expert parallelism distributes experts across GPUs, however, faces key challenges including an unbalanced token routing and expert activation, resulting in communication tail latency and processing inefficiencies. While existing solutions address some of these issues, they fail to resolve the dual challenges of load imbalance and communication skew. The imbalance in token processing load across experts causes uneven processing times on different GPUs, while communication skew between GPUs leads to unbalanced inter-GPU data transfers. These factors degrade the performance of MoE models by increasing tail latency and reducing overall throughput. To address these limitations, we propose an Integer Linear Programming (ILP) formulation to optimize expert placement by jointly considering token load, communication, and computation costs. We exploit the property that there is a token routing dependency across layers, where tokens routed to a specific expert in one layer are likely to be routed to a limited set of experts in the subsequent layer. Our solution, MoETuner, offers an optimal expert-to-GPU assignment that minimizes inter-GPU token routing costs and balances token processing across devices, thereby reducing tail latency and end-to-end execution time. Experimental results demonstrate 9.3% and 17.5% of end-to-end speedups for single-node and multi-node inference respectively, showcasing the potential of our ILP-based optimization for offering expert parallel solutions for next-generation MoEs.

Training Large Language Models (LLMs) on long contexts is severely constrained by prohibitive GPU memory overhead, not training time. The primary culprits are the activations, whose memory footprints scale linearly with sequence length. We introduce OOMB, a highly memory-efficient training system that directly confronts this barrier. Our approach employs a chunk-recurrent training framework with on-the-fly activation recomputation, which maintains a constant activation memory footprint (O(1)) and shifts the primary bottleneck to the growing KV cache. To manage the KV cache, OOMB integrates a suite of synergistic optimizations: a paged memory manager for both the KV cache and its gradients to eliminate fragmentation, asynchronous CPU offloading to hide data transfer latency, and page-level sparse attention to reduce both computational complexity and communication overhead. The synergy of these techniques yields exceptional efficiency. Our empirical results show that for every additional 10K tokens of context, the end-to-end training memory overhead increases by a mere 10MB for Qwen2.5-7B. This allows training Qwen2.5-7B with a 4M-token context on a single H200 GPU, a feat that would otherwise require a large cluster using context parallelism. This work represents a substantial advance in resource efficiency for long-context LLM training. The source code is available at https://github.com/wenhaoli-xmu/OOMB.

As large language models (LLMs) process increasing context windows, the memory usage of KV cache has become a critical bottleneck during inference. The mainstream KV compression methods, including KV pruning and KV quantization, primarily focus on either token or precision dimension and seldom explore the efficiency of their combination. In this paper, we comprehensively investigate the token-precision trade-off in KV cache compression. Experiments demonstrate that storing more tokens in the KV cache with lower precision, i.e., quantized pruning, can significantly enhance the long-context performance of LLMs. Furthermore, in-depth analysis regarding token-precision trade-off from a series of key aspects exhibit that, quantized pruning achieves substantial improvements in retrieval-related tasks and consistently performs well across varying input lengths. Moreover, quantized pruning demonstrates notable stability across different KV pruning methods, quantization strategies, and model scales. These findings provide valuable insights into the token-precision trade-off in KV cache compression. We plan to release our code in the near future.

Long-context LLM agents must access the right evidence from large environments and use it faithfully. However, the popular Needle-in-a-Haystack (NIAH) evaluation mostly measures benign span localization. The needle is near-unique, and the haystack is largely irrelevant. We introduce EverMemBench-S (EMB-S), an adversarial NIAH-style benchmark built on a 326M-token MemoryBank. While the full MemoryBank spans 326M tokens for retrieval-based (RAG) evaluation, we evaluate native long-context models only at scales that fit within each model's context window (up to 1M tokens in this work) to ensure a fair comparison. EMB-S pairs queries with collision-tested near-miss hard negatives and gold evidence sets spanning one or more documents, validated via human screening and LLM verification. We also propose a decoupled diagnostic protocol that reports evidence access (document-ID localization) separately from end-to-end QA quality under full-context prompting. This enables consistent diagnosis for both native long-context prompting and retrieval pipelines. Across a reference-corpus ladder from domain-isolated 64K contexts to a globally shared 326M-token environment, we observe a clear reality gap. Systems that saturate benign NIAH degrade sharply in evidence access under semantic interference. These results indicate that semantic discrimination, not context length alone, is the dominant bottleneck for long-context memory at scale.

Recent advances in Video Large Language Models (VLLMs) have achieved remarkable video understanding capabilities, yet face critical efficiency bottlenecks due to quadratic computational growth with lengthy visual token sequences of long videos. While existing keyframe sampling methods can improve temporal modeling efficiency, additional computational cost is introduced before feature encoding, and the binary frame selection paradigm is found suboptimal. Therefore, in this work, we propose Dynamic Token compression via LLM-guided Keyframe prior (DyToK), a training-free paradigm that enables dynamic token compression by harnessing VLLMs' inherent attention mechanisms. Our analysis reveals that VLLM attention layers naturally encoding query-conditioned keyframe priors, by which DyToK dynamically adjusts per-frame token retention ratios, prioritizing semantically rich frames while suppressing redundancies. Extensive experiments demonstrate that DyToK achieves state-of-the-art efficiency-accuracy tradeoffs. DyToK shows plug-and-play compatibility with existing compression methods, such as VisionZip and FastV, attaining 4.3x faster inference while preserving accuracy across multiple VLLMs, such as LLaVA-OneVision and Qwen2.5-VL. Code is available at https://github.com/yu-lin-li/DyToK .

We present LightVLA, a simple yet effective differentiable token pruning framework for vision-language-action (VLA) models. While VLA models have shown impressive capability in executing real-world robotic tasks, their deployment on resource-constrained platforms is often bottlenecked by the heavy attention-based computation over large sets of visual tokens. LightVLA addresses this challenge through adaptive, performance-driven pruning of visual tokens: It generates dynamic queries to evaluate visual token importance, and adopts Gumbel softmax to enable differentiable token selection. Through fine-tuning, LightVLA learns to preserve the most informative visual tokens while pruning tokens which do not contribute to task execution, thereby improving efficiency and performance simultaneously. Notably, LightVLA requires no heuristic magic numbers and introduces no additional trainable parameters, making it compatible with modern inference frameworks. Experimental results demonstrate that LightVLA outperforms different VLA models and existing token pruning methods across diverse tasks on the LIBERO benchmark, achieving higher success rates with substantially reduced computational overhead. Specifically, LightVLA reduces FLOPs and latency by 59.1% and 38.2% respectively, with a 2.6% improvement in task success rate. Meanwhile, we also investigate the learnable query-based token pruning method LightVLA* with additional trainable parameters, which also achieves satisfactory performance. Our work reveals that as VLA pursues optimal performance, LightVLA spontaneously learns to prune tokens from a performance-driven perspective. To the best of our knowledge, LightVLA is the first work to apply adaptive visual token pruning to VLA tasks with the collateral goals of efficiency and performance, marking a significant step toward more efficient, powerful and practical real-time robotic systems.

Vision-Language Models such as CLIP exhibit strong zero-shot recognition capability by aligning images with textual concepts, yet they often underperform on multi-label recognition where multiple objects co-exist. A key bottleneck is that the [CLS] token, as a single global visual representation, is insufficient to faithfully encode diverse targets with varying scales, contexts, and co-occurrence patterns. To address this limitation, we present a new multi-label image recognition framework, termed PIAA, which formulates prediction as Patch-level Inference followed by Adaptive Aggregation. Specifically, we first enhance patch-wise predictions from two complementary perspectives: (i) mitigating semantic entanglement in the visual encoder to obtain more discriminative patch representations, and (ii) learning an unsupervised visual classifier to narrow the vision-language modality gap. We then introduce an adaptive aggregation module that consolidates patch-level scores into the final multi-label prediction. Notably, the entire pipeline is fully training-free, requiring no gradient updates or parameter fine-tuning. Experiments show that our method achieves strong improvements with minimal extra computation, exceeding a 6% mAP gain on the challenging NUS-WIDE benchmark over representative baselines. Code is available at https://github.com/akang-wang/PIAA.

Despite the strong performance of large audio language models (LALMs) in various tasks, exactly how and where they integrate acoustic features with textual context remains unclear. We adapt causal tracing to investigate the internal information flow of LALMs during audio comprehension. By conducting layer-wise and token-wise analyses across DeSTA, Qwen, and Voxtral, we evaluate the causal effects of individual hidden states. Layer-wise analysis identifies different fusion strategies, from progressive integration in DeSTA to abrupt late-stage fusion in Qwen. Token-wise analysis shows that the final sequence token acts as an informational bottleneck where the network decisively retrieves relevant information from the audio. We also observe an attention-like query mechanism at intermediate token positions that triggers the model to pull task-relevant audio context. These findings provide a clear characterization of when and where multi-modal integration occurs within LALMs.

The rapid advancement toward long-context reasoning and multi-modal intelligence has made the memory footprint of the Key-Value (KV) cache a dominant memory bottleneck for efficient deployment. While the established per-channel quantization effectively accommodates intrinsic channel-wise outliers in Key tensors, its efficacy diminishes under extreme compression. In this work, we revisit the inherent limitations of the per-channel quantization paradigm from both empirical and theoretical perspectives. Our analysis identifies Token Norm Imbalance (TNI) as the primary bottleneck to quantization fidelity. We demonstrate that TNI systematically amplifies errors when shared quantization parameters are required to span token groups exhibiting substantial norm disparities. Instead of relying on intricate quantization pipelines (e.g., TurboQuant), we propose OScaR (Omni-Scaled Canalized Rotation), an accurate and lightweight KV cache compression framework for X-LLMs (i.e., text-only, multi-modal, and omni-modal LLMs). Advancing the per-channel paradigm, OScaR employs Canalized Rotation followed by Omni-Token Scaling to mitigate TNI-induced sequence-dimensional variance both effectively and efficiently, further supported by our optimized system design and CUDA kernels. Extensive evaluations across X-LLMs show that OScaR consistently outperforms existing methods and achieves near-lossless performance under INT2 quantization, establishing it as a robust, low-complexity, and universal framework that defines a new Pareto front. Compared with the BF16 FlashDecoding-v2 baseline, our OScaR implementation achieves a notable up to 3.0x speedup in decoding, reduces memory footprint by 5.3x, and increases throughput by 4.1x. The code for OScaR is publicly available at https://github.com/ZunhaiSu/OScaR-KV-Quant.

Balancing fine-grained local modeling with long-range dependency capture under computational constraints remains a central challenge in sequence modeling. While Transformers provide strong token mixing, they suffer from quadratic complexity, whereas Mamba-style selective state-space models (SSMs) scale linearly but often struggle to capture high-rank and synchronous global interactions. We present a consistency boundary analysis that characterizes when diagonal short-memory SSMs can approximate causal attention and identifies structural gaps that remain. Motivated by this analysis, we propose InfoMamba, an attention-free hybrid architecture. InfoMamba replaces token-level self-attention with a concept bottleneck linear filtering layer that serves as a minimal-bandwidth global interface and integrates it with a selective recurrent stream through information-maximizing fusion (IMF). IMF dynamically injects global context into the SSM dynamics and encourages complementary information usage through a mutual-information-inspired objective. Extensive experiments on classification, dense prediction, and non-vision tasks show that InfoMamba consistently outperforms strong Transformer and SSM baselines, achieving competitive accuracy-efficiency trade-offs while maintaining near-linear scaling.

Generative artificial intelligence (AI) technology is revolutionizing the computing industry. Not only its applications have broadened to various sectors but also poses new system design and optimization opportunities. The technology is capable of understanding and responding in multiple modalities. However, the advanced capability currently comes with significant system resource demands. To sustainably scale generative AI capabilities to billions of users in the world, inference must be fast and efficient. This paper pinpoints key system design and optimization opportunities by characterizing a family of emerging multi-modal generation models on real systems. Auto-regressive token generation is a critical latency performance bottleneck, typically dominated by GPU idle time. In addition to memory-intensive attention across the generative AI models, linear operations constitute significant inference latency due to the feed forward networks in Transformer-based models. We demonstrate that state-of-the-art optimization levers, spanning from applications to system software and hardware, set a 3.88x better baseline.

Transformers have demonstrated success in graph learning, particularly for node-level tasks. However, existing methods encounter an information bottleneck when generating graph-level representations. The prevalent single token paradigm fails to fully leverage the inherent strength of self-attention in encoding token sequences, and degenerates into a weighted sum of node signals. To address this issue, we design a novel serialized token paradigm to encapsulate global signals more effectively. Specifically, a graph serialization method is proposed to aggregate node signals into serialized graph tokens, with positional encoding being automatically involved. Then, stacked self-attention layers are applied to encode this token sequence and capture its internal dependencies. Our method can yield more expressive graph representations by modeling complex interactions among multiple graph tokens. Experimental results show that our method achieves state-of-the-art results on several graph-level benchmarks. Ablation studies verify the effectiveness of the proposed modules.

The deployment of Large Language Models (LLMs) faces a critical bottleneck when handling lengthy inputs: the prohibitive memory footprint of the Key Value (KV) cache. To address this bottleneck, the token pruning paradigm leverages attention sparsity to selectively retain a small, critical subset of tokens. However, existing approaches fall short, with static methods risking irreversible information loss and dynamic strategies employing heuristics that insufficiently capture the query-dependent nature of token importance. We propose FASA, a novel framework that achieves query-aware token eviction by dynamically predicting token importance. FASA stems from a novel insight into RoPE: the discovery of functional sparsity at the frequency-chunk (FC) level. Our key finding is that a small, identifiable subset of "dominant" FCs consistently exhibits high contextual agreement with the full attention head. This provides a robust and computationally free proxy for identifying salient tokens. %making them a powerful and efficient proxy for token importance. Building on this insight, FASA first identifies a critical set of tokens using dominant FCs, and then performs focused attention computation solely on this pruned subset. % Since accessing only a small fraction of the KV cache, FASA drastically lowers memory bandwidth requirements and computational cost. Across a spectrum of long-context tasks, from sequence modeling to complex CoT reasoning, FASA consistently outperforms all token-eviction baselines and achieves near-oracle accuracy, demonstrating remarkable robustness even under constraint budgets. Notably, on LongBench-V1, FASA reaches nearly 100\% of full-KV performance when only keeping 256 tokens, and achieves 2.56times speedup using just 18.9\% of the cache on AIME24.

Vision-language models (VLMs) commonly formulate visual grounding and detection as a coordinate-token generation problem, serializing each 2D box into multiple 1D tokens that are learned and decoded largely independently. This token-by-token decoding mismatches the coupled structure of box geometry and creates a practical inference bottleneck due to strictly sequential generation. We introduce LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). By decoding geometric elements such as bounding boxes and points as atomic units in a single step, LocateAnything preserves intra-box geometric coherence and unlocks substantial parallelism. We show that PBD improves both decoding throughput and localization accuracy. We further develop a scalable data engine and curate LocateAnything-Data, a large-scale dataset with more than 138 million training samples, substantially increasing data diversity for high-precision localization. Extensive evaluations show that LocateAnything advances the speed-accuracy frontier, achieving significantly higher decoding throughput while improving high-IoU localization quality across diverse benchmarks. The results highlight the complementary benefits of Parallel Box Decoding and large-scale training data in enabling efficient and precise unified visual grounding and detection.

Transformer encoder architectures have recently achieved state-of-the-art results on monocular 3D human mesh reconstruction, but they require a substantial number of parameters and expensive computations. Due to the large memory overhead and slow inference speed, it is difficult to deploy such models for practical use. In this paper, we propose a novel transformer encoder-decoder architecture for 3D human mesh reconstruction from a single image, called FastMETRO. We identify the performance bottleneck in the encoder-based transformers is caused by the token design which introduces high complexity interactions among input tokens. We disentangle the interactions via an encoder-decoder architecture, which allows our model to demand much fewer parameters and shorter inference time. In addition, we impose the prior knowledge of human body's morphological relationship via attention masking and mesh upsampling operations, which leads to faster convergence with higher accuracy. Our FastMETRO improves the Pareto-front of accuracy and efficiency, and clearly outperforms image-based methods on Human3.6M and 3DPW. Furthermore, we validate its generalizability on FreiHAND.

The pervasive "memory wall" bottleneck is significantly amplified in modern large-scale Mixture-of-Experts (MoE) architectures. MoE's inherent architectural sparsity leads to sparse arithmetic compute and also introduces substantial activation memory overheads -- driven by large token routing buffers and the need to materialize and buffer intermediate tensors. This memory pressure limits the maximum batch size and sequence length that can fit on GPUs, and also results in excessive data movements that hinders performance and efficient model scaling. We present MoEBlaze, a memory-efficient MoE training framework that addresses these issues through a co-designed system approach: (i) an end-to-end token dispatch and MoE training method with optimized data structures to eliminate intermediate buffers and activation materializing, and (ii) co-designed kernels with smart activation checkpoint to mitigate memory footprint while simultaneously achieving better performance. We demonstrate that MoEBlaze can achieve over 4x speedups and over 50% memory savings compared to existing MoE frameworks.

The autoregressive nature of large language models (LLMs) limits inference speed. Each forward pass generates only a single token and is often bottlenecked by memory bandwidth. Speculative decoding alleviates this issue using a draft-then-verify approach to accelerate token generation. However, the overhead introduced during the draft phase and the training cost of the draft model limit the efficiency and adaptability of speculative decoding. In this work, we introduce PARallel Draft (PARD), a novel speculative decoding method that enables low-cost adaptation of autoregressive draft models into parallel draft models. PARD enhances inference efficiency by predicting multiple future tokens in a single forward pass of the draft phase, and incorporates a conditional drop token method to accelerate training. Its target-independence property allows a single draft model to be applied to an entire family of different models, minimizing the adaptation cost. Our proposed conditional drop token method can improves draft model training efficiency by 3x. On our optimized inference framework, PARD accelerates LLaMA3.1-8B inference by 4.08x, achieving 311.5 tokens per second.

Reinforcement learning (RL) has become a key component in modern large language models, yet the rollout stage remains the key bottleneck in RL training pipelines. Although Multi-Token Prediction (MTP) offers a natural solution to accelerate rollouts through speculative decoding, many studies have observed that MTP acceptance rates degrade significantly during RL training, leading to limited speedup performance. To address this bottleneck, we present Bebop, a systematic study of MTP in LLM post-training, and offer practical recipes to integrate MTP into large-scale RL pipelines. First, we reveal that the MTP acceptance rate is fundamentally bounded by the fluctuation of model entropy, which demonstrates a clear negative linear relationship with the rise of entropy in the RL stage. Second, we show that probabilistic rejection sampling largely alleviates the disturbance introduced by entropy in RL compared to greedy draft sampling. We further identify that the conventional MTP training objectives (cross-entropy or KL) are suboptimal in such settings, and therefore we propose a novel end-to-end TV loss that directly optimizes multi-step rejection sampling acceptance rate, yielding ~10% acceptance rate improvements, achieving up to 95% acceptance rates and up to 25% extra inference throughput gains across mathematical reasoning, code generation, and agentic tasks. Third, we test various online MTP training strategies during RL and show that pre-RL MTP training with e2e TV loss and rejection sampling achieves a consistent acceptance rate and speedup throughout the entire RL, eliminating the need for costly online MTP updating. We provide extensive experiments and analysis that validate our findings. Experimental results show our method achieves up to 1.8x end-to-end acceleration in async RL training of Qwen3.5, Qwen3.6, and Qwen3.7 models.

PromptReps showed that an autoregressive language model can be used directly as a retriever by prompting it to generate dense and sparse representations of a query or passage. Extending this to multiple representatives is inefficient for autoregressive models, since tokens must be generated sequentially, and prior multi-token variants did not reliably improve over single-token decoding. We show that the bottleneck is sequential generation, not the multi-token idea itself. DiffRetriever is a representative-token retriever for diffusion language models: it appends K masked positions to the prompt and reads all K in a single bidirectional forward pass. Across in-domain and out-of-domain evaluation, multi-token DiffRetriever substantially improves over single-token on every diffusion backbone we test, while autoregressive multi-token is flat or negative and pays a latency cost that scales with K where diffusion does not. After supervised fine-tuning, DiffRetriever on Dream is the strongest BEIR-7 retriever in our comparison, ahead of PromptReps, the encoder-style DiffEmbed baseline on the same diffusion backbones, and the contrastively fine-tuned single-vector RepLLaMA. A per-query oracle on the frozen base model exceeds contrastive fine-tuning at the same fixed budget, pointing to adaptive budget selection as future work. Code is available at https://github.com/ielab/diffretriever.

Multimodal protein language models (PLMs) integrate sequence and token-based structural information, serving as a powerful foundation for protein modeling, generation, and design. However, the reliance on tokenizing 3D structures into discrete tokens causes substantial loss of fidelity about fine-grained structural details and correlations. In this paper, we systematically elucidate the design space of multimodal PLMs to overcome their limitations. We identify tokenization loss and inaccurate structure token predictions by the PLMs as major bottlenecks. To address these, our proposed design space covers improved generative modeling, structure-aware architectures and representation learning, and data exploration. Our advancements approach finer-grained supervision, demonstrating that token-based multimodal PLMs can achieve robust structural modeling. The effective design methods dramatically improve the structure generation diversity, and notably, folding abilities of our 650M model by reducing the RMSD from 5.52 to 2.36 on PDB testset, even outperforming 3B baselines and on par with the specialized folding models.

Using a diffusion model for parallel drafting is a promising approach for speculative decoding. By predicting tokens at multiple future positions in a single forward pass, diffusion drafters substantially reduce drafting latency. However, this shifts the bottleneck to verification: verifying a single sequence limits acceptance length, while verifying large draft trees incurs excessive target-model latency. We identify a key mismatch in existing draft-tree methods: existing diffusion-tree methods rank nodes by the marginal probability, ignoring that verification is prefix-conditioned. As a result, they may verify unreachable descendants of rejected prefixes, increasing latency with limited acceptance gains. To address this, we propose TAPS, a target-aware prefix selection method that turns diffusion marginals into path-conditioned acceptance estimates. TAPS then selects a compact prefix-closed subtree under a fixed verification budget, improving the acceptance-cost tradeoff rather than simply expanding the draft tree. Experiments across diverse datasets and model families demonstrate that TAPS achieves up to 7.9x lossless end-to-end speedup over vanilla autoregressive decoding, outperforming state-of-the-art DFlash and DDTree by 1.36x and 1.74x respectively. Our work is available at https://anonymous.4open.science/r/TAPS-EMNLP2026-53DD

Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, especially when guided by explicit chain-of-thought (CoT) reasoning that verbalizes intermediate steps. While CoT improves both interpretability and accuracy, its dependence on natural language reasoning limits the model's expressive bandwidth. Latent reasoning tackles this bottleneck by performing multi-step inference entirely in the model's continuous hidden state, eliminating token-level supervision. To advance latent reasoning research, this survey provides a comprehensive overview of the emerging field of latent reasoning. We begin by examining the foundational role of neural network layers as the computational substrate for reasoning, highlighting how hierarchical representations support complex transformations. Next, we explore diverse latent reasoning methodologies, including activation-based recurrence, hidden state propagation, and fine-tuning strategies that compress or internalize explicit reasoning traces. Finally, we discuss advanced paradigms such as infinite-depth latent reasoning via masked diffusion models, which enable globally consistent and reversible reasoning processes. By unifying these perspectives, we aim to clarify the conceptual landscape of latent reasoning and chart future directions for research at the frontier of LLM cognition. An associated GitHub repository collecting the latest papers and repos is available at: https://github.com/multimodal-art-projection/LatentCoT-Horizon/.

We introduce DiffusionGemma, an experimental open-weight language model that uses discrete diffusion to generate text at exceptionally high speed. Rather than decoding one token at a time, DiffusionGemma iteratively refines blocks of 256 tokens in parallel, avoiding the sequential decoding bottleneck of conventional autoregressive (AR) large language models. Instead of training from scratch, we obtain DiffusionGemma by fine-tuning the mixture-of-experts Gemma 4 model with 3.8B activated and 25.2B total parameters. Our compute-efficient two-stage training pipeline uses fewer than 10% of the starting AR model's total training token budget. The first stage uses supervised fine-tuning to teach bidirectional denoising, while the second stage combines reinforcement learning with sampler distillation to jointly improve generation quality and inference efficiency. DiffusionGemma establishes a new Pareto frontier for the trade-off between generation speed and model capability. Averaged across our full evaluation suite, it generates around 20 tokens per forward pass and achieves roughly 1,500 output tokens per second on a single NVIDIA H100 GPU, which is substantially faster than AR models even with state-of-the-art speculative decoding. DiffusionGemma also retains the starting model's support for thinking mode, multimodal inputs, and long contexts. Despite diffusion fine-tuning, it remains capable of AR generation with only minor performance degradation, suggesting a path toward hybrid diffusion-AR decoding.

Visual-prompt-guided edit transfer aims to learn image transformations directly from example pairs, offering more precise and controllable editing than purely text-driven approaches. However, existing diffusion transformer-based methods often fail to faithfully reproduce the demonstrated edits due to structural mismatches between the task and the backbone, including a pretrained bias toward textual conditioning and inherent stochastic instability during sampling. To bridge this gap, we present EditTransfer++, a framework that combines progressively structured training with an efficient conditioning scheme to improve both visual prompt faithfulness and inference efficiency. We first mitigate textual dominance with a text-decoupled training strategy that removes text conditioning during fine-tuning, compelling the model to infer transformations solely from visual evidence while still supporting optional text guidance at inference. On top of this visually grounded model, a best-worst contrastive refinement mechanism reshapes the denoising trajectories to suppress unfaithful generations and improve consistency across random seeds. To alleviate the computational bottleneck of high-resolution in-context editing, we further introduce a condition compression and reuse strategy that reduces token redundancy and enables efficient generation of images with a 1024-pixel long edge. Extensive experiments on existing benchmarks and the proposed EditTransfer-Bench show that EditTransfer++ achieves state-of-the-art visual prompt faithfulness with substantially faster inference than prior methods, suggesting a promising direction for scalable prompt-guided image editing and broader visual in-context learning.

In this short report, we introduce joint multi-token prediction (JTP), a lightweight modification of standard next-token prediction designed to enrich hidden state representations by jointly predicting multiple future tokens. Unlike previous multi-token prediction approaches, JTP strategically employs teacher forcing of future-tokens through a carefully designed representation bottleneck, allowing the model to encode rich predictive information with minimal computational overhead during training. We show that the JTP approach achieves a short-horizon belief state representation, while popular alternatives for multi-token prediction fail to do so. We demonstrate the effectiveness of our method on the synthetic star graph navigation task from from Bachmann and Nagarajan [2024], highlighting a significant performance improvement over existing methods. This manuscript presents promising preliminary results intended to stimulate further research.

Autonomous LLM agents require structured long-term memory, yet current "append-and-evolve" systems like A-MEM face O(N^2) write-latency and excessive token costs. We introduce D-MEM (Dopamine-Gated Agentic Memory), a biologically inspired architecture that decouples short-term interaction from cognitive restructuring via a Fast/Slow routing system based on Reward Prediction Error (RPE). A lightweight Critic Router evaluates stimuli for Surprise and Utility. Routine, low-RPE inputs are bypassed or cached in an O(1) fast-access buffer. Conversely, high-RPE inputs, such as factual contradictions or preference shifts, trigger a "dopamine" signal, activating the O(N) memory evolution pipeline to reshape the agent's knowledge graph. To evaluate performance under realistic conditions, we introduce the LoCoMo-Noise benchmark, which injects controlled conversational noise into long-term sessions. Evaluations demonstrate that D-MEM reduces token consumption by over 80%, eliminates O(N^2) bottlenecks, and outperforms baselines in multi-hop reasoning and adversarial resilience. By selectively gating cognitive restructuring, D-MEM provides a scalable, cost-efficient foundation for lifelong agentic memory.

Autoregressive large language models (LLMs) are bottlenecked by sequential decoding, where each new token typically requires executing all transformer layers. Existing dynamic-depth and layer-skipping methods reduce this cost, but often rely on auxiliary routing mechanisms or incur accuracy degradation when bypassed layers are left uncompensated. We present LoRA-Drop, a plug-and-play inference framework that accelerates decoding by applying a temporal compute schedule to a fixed subset of intermediate layers: on most decoding steps, selected layers reuse the previous-token hidden state and apply a low-rank LoRA correction, while periodic refresh steps execute the full model to prevent drift. LoRA-Drop requires no routing network, is compatible with standard KV caching, and can reduce KV-cache footprint by skipping KV updates in droppable layers during LoRA steps and refreshing periodically. Across LLaMA2-7B, LLaMA3-8B, Qwen2.5-7B, and Qwen2.5-14B, LoRA-Drop achieves up to 2.6times faster decoding and 45--55\% KV-cache reduction while staying within 0.5 percentage points (pp) of baseline accuracy. Evaluations on reasoning (GSM8K, MATH, BBH), code generation (HumanEval, MBPP), and long-context/multilingual benchmarks (LongBench, XNLI, XCOPA) identify a consistent safe zone of scheduling configurations that preserves quality while delivering substantial efficiency gains, providing a simple path toward adaptive-capacity inference in LLMs. Codes are available at https://github.com/hosseinbv/LoRA-Drop.git.

Diffusion large language models promise parallel token generation, yet inference remains bottlenecked by deciding which masked tokens can be safely committed together. Fast-dLLM addressed this with KV caching and confidence-guided parallel decoding, but its decoding theory uses a homogeneous high-confidence assumption that effectively reduces each candidate set to its weakest selected token. We argue that this leaves speed on the table because real decoding steps exhibit heterogeneous confidence profiles. We propose Fast-dLLM++, a training-free extension that introduces Fréchet profile decoding: selecting parallel commit sets from the full sorted confidence profile rather than a single worst-case confidence. The resulting rule is a heterogeneous-confidence generalization of Fast-dLLM's factor selector and it recovers the previous rule exactly in the equal-confidence case and adds a provable heterogeneity bonus when the selected tokens have uneven confidences. Fast-dLLM++ leaves the model, diffusion process, and cache implementation entirely unchanged, making it a drop-in replacement for existing Fast-dLLM decoding. Experiments on GSM8K, MATH, HumanEval, and MBPP with the LLaDA-8B model show that the theoretical improvement translates directly into empirical gains: profile-aware selection improves the accuracy--throughput frontier by exploiting safe parallelism that weakest-token rules miss, achieving up to 37\% higher throughput at comparable accuracy. Our code release is at https://github.com/Ringo-Star/FastdLLM_plusplus.

Resolving real-world software engineering (SWE) issues with autonomous agents requires complex, long-horizon reasoning. Current pipelines are bottlenecked by unoptimized demonstration data, sparse execution rewards, and computationally prohibitive inference scaling, which collectively exacerbate token bloat, reward hacking, and policy degradation. We present SWE-TRACE (Trajectory Reduction and Agentic Criteria Evaluation), a unified framework optimizing the SWE agent lifecycle across data curation, reinforcement learning (RL), and test-time inference. First, we introduce an LLM multi-task cascading method, utilizing stepwise oracle verification to distill a 60K-instance Supervised Fine-Tuning (SFT) corpus strictly biased toward token-efficient, shortest-path trajectories. Second, to overcome the instability of sparse outcome rewards, we design a MemoryAugmented Agentic RL pipeline featuring a Rubric-Based Process Reward Model (PRM). An auxiliary Rubric-Agent provides dense, fine-grained heuristic feedback on intermediate steps, guiding the model through long-horizon tasks. Finally, we bridge training and inference by repurposing the PRM for heuristic-guided Test-Time Scaling (TTS). By dynamically evaluating and pruning action candidates at each step, SWE-TRACE achieves superior search efficiency without the latency overhead of standard parallel sampling. Extensive experiments on standard SWE benchmarks demonstrate that SWE-TRACE significantly advances the state-of-the-art, maximizing resolution rates while drastically reducing both token consumption and inference latency.

While safety alignment for Multimodal Large Language Models (MLLMs) has gained significant attention, current paradigms primarily target malicious intent or situational violations. We propose shifting the safety frontier toward consequence-driven safety, a paradigm essential for the robust deployment of autonomous and embodied agents. To formalize this shift, we introduce OOD-MMSafe, a benchmark comprising 455 curated query-image pairs designed to evaluate a model's ability to identify latent hazards within context-dependent causal chains. Our analysis reveals a pervasive causal blindness among frontier models, with the highest 67.5% failure rate in high-capacity closed-source models, and identifies a preference ceiling where static alignment yields format-centric failures rather than improved safety reasoning as model capacity grows. To address these bottlenecks, we develop the Consequence-Aware Safety Policy Optimization (CASPO) framework, which integrates the model's intrinsic reasoning as a dynamic reference for token-level self-distillation rewards. Experimental results demonstrate that CASPO significantly enhances consequence projection, reducing the failure ratio of risk identification to 7.3% for Qwen2.5-VL-7B and 5.7% for Qwen3-VL-4B while maintaining overall effectiveness.

Large Vision-Language Models (LVLMs) map visual inputs into dense token sequences, imposing a quadratic computational bottleneck for inference. Elastic visual-token compression addresses this by training a single model that can run at multiple visual-token budgets. However, existing approaches struggle under aggressive compression. Spatial-only compression, as in nested pooling, behaves as an imperfect low-pass filter and induces spectral aliasing that obscures fine-grained detail. Query-only compression, as in nested query resampling, replaces explicit grid-aligned tokens with non-local summaries and substantially degrades spatial grounding. To resolve this representational conflict, we introduce PARCEL (Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding), a visual tokenization architecture that dynamically partitions the labor of feature extraction. PARCEL establishes spatial pool tokens as low-frequency layout anchors and conditions elastic query tokens on these anchors through Pool-Conditioned Query Resampling. This encourages query tokens to focus on complementary visual features rather than redundant spatial mapping. Extensive evaluations across 27 benchmarks show that PARCEL improves the performance-efficiency Pareto frontier, consistently outperforming existing matryoshka baselines across visual-token budgets while preserving the "train once, deploy anywhere" paradigm.

Large language models (LLMs) have driven significant advancements across diverse NLP tasks, with long-context models gaining prominence for handling extended inputs. However, the expanding key-value (KV) cache size required by Transformer architectures intensifies the memory constraints, particularly during the decoding phase, creating a significant bottleneck. Existing sparse attention mechanisms designed to address this bottleneck have two limitations: (1) they often fail to reliably identify the most relevant tokens for attention, and (2) they overlook the spatial coherence of token selection across consecutive Transformer layers, which can lead to performance degradation and substantial overhead in token selection. This paper introduces TidalDecode, a simple yet effective algorithm and system for fast and accurate LLM decoding through position persistent sparse attention. TidalDecode leverages the spatial coherence of tokens selected by existing sparse attention methods and introduces a few token selection layers that perform full attention to identify the tokens with the highest attention scores, while all other layers perform sparse attention with the pre-selected tokens. This design enables TidalDecode to substantially reduce the overhead of token selection for sparse attention without sacrificing the quality of the generated results. Evaluation on a diverse set of LLMs and tasks shows that TidalDecode closely matches the generative performance of full attention methods while reducing the LLM decoding latency by up to 2.1x.

Recent conversational memory systems invest heavily in LLM-based structuring at ingestion time and learned retrieval policies at query time. We show that neither is necessary. SmartSearch retrieves from raw, unstructured conversation history using a fully deterministic pipeline: NER-weighted substring matching for recall, rule-based entity discovery for multi-hop expansion, and a CrossEncoder+ColBERT rank fusion stage -- the only learned component -- running on CPU in ~650ms. Oracle analysis on two benchmarks identifies a compilation bottleneck: retrieval recall reaches 98.6%, but without intelligent ranking only 22.5% of gold evidence survives truncation to the token budget. With score-adaptive truncation and no per-dataset tuning, SmartSearch achieves 93.5% on LoCoMo and 88.4% on LongMemEval-S, exceeding all known memory systems under the same evaluation protocol on both benchmarks while using 8.5x fewer tokens than full-context baselines.

While Chain-of-Thought (CoT) reasoning significantly enhances the performance of Multimodal Large Language Models (MLLMs), its autoregressive nature incurs prohibitive latency constraints. Current efforts to mitigate this via token compression often fail by blindly applying text-centric metrics to multimodal contexts. We identify a critical failure mode termed Visual Amnesia, where linguistically redundant tokens are erroneously pruned, leading to hallucinations. To address this, we introduce V-Skip that reformulates token pruning as a Visual-Anchored Information Bottleneck (VA-IB) optimization problem. V-Skip employs a dual-path gating mechanism that weighs token importance through both linguistic surprisal and cross-modal attention flow, effectively rescuing visually salient anchors. Extensive experiments on Qwen2-VL and Llama-3.2 families demonstrate that V-Skip achieves a 2.9times speedup with negligible accuracy loss. Specifically, it preserves fine-grained visual details, outperforming other baselines over 30\% on the DocVQA.

On-policy distillation transfers reasoning capabilities by training a student model on its own generated trajectories using token-level feedback from a teacher. However, we identify a critical bottleneck, Supervision Fidelity Decay (SFD): as student-generated prefixes lengthen, the teacher's next-token distribution becomes less confident and less discriminative. Consequently, the teacher-dependent corrective signal in reverse-KL distillation weakens, causing student drift to compound across long reasoning chains. To mitigate SFD, we introduce Lookahead Group Reward (\ours{)}. Building on the insight that next-step teacher confidence reflects the discriminative strength of future reverse-KL supervision, evaluates the student's top-K candidate tokens by the teacher confidence they induce at the subsequent step and assigns a group-normalized reward. To maintain computational efficiency, we further design an entropy-triggered tree-attention mechanism. Across six math and code benchmarks, improves mean@8 by 2.57 points over OPD for a 7B student, with gains increasing in longer-generation and reaching +4.92 points on AIME-26 at 39k tokens.

Test-Time Training (TTT) language models achieve theoretically infinite context windows with an O(1) memory footprint by replacing the standard exact-attention KV-cache with hidden state ``fast weights'' W_fast updated via self-supervised learning during inference. However, pure TTT architectures suffer catastrophic failures on exact-recall tasks (e.g., Needle-in-a-Haystack). Because the fast weights aggressively compress the context into an information bottleneck, highly surprising or unique tokens are rapidly overwritten and forgotten by subsequent token gradient updates. We introduce SR-TTT (Surprisal-Aware Residual Test-Time Training), which resolves this recall failure by augmenting the TTT backbone with a loss-gated sparse memory mechanism. By dynamically routing only incompressible, highly surprising tokens to a traditional exact-attention Residual Cache, SR-TTT preserves O(1) memory for low-entropy background context while utilizing exact attention exclusively for critical needles. Our complete implementation, training scripts, and pre-trained weights are open-source and available at: https://github.com/swamynathanvp/Surprisal-Aware-Residual-Test-Time-Training.

By treating visual tokens from visual encoders as text tokens, Multimodal Large Language Models (MLLMs) have achieved remarkable progress across diverse visual understanding tasks, leveraging the robust architectures of Large Language Models (LLMs). However, as token counts grow, the quadratic scaling of computation in LLMs introduces a significant efficiency bottleneck, impeding further scalability. Although recent approaches have explored pruning visual tokens or employing lighter LLM architectures, the computational overhead from an increasing number of visual tokens remains a substantial challenge. In this study, we investigate the redundancy in visual computation at both the parameter and computational pattern levels within LLaVA, a representative MLLM, and introduce a suite of streamlined strategies to enhance efficiency. These include neighbor-aware visual token attention, pruning of inactive visual attention heads, and selective layer dropping for visual computations. By implementing these strategies in LLaVA, we achieve a reduction in computational demands of 88% while maintaining model performance across key benchmarks. Additionally, we validate the existence of visual computational redundancy in other MLLMs, such as Qwen2-VL-7B and InternVL-2.0-4B/8B/26B. These results present a novel pathway for MLLMs to handle dense visual tokens with minimal computational costs. Code and model checkpoints will be released to support further research.

Long-video understanding with multimodal language models suffers from three compounding bottlenecks: heavy decode cost to obtain dense RGB frames, quadratic token growth with frame count, and weak motion perception under sparse keyframe sampling. We present HY-Himmel, a hierarchical video-language framework that allocates semantic and motion capacity separately. A small set of sparse anchor I-frames is routed to the expensive host ViT to ground object identity and scene layout, while the far denser inter-frame intervals are encoded by a lightweight compressed-domain tri-stream adapter that distils motion evidence from motion-vector maps, residual maps, and I-frame context into aligned motion tokens. These tokens are injected into the LLM via a differentiable placeholder mechanism after a dedicated Stage-1 contrastive alignment that places the motion representation in a geometry compatible with the frozen visual backbone. On Video-MME, HY-Himmel surpasses the dense 32-frame baseline by +2.3 pp (61.2 to 63.5%) while using 3.6x fewer context tokens. Extensive ablations over stream composition, motion encoder family, fusion mode, alignment objective, anchor count, LoRA rank, and video duration confirm that the full tri-stream is necessary and sufficient for the observed gains.

Practical large language model (LLM) services may involve a long system prompt, which specifies the instructions, examples, and knowledge documents of the task and is reused across numerous requests. However, the long system prompt causes throughput/latency bottlenecks as the cost of generating the next token grows w.r.t. the sequence length. This paper aims to improve the efficiency of LLM services that involve long system prompts. Our key observation is that handling these system prompts requires heavily redundant memory accesses in existing causal attention computation algorithms. Specifically, for batched requests, the cached hidden states (i.e., key-value pairs) of system prompts are transferred from off-chip DRAM to on-chip SRAM multiple times, each corresponding to an individual request. To eliminate such a redundancy, we propose RelayAttention, an attention algorithm that allows reading these hidden states from DRAM exactly once for a batch of input tokens. RelayAttention is a free lunch: it maintains the generation quality while requiring no model retraining, as it is based on a mathematical reformulation of causal attention.

The increasing complexity of AI tasks has shifted the paradigm from monolithic models toward multi-agent large language model (LLM) systems. However, these collaborative architectures introduce a critical bottleneck: redundant prefill computation for shared content generated by previous agents, which significantly increases KV cache memory usage and time-to-first-token (TTFT). While various KV cache methods have been proposed to mitigate prefill redundancy, they either fail to maintain accuracy on agent-generated outputs or exhibit low reuse rates due to rigid constraints. We present RelayCaching, a training-free inference method that directly reuses decoding phase KV caches from previous agents in subsequent prefill phases. Our key insight is that KV caches for identical content are highly consistent across phases, while prefix-induced deviations are sparse and localized within a limited range of layers and token positions. By selectively recomputing KV caches at these positions, RelayCaching preserves model accuracy with minimal overhead, yielding a superior accuracy-efficiency trade-off over existing methods. Experiments on diverse collaborative LLM tasks spanning mathematical reasoning, general knowledge, and code generation demonstrate that RelayCaching achieves over 80% KV cache reuse, reduces TTFT by up to 4.7times compared to the standard pipeline, all with negligible accuracy degradation.