Datasets:
arxiv float64 2k 2.61k | architecture stringlengths 3 15 | stratum stringclasses 5
values | category stringclasses 2
values | claim stringlengths 30 200 | claim_page int64 1 7 | claim_quote stringlengths 53 247 | ablation_level stringclasses 3
values | ablation_page float64 0 47 ⌀ | ablation_quote stringlengths 64 204 ⌀ | decided_by stringclasses 4
values |
|---|---|---|---|---|---|---|---|---|---|---|
2,007.14062 | big_bird | 2020-2022 | ARCH | BigBird sparse attention mechanism (random + window + global tokens) whose complexity is linear in sequence length | 1 | To remedy this, we propose, B IG B IRD, a sparse attention mechanism that reduces this quadratic dependency to linear. | PARTIAL | 4 | We found that random blocks and local window were insufficient in capturing all the context necessary to compete with the performance of BERT. | third model (pair) |
2,007.14062 | big_bird | 2020-2022 | ARCH | Global tokens (ITC/ETC) attending to the entire sequence are critical for expressivity and performance | 1 | our theoretical analysis reveals some of the benefits of having O(1) global tokens (such as CLS), that attend to the entire sequence as part of the sparse attention mechanism. | NO | null | null | third model (pair) |
2,003.10555 | electra | 2020-2022 | TRAIN | Replaced token detection: a new, more sample-efficient pre-training task in which a discriminator predicts, for every input token, whether a small generator replaced it | 1 | As an alternative, we propose a more sample-efficient pre-training task called replaced token detection. | YES | 6 | To provide a fair comparison, we also train a BERT-Small model using the same hyperparameters. | third model (A_only) |
2,003.10555 | electra | 2020-2022 | TRAIN | Learning from all input tokens rather than the masked subset makes the task more compute-efficient than MLM | 1 | A key advantage of our discriminative task is that the model learns from all input tokens instead of just the small masked-out subset, making it more computationally efficient. | YES | 8 | First, we find that ELECTRA is greatly benefiting from having a loss defined over all input tokens rather than just a subset: ELECTRA 15% performs much worse than ELECTRA. | third model (A_only) |
2,001.0821 | mbart | 2020-2022 | TRAIN | mBART is the first method to pre-train a complete sequence-to-sequence model by denoising full texts in multiple languages | 1 | mBART is the first method for pre-training a complete sequence-to-sequence model by | PARTIAL | 4 | Table 2: Low/Medium Resource Machine Translation Pre-training consistently improves over a randomly ini- | both coders |
2,206.12131 | mvp | 2020-2022 | TRAIN | Pre-train task-specific soft (prefix) prompts per task to stimulate task capacity and mitigate multi-task neutralization | 2 | To mitigate this potential issue, we propose to learn | NO | null | null | third model (pair) |
2,206.12131 | mvp | 2020-2022 | TRAIN | Supervised pre-training (with labeled data) for NLG-oriented PLMs, rather than unsupervised pre-training | 2 | To fill this gap, we explore supervised
pre-training and multi-task learning for deriving | NO | null | null | third model (A_only) |
2,001.04063 | prophetnet | 2020-2022 | TRAIN | Novel self-supervised objective: future n-gram prediction, i.e. predicting the next n tokens at once at each time step | 3 | For simplicity, we take bigram (n = 2) as an example to introduce ProphetNet, whose modeling target is | NO | null | null | third model (A_only) |
2,001.04063 | prophetnet | 2020-2022 | ARCH | Proposed n-stream self-attention in the decoder (extends XLNet two-stream attention), with main and predicting streams that can be switched off at inference | 4 | N-stream self-attention mechanism which contains a main stream self-attention and n predicting stream self-attention. | PARTIAL | 7 | Transformer (Raffel et al., 2019) 39.19 17.60 36.69 ProphetNetw/o pre-train 40.66 18.05 37.79 | third model (A_only) |
2,205.06266 | xmod | 2020-2022 | ARCH | Language-specific modular components pre-trained from the start (X-MOD) mitigate negative interference between languages and achieve positive transfer | 2 | that our modular architecture not only is effective at mitigating interference between languages, but also | YES | 7 | Table 2: Pre-trained language results for the modular and shared model variants, pre-trained on the set of 60 | third model (pair) |
2,309.11419 | kosmos2_5 | 2023 | TRAIN | Two complementary pre-training tasks: document-level text recognition (spatially-aware text with bounding boxes) and image-to-markdown generation, unified via task prompts | 2 | KOSMOS -2.5 is a multimodal document foundation model that takes text images as input and generates spatially-aware | NO | null | null | third model (A_only) |
2,312.00752 | mamba | 2023 | ARCH | Selection mechanism: making SSM parameters (Δ, B, C) input-dependent (selective SSM / S6) | 1 | simply letting the SSM parameters be functions of the input addresses their weakness with discrete modalities | YES | 32 | Replacing any of these with a selective SSM (S6) significantly improves performance, validating the motivation of Section 3. | both coders |
2,312.00752 | mamba | 2023 | ARCH | Simplified homogeneous Mamba block architecture without attention or MLP blocks | 1 | We integrate these selective SSMs into a simplified end-to-end neural network architecture without attention or even MLP blocks (Mamba) | YES | 33 | The Mamba architecture performs similarly to the H3 architecture (and seems slightly better when using a selective layer). | both coders |
2,310.06825 | mistral | 2023 | ARCH | Uses sliding window attention (SWA) to handle sequences of arbitrary length at reduced inference cost | 1 | coupled with sliding window attention (SWA) to effectively handle sequences of arbitrary length with a reduced inference cost. | NO | null | null | both coders |
2,310.06825 | mistral | 2023 | ARCH | Uses grouped-query attention (GQA) for faster inference and lower decoding memory | 1 | GQA significantly accelerates the inference speed, and also reduces the memory requirement during decoding, allowing for higher batch sizes hence higher throughput | NO | null | null | both coders |
2,410.05993 | aria | 2024 | ARCH | First open mixture-of-experts model that is multimodal native | 1 | In this report, we present A RIA, the first open mixture-of-experts (MoE) model that is multimodal native. | NO | null | null | both coders |
2,410.05993 | aria | 2024 | TRAIN | 4-stage training pipeline that progressively adds capabilities | 2 | We design a 4-stage training pipeline, including language pre-training, multimodal pre-training, multimodal long-context pre-training, and multimodal post-training. | NO | null | null | both coders |
2,410.05993 | aria | 2024 | ARCH | Lightweight 438M visual encoder handles variable-length, variable-size and variable-aspect visual input | 1 | Visual input of variable length, size, and aspect is encoded as visual tokens using a | NO | null | null | both coders |
2,412.19437 | deepseek_v3 | 2024 | TRAIN | Multi-Token Prediction training objective improves model performance and enables speculative decoding | 5 | We investigate a Multi-Token Prediction (MTP) objective and prove it beneficial to model performance. It can also be used for speculative decoding for inference acceleration. | YES | 26 | On top of them, keeping the training data and the other architectures the same, we append a 1-depth MTP module onto them and train two models with the MTP strategy for comparison. | both coders |
2,412.19437 | deepseek_v3 | 2024 | ARCH | Pioneers an auxiliary-loss-free load balancing strategy for MoE that minimizes performance degradation from load balancing | 5 | we pioneer an auxiliary-loss-free strategy for load balancing, which minimizes the performance degradation that arises from encouraging load balancing. | YES | 27 | On top of these two baseline models, keeping the training data and the other architectures the same, we remove all auxiliary losses and introduce the auxiliary-loss-free balancing strategy for comparison. | both coders |
2,412.19437 | deepseek_v3 | 2024 | TRAIN | FP8 mixed-precision framework, with first validation on an extremely large-scale model. | 5 | We design an FP8 mixed precision training framework and, for the first time, validate the
feasibility and effectiveness of FP8 training on an extremely large-scale model. | PARTIAL | 47 | We validate our FP8 mixed precision framework with a comparison to BF16 training on top of
two baseline models across different scales. | third model (B_only) |
2,409.18869 | emu3 | 2024 | ARCH | Emu3: multimodal models trained solely with next-token prediction; a single transformer trained from scratch on mixed image/text/video token sequences, without diffusion or compositional architectures | 1 | By tokenizing images, text, and videos into a discrete space, we train a single transformer from scratch on a mixture of multimodal sequences. | NO | null | null | both coders |
2,410.05355 | falcon_mamba | 2024 | ARCH | Falcon Mamba 7B is a new base LLM with a pure Mamba (attention-free) architecture and the first SSLM in the Falcon series, presented as the first competitive attention-free 7B model | 2 | We introduce Falcon Mamba 7B — a base (pre-trained) model with pure mamba architecture design, and the first State Space Language Model (SSLM) in the FalconLLM series. | NO | null | null | both coders |
2,410.05355 | falcon_mamba | 2024 | ARCH | A pure Mamba design can match or beat Transformer and hybrid designs | 1 | we demonstrate that even the pure Mamba design can achieve similar, even superior results compared to the Transformer and hybrid designs. | PARTIAL | 0 | null | third model (A_only) |
2,408.00118 | gemma2 | 2024 | TRAIN | Training 2B and 9B models with knowledge distillation instead of next-token prediction | 1 | and 9B models with knowledge distillation (Hinton et al., 2015) instead of next token prediction. | YES | 5 | In Table 6, we show that distilling from a larger model improves | both coders |
2,408.00118 | gemma2 | 2024 | ARCH | Interleaving local sliding-window and global attention layers (known technique applied) | 1 | several known technical modifications to the Transformer architecture, such as interleaving local-global attentions (Beltagy et al., 2020a) | NO | null | null | both coders |
2,405.2106 | mamba2 | 2024 | ARCH | Mamba-2 architecture combining parallel block with SSD as inner SSM layer | 2 | The combination of the modified parallel Mamba block, together with using SSD as the inner SSM layer, results in the Mamba-2 architecture. | PARTIAL | 33 | Surprisingly, it is significantly better than Mamba-1 even when the state sizes are controlled (N = 16). | both coders |
2,405.2106 | mamba2 | 2024 | ARCH | Introduce heads from multi-head attention to SSMs (head structure design, MVA/MIS pattern) | 2 | For example, we introduce the analog of heads from multi-head attention (MHA) to SSMs. | YES | 35 | Table 5 ablates choices of the multi-head structure for the Mamba-2 architecture. | third model (pair) |
2,405.2106 | mamba2 | 2024 | ARCH | Modified Mamba block with grouped-value head structure and parallel data-dependent projections at block start | 2 | The main ideas include introducing grouped-value attention (GVA) head structure, and moving all data-dependent projections to occur in parallel at the beginning of the block. | PARTIAL | 35 | Table 4 ablates these architecture changes to the block, which occur outside of the core SSM layer. | third model (pair) |
2,405.2106 | mamba2 | 2024 | ARCH | Generalize linear attention to structured masked attention. | 2 | tensor contractions, and then generalize it to a new family
of structured masked attention (SMA) (Section 4). | NO | null | null | third model (B_only) |
2,410.00037 | moshi | 2024 | TRAIN | Inner Monologue: predicting time-aligned text tokens as prefix to audio tokens | 3 | we introduce Inner Monologue, a new training and inference setup for audio language models that significantly improves the factuality and linguistic quality of generated speech | YES | 26 | Finally, the most drastic improvement to the quality and length of generated speech comes from enabling Inner Monologue. | both coders |
2,410.00037 | moshi | 2024 | ARCH | New hierarchical streaming architecture (Helium + small Depth Transformer, RQ-Transformer) for audio tokens | 3 | a new architecture for audio language modeling, which combines Helium with a smaller Transformer (Vaswani et al., 2017) model to predict audio | YES | 26 | In that context, modeling RVQ tokens with an RQ-Transformer significantly improves perplexity over using separate classification heads. | third model (A_only) |
2,410.00037 | moshi | 2024 | ARCH | First multi-stream audio language model for full-duplex dialogue | 2 | We furthermore introduce the first multi-stream audio language model, i.e. a model that explicitly processes the input and output audio streams jointly into two autoregressive token streams. | NO | null | null | third model (A_only) |
2,410.00037 | moshi | 2024 | TRAIN | Introduce improved training tricks for Mimi's semantic-acoustic codec. | 3 | first level of acoustic tokens and introduce improved training tricks. | YES | 24 | This human evaluation shows a significant improve-
ment from using adversarial losses only, with a MUSHRA score of 81.0 against 58.8 when
using the mix of loss functions used in Encodec. | third model (B_only) |
2,409.0206 | olmoe | 2024 | ARCH | Dropless token-choice routing outperforms expert-choice routing | 3 | we find dropless [58] token-based routing [154] outperforms expert-based routing [219] | YES | 11 | Both models have an 8-expert MoE in every 2nd layer. For TC, 2 experts are activated per token | both coders |
2,409.0206 | olmoe | 2024 | ARCH | Fine-grained routing with 64 small experts, 8 active, is critical | 3 | One critical design decision for making MoEs performant is the use of fine-grained routing with granular experts [42]: we employ 64 small experts in each layer | YES | 10 | We vary the number of experts in tandem with the FFN dimension to ensure that active and total parameters and thus compute cost remain the same. | both coders |
2,409.0206 | olmoe | 2024 | ARCH | Shared experts are ineffective | 3 | Our findings also include those that challenge prior work, such as the ineffectiveness of shared experts | YES | 10 | While both settings lead to similar performance, sharing an expert performs slightly worse. | both coders |
2,409.0206 | olmoe | 2024 | TRAIN | Sparse upcycling from a dense model has limited benefits except at small compute budgets | 3 | challenge prior work, such as the ineffectiveness of shared experts [42] and the limited benefits of sparsely | YES | 12 | We find that after 500B tokens, an otherwise equivalent MoE trained from scratch already catches up with the upcycled model | third model (pair) |
2,404.07839 | recurrent_gemma | 2024 | ARCH | RecurrentGemma uses the Griffin architecture, which combines linear recurrences with local attention and has a fixed-size state | 1 | Griffin combines linear recurrences with local attention to achieve excellent performance on language. It has a fixed-sized state, which reduces memory use and enables efficient inference on long sequences. | PARTIAL | 2 | Gemma-2B was trained on 3T tokens and Gemma-7B was trained on 6T tokens, while both RecurrentGemma-2B and RecurrentGemma-9B were trained on 2T tokens. | both coders |
2,411.15242 | zamba2 | 2024 | ARCH | Rotary position embeddings in the shared attention blocks improve performance | 2 | We apply Rotary Position Embeddings (Su et al., 2023) | NO | null | null | both coders |
2,411.15242 | zamba2 | 2024 | ARCH | Two alternating shared attention blocks instead of a single shared block, which improves performance against parameter-matched baselines | 2 | We use two alternating shared attention blocks instead | NO | null | null | both coders |
2,411.15242 | zamba2 | 2024 | ARCH | Non-shared LoRAs applied to the shared transformer blocks | 2 | et al., 2021) to the shared transformer blocks (Attention, | NO | null | null | both coders |
2,411.15242 | zamba2 | 2024 | ARCH | Switched from a Mamba1 to a Mamba2 backbone, which allows a larger state size at the same throughput | 2 | for a larger state-size than is possible in Mamba1, which | NO | null | null | third model (A_only) |
2,507.11407 | exaone4 | 2025 | ARCH | Hybrid global/local (sliding window) attention to reduce compute with minimal degradation | 1 | a hybrid architecture combining global attention and local attention is adopted. This approach minimizes performance degradation while reducing computational costs | NO | null | null | both coders |
2,508.06471 | glm4_moe | 2025 | TRAIN | Multi-stage pre-/mid-training on 23T tokens | 1 | Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning | NO | null | null | both coders |
2,508.06471 | glm4_moe | 2025 | ARCH | GLM-4.5 is an open-source MoE LLM with 355B total parameters and 32B activated parameters | 1 | We present GLM-4.5, an open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters | NO | null | null | both coders |
2,506.21734 | hrm_text | 2025 | TRAIN | One-step gradient approximation removes the need for BPTT and uses O(1) memory | 3 | we propose a one-step gradient approximation for training HRM, which offers | NO | null | null | both coders |
2,506.21734 | hrm_text | 2025 | ARCH | Hierarchical Reasoning Model: a new recurrent architecture with two coupled modules, a slow high-level module and a fast low-level module | 1 | we propose the Hierarchical Reasoning Model (HRM), a novel recurrent architecture that attains significant computational depth | PARTIAL | 12 | retains the exact training setup of HRM but swaps in a Transformer architecture | both coders |
2,506.21734 | hrm_text | 2025 | ARCH | Hierarchical convergence avoids the premature convergence of standard recurrent models | 3 | avoids the rapid convergence of standard recurrent models through a process we term | PARTIAL | 3 | Standard architectures saturates, failing to benefit from increased depth. HRM overcomes this fundamental limitation | third model (A_only) |
2,506.21734 | hrm_text | 2025 | TRAIN | Deep supervision across segments, with the hidden state detached between segments | 7 | we incorporate a deep supervision mechanism into HRM, as detailed next. | NO | null | null | third model (A_only) |
2,506.21734 | hrm_text | 2025 | TRAIN | Adaptive computation time (halting learned with Q-learning), which also gives inference-time scaling | 7 | Inspired by the above mechanism, we incorporate an adaptive halting strategy into HRM that | PARTIAL | 8 | Figure 5 presents a performance comparison between two HRM variants: one incorporating ACT and another employing a fixed computational step count | third model (A_only) |
2,510.26692 | kimi_linear | 2025 | ARCH | Kimi Delta Attention (KDA): linear attention refining gated delta rule with fine-grained channel-wise gating | 2 | Kimi Delta Attention (KDA): a linear attention mechanism that refines the gated delta rule with improved recurrent | YES | 9 | hybrid Gated DeltaNet (GDN-H) baseline, all of which share the same architecture, parameter count, and training setup | both coders |
2,510.26692 | kimi_linear | 2025 | ARCH | Kimi Linear hybrid architecture with 3:1 KDA-to-global attention ratio | 2 | The Kimi Linear architecture: a hybrid design adopting a 3:1 KDA-to-global attention ratio, reducing memory | PARTIAL | 8 | All models were trained with the same FLOPs budget and hyperparameters for a fair comparison. | third model (pair) |
2,509.01322 | longcat_flash | 2025 | TRAIN | Model-growth initialization from half-scale checkpoint | 4 | We initialize the model using a model-growth mechanism based on a refined half-scale checkpoint, achieving improved performance compared to conventional initialization methods | PARTIAL | 10 | Validataion loss curve of a 6B activated MoE model in model growth experiments | both coders |
2,509.01322 | longcat_flash | 2025 | ARCH | Shortcut-connected MoE (ScMoE) to expand computation-communication overlap | 4 | we incorporate the Shortcut-connected MoE (ScMoE) [Cai et al., 2024] design to expand the computation-communication overlap window | PARTIAL | 8 | Training loss curves comparing baseline models (without ScMoE) against their ScMoE-enhanced counterparts across four different model configurations | both coders |
2,509.01322 | longcat_flash | 2025 | TRAIN | Hyperparameter transfer from smaller proxy models | 4 | We successfully apply a hyperparameter transfer strategy to such a large model, predicting optimal hyperparameter configurations by leveraging results from smaller proxy models | NO | null | null | both coders |
2,509.01322 | longcat_flash | 2025 | ARCH | Zero-computation experts for dynamic per-token computation budget (18.6B-31.3B active) | 4 | we introduce the zero-computation experts mechanism in MoE blocks to allocate a dynamic computation budget to important tokens based on their significance | PARTIAL | 7 | Validation loss curve comparing models with/without zero-computation experts under matched computation | both coders |
2,509.01322 | longcat_flash | 2025 | TRAIN | Multi-pronged stability suite: router-gradient balancing, hidden z-loss, optimizer configuration | 4 | A multi-pronged stability suite incorporates principled router-gradient balancing, a hidden z-loss to suppress massive activations, and fine-tuned optimizer configurations | PARTIAL | 15 | The introduction of a negligible-coefficient hidden z-loss stabilizes the norm curve without degrading training loss. | third model (pair) |
2,509.01322 | longcat_flash | 2025 | TRAIN | PID-controlled expert bias maintains the average computation budget. | 4 | To ensure consistent computation load, we employ expert
bias adjusted by a PID-controller, maintaining an average of ∼27 billion activated parameters per token. | NO | null | null | third model (B_only) |
2,509.01322 | longcat_flash | 2025 | TRAIN | Hidden z-loss suppresses massive activations for training stability. | 4 | a hidden z-loss to suppress massive activations, and fine-tuned optimizer
configurations. | PARTIAL | 15 | The introduction of a negligible-coefficient hidden z-loss stabilizes the norm curve without degrading training loss. | third model (B_only) |
2,509.01322 | longcat_flash | 2025 | TRAIN | Fine-tuned optimizer configurations improve large-scale training stability. | 4 | A multi-pronged stability suite incorporates
principled router-gradient balancing, a hidden z-loss to suppress massive activations, and fine-tuned optimizer
configurations. | PARTIAL | 16 | Exploring the impact of the Root Mean Square (RMS) norm of gradients and epsilon on loss across different
model sizes. | third model (B_only) |
2,509.01322 | longcat_flash | 2025 | TRAIN | Deterministic computation ensures exact reproducibility and enables silent-data-corruption detection. | 4 | To enhance the reliability of large-scale cluster training, we introduce deterministic computation.
This guarantees the exact reproducibility of experiments and enables the detection of SDC (Silent Data Corruption)
during the training process. | NO | null | null | third model (B_only) |
2,501.08313 | minimax | 2025 | ARCH | Hybrid architecture: one softmax-attention block after every seven lightning-attention (transnormer) blocks | 2 | In the architecture, one transformer block with softmax attention follows every seven transnormer blocks (Qin et al., 2022a) with lightning attention. | YES | 12 | For every 8 consecutive layers in the base model, we systematically replace softmax attention with lightning attention in the first 7 layers. | both coders |
2,501.08313 | minimax | 2025 | TRAIN | Novel global router load-balancing strategy for MoE training stability | 4 | To ensure load balancing in the MoE blocks, we propose a novel load balancing strategy inspired by GShard (Lepikhin et al., 2021), which we refer to the global router. | NO | null | null | third model (A_only) |
2,501.08313 | minimax | 2025 | ARCH | Integrate lightning attention with MoE to maximize parameter and computational capacity. | 1 | To maximize computational capacity, we integrate it
with Mixture of Experts (MoE), creating a model with 32 experts and 456 billion total parameters, of | PARTIAL | 5 | a dense model with 7 billion parameters and an MoE model with 2 billion activation parameters
out of a total of 24 billion parameters. | third model (B_only) |
2,503.01743 | phi4_multimodal | 2025 | ARCH | Group query attention for more efficient long-sequence generation | 1 | as well as group query attention for more efficient long-sequence generation | NO | null | null | both coders |
2,503.01743 | phi4_multimodal | 2025 | ARCH | Novel mixture-of-LoRAs modality extension with modality-specific LoRAs and routers on a frozen base LM, unifying text, vision and speech in one model | 1 | Its novel modality extension approach leverages LoRA adapters and modality-specific routers to allow multiple inference modes combining various modalities without interference. | NO | null | null | third model (pair) |
2,512.24618 | youtu | 2025 | TRAIN | Commonsense-STEM-Agent multi-stage curriculum progressively shifting data distribution | 1 | By progressively shifting the pre-training data distribution from general commonsense to complex STEM and agentic tasks, we ensure the model acquires deep cognitive abilities | NO | null | null | both coders |
2,512.24618 | youtu | 2025 | ARCH | Dense Multi-Latent Attention (MLA) architecture for a lightweight 1.96B model | 1 | Built on a dense Multi-Latent Attention (MLA) architecture with a novel STEM-oriented vocabulary, Youtu-LLM supports a 128k context window. | YES | 16 | We compared GQA and MLA by pre-training 1B-parameter models from scratch on 500B tokens. | both coders |
2,512.24618 | youtu | 2025 | TRAIN | Principled agent-centric training paradigm (agentic mid-training) inducing native agentic capabilities | 3 | We propose a principled training paradigm that enhances native agentic capabilities through innovations in tokenizer design, data allocation, and multi-stage learning, guided by an agent-centric philosophy. | PARTIAL | 29 | Table 14. Performance of post-trained models with and without agentic mid-training (AMT). | both coders |
2,606.19348 | deepseek_v4 | 2026 | TRAIN | Muon optimizer for faster convergence and training stability | 4 | we introduce the Muon (Jordan et al., 2024; Liu et al., 2025) optimizer to the training of DeepSeek-V4 series, leading to faster convergence and improved training stability. | NO | null | null | both coders |
2,606.19348 | deepseek_v4 | 2026 | ARCH | Manifold-Constrained Hyper-Connections (mHC) replacing/upgrading conventional residual connections | 4 | To strengthen modeling capability, we incorporate Manifold-Constrained Hyper-Connections (mHC) (Xie et al., 2026) that upgrade conventional residual connections. | NO | null | null | both coders |
2,606.19348 | deepseek_v4 | 2026 | ARCH | Hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) for long-context efficiency | 1 | a hybrid attention architecture that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to improve long-context efficiency | NO | null | null | third model (A_only) |
The novelty exposed: 40 coded LLM technical reports
Dataset for the paper The novelty exposed: ablation, reproducibility and transparency in 40 technical reports of open-weight large language models (LLMs) (Bahaeddin Türkoğlu, 2026).
Code, protocol and raw coder outputs: https://github.com/bturkoglu/llm-novelty-exposed (archived: https://doi.org/10.5281/zenodo.24340455)
Configurations
| Config | Rows | Content |
|---|---|---|
claims (default) |
73 | Architectural/training novelty claims in 40 randomly sampled reports: category, claim, page, verbatim quote, ablation level (YES = controlled with/without comparison, PARTIAL = confounded, NO), supporting quote, and whether the value came from coder agreement or the resolver |
reports |
40 | One row per report: number of claims and their ablation levels, training runs stated, baselines rerun, scores copied, third-party evaluation, data mixture / token / compute disclosure, reporting rules met |
sampling_frame |
171 | Causal LM architectures in Hugging Face transformers (commit 0d7a6219): date added, linked arXiv report, eligibility and reason, stratum year, sampled or not |
coding/final_verified.json holds the full final coding with every quote checked against the report text by script.
from datasets import load_dataset
claims = load_dataset("bturkoglu/llm-novelty-exposed", "claims", split="train")
How it was made
40 reports were drawn at random (seed 20261114, stratified by year) from the eligible architectures. Two independent model coders from different vendors (Claude Opus 5.5, GPT-6.1 Sol) coded each report with the same pre-registered protocol; every quote was verified against the page-numbered text by script; disagreements were resolved by a third model family (Grok 4.7) under the same rule. Inter-coder agreement is reported in the paper (Table A1).
Limitations
The coding records what the reports state; it does not judge model quality, and an untested claim may still be true. Coders are language models, not human annotators. Report PDFs are not included; use the arXiv ids.
Licence
CC BY 4.0. Quotes are short excerpts from the cited reports; their copyright stays with the original authors.
Citation
@misc{turkoglu2026novelty,
author = {T{\"u}rko{\u{g}}lu, Bahaeddin},
title = {The novelty exposed: ablation, reproducibility and transparency in 40 technical reports of
open-weight large language models ({LLMs})},
year = {2026},
eprint = {2611.20003},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2611.20003}
}
- Downloads last month
- -