new

Get trending papers in your email inbox!

Subscribe

Daily Papers

byAK and the research community

Nov 20

Late Entry in Looped Language Models

Looped language models deepen reasoning at a compact parameter count by iterating a shared layer stack T times per token, turning recurrence depth directly into sequential decoding latency. Conventional early-exit decoding halts after initial passes without cache completion, truncating cached key-value states of past tokens to shallow depth. We study the opposite strategy, late entry, where new tokens skip initial passes and enter directly into the final K < T passes over a deep context cache. Once lightweight fine-tuning resolves hidden-state norm mismatch at the entry boundary, deep cached context substitutes for skipped token-local passes, allowing late entry to outperform matched early exit across model families and pass budgets. To turn this asymmetry into wall-clock speedups without stalling to build deep caches, candidate blocks verify through only the final K passes on the critical path while previously committed tokens complete passes 1 -> T alongside subsequent steps at zero extra serial calls. Projecting these finished T-pass wake states inside the existing output GEMM (Wake verification, governed by logit-margin tolerance tau in [0, inf]) unifies greedy-exact T-pass decoding (tau = 0, matching K = T block decoding within bf16 reduction order), margin-gated verification (tau > 0), and non-rollback lossy late entry (tau = inf) inside a single static CUDA graph. On Ouro-2.6B (T = 4), K = 2 Wake reaches 5.2x greedy-exact speedup at tau = 0 (+26% over K = 4 block decoding on GSM8K) and 6.8x at tau = inf (+16.9 points over trained early exit at matched speed). Scaling test-time recurrence across T in {8, 16, 32, 64} on Huginn-3.5B cuts K = T/4 rollbacks 21.5x with depth, driving greedy-exact (tau = 0) speedups to 6.1x (T = 32) and 8.4x (T = 64, +153% over K = 64 block) and tau = inf speedups to 10.6 to 12.5x.

FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness

Transformers are slow and memory-hungry on long sequences, since the time and memory complexity of self-attention are quadratic in sequence length. Approximate attention methods have attempted to address this problem by trading off model quality to reduce the compute complexity, but often do not achieve wall-clock speedup. We argue that a missing principle is making attention algorithms IO-aware -- accounting for reads and writes between levels of GPU memory. We propose FlashAttention, an IO-aware exact attention algorithm that uses tiling to reduce the number of memory reads/writes between GPU high bandwidth memory (HBM) and GPU on-chip SRAM. We analyze the IO complexity of FlashAttention, showing that it requires fewer HBM accesses than standard attention, and is optimal for a range of SRAM sizes. We also extend FlashAttention to block-sparse attention, yielding an approximate attention algorithm that is faster than any existing approximate attention method. FlashAttention trains Transformers faster than existing baselines: 15% end-to-end wall-clock speedup on BERT-large (seq. length 512) compared to the MLPerf 1.1 training speed record, 3times speedup on GPT-2 (seq. length 1K), and 2.4times speedup on long-range arena (seq. length 1K-4K). FlashAttention and block-sparse FlashAttention enable longer context in Transformers, yielding higher quality models (0.7 better perplexity on GPT-2 and 6.4 points of lift on long-document classification) and entirely new capabilities: the first Transformers to achieve better-than-chance performance on the Path-X challenge (seq. length 16K, 61.4% accuracy) and Path-256 (seq. length 64K, 63.1% accuracy).

  • 5 authors
·
May 27, 2022 4

You Only Judge Once: Multi-response Reward Modeling in a Single Forward Pass

We present a discriminative multimodal reward model that scores all candidate responses in a single forward pass. Conventional discriminative reward models evaluate each response independently, requiring multiple forward passes, one for each potential response. Our approach concatenates multiple responses with separator tokens and applies cross-entropy over their scalar scores, enabling direct comparative reasoning and efficient N-way preference learning. The multi-response design also yields up to Ntimes wall-clock speedup and FLOPs reduction over conventional single-response scoring. To enable N-way reward evaluation beyond existing pairwise benchmarks, we construct two new benchmarks: (1) MR^2Bench-Image contains human-annotated rankings over responses from 8 diverse models; (2) MR^2Bench-Video is a large-scale video-based reward benchmark derived from 94K crowdsourced pairwise human judgments over video question-answering spanning 19 models, denoised via preference graph ensemble. Both benchmarks provide 4-response evaluation variants sampled from the full rankings. Built on a 4B vision-language backbone with LoRA fine-tuning and a lightweight MLP value head, our model achieves state-of-the-art results on six multimodal reward benchmarks, including MR^2Bench-Image, MR^2Bench-Video, and four other existing benchmarks. Our model outperforms existing larger generative and discriminative reward models. We further demonstrate that our reward model, when used in reinforcement learning with GRPO, produces improved policy models that maintain performance across standard multimodal benchmarks while substantially improving open-ended generation quality, outperforming a single-response discriminative reward model (RM) baseline by a large margin in both training stability and open-ended generation quality.

X-Cache: Cross-Chunk Block Caching for Few-Step Autoregressive World Models Inference

Real-time world simulation is becoming a key infrastructure for scalable evaluation and online reinforcement learning of autonomous driving systems. Recent driving world models built on autoregressive video diffusion achieve high-fidelity, controllable multi-camera generation, but their inference cost remains a bottleneck for interactive deployment. However, existing diffusion caching methods are designed for offline video generation with multiple denoising steps, and do not transfer to this scenario. Few-step distilled models have no inter-step redundancy left for these methods to reuse, and sequence-level parallelization techniques require future conditioning that closed-loop interactive generation does not provide. We present X-Cache, a training-free acceleration method that caches along a different axis: across consecutive generation chunks rather than across denoising steps. X-Cache maintains per-block residual caches that persist across chunks, and applies a dual-metric gating mechanism over a structure- and action-aware block-input fingerprint to independently decide whether each block should recompute or reuse its cached residual. To prevent approximation errors from permanently contaminating the autoregressive KV cache, X-Cache identifies KV update chunks (the forward passes that write clean keys and values into the persistent cache) and unconditionally forces full computation on these chunks, cutting off error propagation. We implement X-Cache on X-world, a production multi-camera action-conditioned driving world model built on multi-block causal DiT with few-step denoising and rolling KV cache. X-Cache achieves 71% block skip rate with 2.6x wall-clock speedup while maintaining minimum degradation.

  • 13 authors
·
May 4

vla-eval: A Unified Evaluation Harness for Vision-Language-Action Models

Vision-Language-Action (VLA) models are increasingly evaluated across multiple simulation benchmarks, yet adding each benchmark to an evaluation pipeline requires resolving incompatible dependencies, matching underspecified evaluation protocols, and reverse-engineering undocumented preprocessing. This burden scales with the number of models and benchmarks, making comprehensive evaluation impractical for most teams. We present vla-eval, an open-source evaluation harness that eliminates this per-benchmark cost by decoupling model inference from benchmark execution through a WebSocket+msgpack protocol with Docker-based environment isolation. Models integrate once by implementing a single predict() method; benchmarks integrate once via a four-method interface; the full cross-evaluation matrix works automatically. The framework supports 14 simulation benchmarks and six model servers. Parallel evaluation via episode sharding and batch inference achieves up to 47x wall-clock speedup, completing 2,000 LIBERO episodes in ~18 minutes. To validate the framework, we reproduce published scores across six VLA codebases and three benchmarks, documenting previously undocumented pitfalls. We additionally release a VLA leaderboard aggregating 657 published results across 17 benchmarks. Framework, evaluation configs, and all reproduction results are publicly available at https://github.com/allenai/vla-evaluation-harness and https://allenai.github.io/vla-evaluation-harness/leaderboard.

  • 7 authors
·
Apr 16

DiffuPlex: Accelerating Full-Duplex Spoken Dialog Models via Rolling Masked Diffusion

Recent full-duplex spoken dialog models enable simultaneous listening and speaking, but fine-grained models still advance their backbone autoregressively at every interaction frame. We introduce DiffuPlex, a rolling masked diffusion framework that reduces this sequential computation by predicting multiple future user and assistant frames in a single backbone wake. DiffuPlex consumes only a confident prefix of each predicted future while interaction continues at the original frame rate. As user speech arrives, it checks the corresponding user predictions and, when the interaction diverges, preserves already played assistant content while revising only the unplayed future. We consider two inference policies over the same predictor: DiffuPlex-LISTEN consumes multiple future frames when they predict assistant silence, whereas DiffuPlex-SPEAK can also consume predicted assistant speech. Across full-duplex interaction and spoken-language evaluations, DiffuPlex substantially reduces sequential backbone computation while largely preserving interaction behavior and general capability. DiffuPlex-LISTEN and DiffuPlex-SPEAK achieve 1.46times and 1.59times deployment-path wall-clock speedups and 1.61times and 1.80times Core LM speedups, with all measured backbone invocations completing within the 80ms interaction interval. Human evaluation shows that LISTEN preserves speech naturalness and conversational quality, while SPEAK retains conversational quality with some degradation in speech naturalness.

  • 1 authors
·
Oct 8

DiffuSpec: Unlocking Diffusion Language Models for Speculative Decoding

As large language models (LLMs) scale up, accuracy improves, but the autoregressive (AR) nature of decoding increases latency since each token requires a serial forward pass. Speculative decoding addresses this by employing a fast drafter to propose multi-token drafts, which are then verified in parallel by the target model. However, many deployments still rely on AR drafters, where sequential passes limit wall-clock gains. We revisit the drafting stage and present DiffuSpec, a training-free drop-in framework that uses a pretrained diffusion language model (DLM) to produce multi-token drafts in a single forward pass, while remaining compatible with standard AR verifiers. Because DLM drafts are generated under bidirectional conditioning, parallel per-position candidates form a token lattice in which the locally highest-probability token at each position need not form a causal left-to-right path. Moreover, DLM drafting requires pre-specifying a draft length, inducing a speed-quality trade-off. To address these challenges, we introduce two practical components: (i) a causal-consistency path search (CPS) over this lattice that extracts a left-to-right path aligned with AR verification; and (ii) an adaptive draft-length (ADL) controller that adjusts next proposal size based on recent acceptance feedback and realized generated length. Across benchmarks, DiffuSpec yields up to 3x wall-clock speedup, establishing diffusion-based drafting as a robust alternative to autoregressive drafters for speculative decoding.

  • 7 authors
·
Sep 28, 2025

DPA4: Pushing the Accuracy-Cost Frontier of Interatomic Potentials with EMFA SO(2) Convolution

Machine-learning interatomic potentials now approach quantum-mechanical accuracy on standard benchmarks, but the training cost of the most expressive equivariant architectures has become a serious bottleneck. We introduce DPA4, an SE(3)-equivariant interatomic-potential architecture with an EMFA (Edge-conditioned, Multi-Focus, Attention) SO(2)-equivariant convolution that combines a low-rank edge-node SO(2)-equivariant product, a multi-focus design for message nonlinearity, and envelope-gated attention for message aggregation. A Lebedev-grid projection further preserves SO(3)-equivariance in the nonlinearity to machine precision. A compiler-friendly conservative energy-gradient training path provides up to sim3 times wall-clock speedup under torch compile. On the compliant Matbench Discovery benchmark, DPA4-Pro attains the best Combined Performance Score (CPS) on the leaderboard, while the 2.76M-parameter DPA4-Air exceeds the accuracy of the 30.1M-parameter eSEN-30M-MP baseline with 10.9times fewer parameters and 42.9times less training compute. On SPICE-MACE-OFF, the 5.4M-parameter DPA4-Plus lowers the aggregate molecular energy and force errors of the 6.5M-parameter eSEN baseline by 29% and 30%, while the 2.7M-parameter DPA4-Air still surpasses that baseline with sim2.4times fewer parameters. Together these results place DPA4 on a new accuracy-cost Pareto frontier on Matbench Discovery and position it as a strong candidate backbone for future multi-task large atomistic model (LAM) pretraining.

  • 7 authors
·
Jun 9

R2R: Efficiently Navigating Divergent Reasoning Paths with Small-Large Model Token Routing

Large Language Models (LLMs) achieve impressive reasoning capabilities at the cost of substantial inference overhead, posing substantial deployment challenges. Although distilled Small Language Models (SLMs) significantly enhance efficiency, their performance suffers as they fail to follow LLMs' reasoning paths. Luckily, we reveal that only a small fraction of tokens genuinely diverge reasoning paths between LLMs and SLMs. Most generated tokens are either identical or exhibit neutral differences, such as minor variations in abbreviations or expressions. Leveraging this insight, we introduce **Roads to Rome (R2R)**, a neural token routing method that selectively utilizes LLMs only for these critical, path-divergent tokens, while leaving the majority of token generation to the SLM. We also develop an automatic data generation pipeline that identifies divergent tokens and generates token-level routing labels to train the lightweight router. We apply R2R to combine R1-1.5B and R1-32B models from the DeepSeek family, and evaluate on challenging math, coding, and QA benchmarks. With an average activated parameter size of 5.6B, R2R surpasses the average accuracy of R1-7B by 1.6x, outperforming even the R1-14B model. Compared to R1-32B, it delivers a 2.8x wall-clock speedup with comparable performance, advancing the Pareto frontier of test-time scaling efficiency. Our code is available at https://github.com/thu-nics/R2R.

  • 9 authors
·
May 27, 2025 2

CD4LM: Consistency Distillation and aDaptive Decoding for Diffusion Language Models

Autoregressive large language models achieve strong results on many benchmarks, but decoding remains fundamentally latency-limited by sequential dependence on previously generated tokens. Diffusion language models (DLMs) promise parallel generation but suffer from a fundamental static-to-dynamic misalignment: Training optimizes local transitions under fixed schedules, whereas efficient inference requires adaptive "long-jump" refinements through unseen states. Our goal is to enable highly parallel decoding for DLMs with low number of function evaluations while preserving generation quality. To achieve this, we propose CD4LM, a framework that decouples training from inference via Discrete-Space Consistency Distillation (DSCD) and Confidence-Adaptive Decoding (CAD). Unlike standard objectives, DSCD trains a student to be trajectory-invariant, mapping diverse noisy states directly to the clean distribution. This intrinsic robustness enables CAD to dynamically allocate compute resources based on token confidence, aggressively skipping steps without the quality collapse typical of heuristic acceleration. On GSM8K, CD4LM matches the LLaDA baseline with a 5.18x wall-clock speedup; across code and math benchmarks, it strictly dominates the accuracy-efficiency Pareto frontier, achieving a 3.62x mean speedup while improving average accuracy. Code is available at https://github.com/yihao-liang/CDLM

  • 10 authors
·
Jan 5

HiAR: Efficient Autoregressive Long Video Generation via Hierarchical Denoising

Autoregressive (AR) diffusion offers a promising framework for generating videos of theoretically infinite length. However, a major challenge is maintaining temporal continuity while preventing the progressive quality degradation caused by error accumulation. To ensure continuity, existing methods typically condition on highly denoised contexts; yet, this practice propagates prediction errors with high certainty, thereby exacerbating degradation. In this paper, we argue that a highly clean context is unnecessary. Drawing inspiration from bidirectional diffusion models, which denoise frames at a shared noise level while maintaining coherence, we propose that conditioning on context at the same noise level as the current block provides sufficient signal for temporal consistency while effectively mitigating error propagation. Building on this insight, we propose HiAR, a hierarchical denoising framework that reverses the conventional generation order: instead of completing each block sequentially, it performs causal generation across all blocks at every denoising step, so that each block is always conditioned on context at the same noise level. This hierarchy naturally admits pipelined parallel inference, yielding a 1.8 wall-clock speedup in our 4-step setting. We further observe that self-rollout distillation under this paradigm amplifies a low-motion shortcut inherent to the mode-seeking reverse-KL objective. To counteract this, we introduce a forward-KL regulariser in bidirectional-attention mode, which preserves motion diversity for causal inference without interfering with the distillation loss. On VBench (20s generation), HiAR achieves the best overall score and the lowest temporal drift among all compared methods.

SNLP: Layer-Parallel Inference via Structured Newton Corrections

Autoregressive language models execute Transformer layers sequentially, creating a latency bottleneck that is not removed by conventional tensor or pipeline parallelism. We study whether this layerwise dependency can be relaxed by treating the hidden-state trace across layers as the solution of a nonlinear residual equation and solving it with parallel Newton-style updates. While this view is principled, exact Newton corrections require expensive Jacobian-vector products and naive fixed-point iterations are unstable on trained Transformers. We introduce Structured Newton Layer Parallelism (SNLP), a training and inference framework that replaces exact layer Jacobians with cheap architecture-induced surrogate dynamics. In residual Transformers, this yields Identity Newton (IDN), where the correction reduces to a prefix-sum-like update; in mHC-style architectures, HC Newton (HCN) uses the model's residual mixing matrix. We further introduce SNLP-aware regularization, which trains models to make one or a few structured Newton iterations accurately approximate the sequential forward. Experiments on nanochat-scale Transformers show that SNLP regularization improves layer-parallel compatibility and can also improve standard sequential perplexity, reducing baseline PPL by 4.7%-23.4%. At inference time, SNLP combined with layer fusion and chunkwise decomposition achieves practical wall-clock speedups: on a 0.5B Nanochat model, it reaches 2.3x speedup while still improving PPL by 6.1%. These results suggest that layer-parallel inference is not merely a numerical approximation to sequential execution, but can act as a useful solver-induced inference bias. We also characterize limitations: off-the-shelf pretrained models are less amenable to this procedure, and exact convergence recovers the sequential computation rather than providing monotonic inference-time scaling.

RedHatAI Red Hat AI
·
May 17 1

AuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing

We introduce AuK, an open-source foundational model that unifies speech generation and editing through a common interface of natural-language instructions and audio context. To support this broad capability set, we construct approximately 3.03 billion instruction--audio instances and 1.95 million hours of effective supervision across five task families: speech generation, content editing, enhancement and separation, paralinguistic editing, and acoustic editing. AuK combines a multimodal large language model for semantic conditioning, an VAE jointly trained on speech, general audio, and music for acoustic conditioning, and a hybrid rectified-flow Transformer that performs dual-stream MMDiT blocks followed by unified single-stream DiT blocks for generation. Training begins with generation-only warm-up and proceeds to joint generation--editing pre-training. We then apply complementary post-training strategies: human-feedback preference optimization for open-ended editing and reward-based reinforcement learning for speech generation. To reduce inference cost, we further distill the model with consistency initialization and task-routed Decoupled DMD. The resulting AuK-Flash performs 4-step inference without classifier-free guidance and achieves a 4.5 wall-clock speedup over the full model under matched conditions. Experiments demonstrate leading performance on zero-shot and instruction-controlled speech generation and general instruction-guided editing, while remaining competitive on signal-level restoration tasks. We release both the source code and model weights to support reproducibility and further research.

TS-DP: Reinforcement Speculative Decoding For Temporal Adaptive Diffusion Policy Acceleration

Diffusion Policy (DP) excels in embodied control but suffers from high inference latency and computational cost due to multiple iterative denoising steps. Existing diffusion model (DMs) acceleration methods improve inference speed by reducing computational overhead, as DMs for image generation are computation-bound. In this work, we reveal that DP are I/O bound, and thus propose using speculative decoding, a parallel computing framework, to accelerate DP for the first time. However, it is non-trivial to address the following challenges: (1) How to preserve the denoising quality of the base model at reduced computational cost under time-varying task difficulty in embodied settings; and (2) How to dynamically allocate computation according to task difficulty in such environments. Therefore, we propose Temporal-aware Reinforcement-based Speculative Diffusion Policy (TS-DP), the first framework that enables speculative decoding for DP with temporal adaptivity. First, to handle dynamic environments where task difficulty varies over time, we distill a Transformer-based drafter to imitate the base model and replace its costly denoising calls. Second, an RL-based scheduler further adapts to the time-varying task difficulty by adjusting speculative parameters to maintain accuracy while improving efficiency. Extensive experiments show that TS-DP achieves up to a 4.17x NFE reduction and a 3.2x wall-clock speedup with a draft acceptance rate of over 85% in simulation. Real-world experiments further demonstrate a 3.8x NFE reduction while improving success rates on real-time tasks (e.g., a 23% gain on the conveyor-belt task), reaching an inference frequency of 26 Hz and enabling real-time control while maintaining comparable task performance. The code is available at https://github.com/ChildTang/TS-DP.

  • 12 authors
·
Oct 26

MiniMax Sparse Attention

Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale. We introduce MiniMax Sparse Attention (MSA), a blockwise sparse attention built upon Grouped Query Attention (GQA). A lightweight Index Branch scores key-value blocks and independently selects a Top-k subset for each GQA group, enabling group-specific sparse retrieval while maintaining efficient block-level execution; the Main Branch then performs exact block-sparse attention over only the selected blocks. Designed around a principle of simplicity and scalability, MSA is deliberately streamlined, making it straightforward to deploy efficiently across a broad range of GPUs. To translate sparsity into practical speedups, we co-design MSA with a GPU execution path that uses exp-free Top-k selection and KV-outer sparse attention to improve tensor-core utilization under block-granular access. On a 109B-parameter model with native multimodal training, MSA performs on par with GQA while reducing per-token attention compute by 28.4x at 1M context. Paired with our co-designed kernel, MSA achieves 14.2x prefill and 7.6x decoding wall-clock speedups on H800. Our inference kernel is available at: https://github.com/MiniMax-AI/MSA. A production-grade natively multimodal model powered by MSA has been publicly released at: https://huggingface.co/MiniMaxAI/MiniMax-M3.

MiniMaxAI MiniMax
·
Jun 10 7

Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion

Autoregressive (AR) video diffusion is a powerful paradigm for streaming and interactive video generation. However, its reliance on softmax self-attention leads to quadratic compute complexity in sequence length and memory usage due to key-value caching, which limits its scalability to long video horizons. Existing remedies (e.g., sparse attention and KV-cache compression) reduce per-step cost but still rely on a linearly growing cache or irreversibly discard past context, and thus fail to address linear memory growth and streaming context management. To address this scalability bottleneck, we propose ARL2 (Attend Locally, Remember Linearly), a hybrid attention module that replaces quadratic cross-frame attention with a fixed-size recurrent state. We decompose self-attention into two branches: an intra-frame softmax branch for spatial detail and local dependencies, and an inter-frame gated recurrent linear branch that maintains a fixed-size state for streaming context. Our key insight is that softmax attention captures fine-grained local interactions, while a recurrent state provides controllable long-range memory. This design achieves linear-time scaling with constant memory while improving temporal consistency over the full-softmax model. To prevent noisy intermediate states from corrupting memory, we update the recurrent state only after the denoised pass. To avoid within-frame information asymmetry, all tokens share the same pre-update state rather than sequential updates. To the best of our knowledge, this is the first work to convert a pretrained AR video diffusion model into a hybrid linear attention architecture, through an efficient two-stage training scheme for AR video. With 75% of layers replaced by hybrid linear attention, the model achieves up to 2.26 wall-clock speedup and 54% memory reduction, while maintaining comparable quality with improving temporal consistency.

  • 3 authors
·
May 19

SpecTr: Fast Speculative Decoding via Optimal Transport

Autoregressive sampling from large language models has led to state-of-the-art results in several natural language tasks. However, autoregressive sampling generates tokens one at a time making it slow, and even prohibitive in certain tasks. One way to speed up sampling is speculative decoding: use a small model to sample a draft (block or sequence of tokens), and then score all tokens in the draft by the large language model in parallel. A subset of the tokens in the draft are accepted (and the rest rejected) based on a statistical method to guarantee that the final output follows the distribution of the large model. In this work, we provide a principled understanding of speculative decoding through the lens of optimal transport (OT) with membership cost. This framework can be viewed as an extension of the well-known maximal-coupling problem. This new formulation enables us to generalize the speculative decoding method to allow for a set of k candidates at the token-level, which leads to an improved optimal membership cost. We show that the optimal draft selection algorithm (transport plan) can be computed via linear programming, whose best-known runtime is exponential in k. We then propose a valid draft selection algorithm whose acceptance probability is (1-1/e)-optimal multiplicatively. Moreover, it can be computed in time almost linear with size of domain of a single token. Using this new draft selection algorithm, we develop a new autoregressive sampling algorithm called SpecTr, which provides speedup in decoding while ensuring that there is no quality degradation in the decoded output. We experimentally demonstrate that for state-of-the-art large language models, the proposed approach achieves a wall clock speedup of 2.13X, a further 1.37X speedup over speculative decoding on standard benchmarks.

  • 6 authors
·
Oct 23, 2023

Fast and Accurate Causal Parallel Decoding using Jacobi Forcing

Multi-token generation has emerged as a promising paradigm for accelerating transformer-based large model inference. Recent efforts primarily explore diffusion Large Language Models (dLLMs) for parallel decoding to reduce inference latency. To achieve AR-level generation quality, many techniques adapt AR models into dLLMs to enable parallel decoding. However, they suffer from limited speedup compared to AR models due to a pretrain-to-posttrain mismatch. Specifically, the masked data distribution in post-training deviates significantly from the real-world data distribution seen during pretraining, and dLLMs rely on bidirectional attention, which conflicts with the causal prior learned during pretraining and hinders the integration of exact KV cache reuse. To address this, we introduce Jacobi Forcing, a progressive distillation paradigm where models are trained on their own generated parallel decoding trajectories, smoothly shifting AR models into efficient parallel decoders while preserving their pretrained causal inference property. The models trained under this paradigm, Jacobi Forcing Model, achieves 3.8x wall-clock speedup on coding and math benchmarks with minimal loss in performance. Based on Jacobi Forcing Models' trajectory characteristics, we introduce multi-block decoding with rejection recycling, which enables up to 4.5x higher token acceptance count per iteration and nearly 4.0x wall-clock speedup, effectively trading additional compute for lower inference latency. Our code is available at https://github.com/hao-ai-lab/JacobiForcing.

  • 8 authors
·
Dec 16, 2025 2

Verification-Aware Training for Speculative Decoding

Speculative decoding accelerates large language model inference by using a draft model to generate candidate tokens, which are verified by the target model in a single forward pass. Verification proceeds sequentially and discards every position from the first rejection onward, yet existing draft training relies on token-level imitation of the target with a fixed per-position weighting that reflects neither property. We introduce Verification-Aware Training (VAT), a plug-in framework that simulates verification at every training step and turns the resulting accept and reject patterns into supervision. VAT consists of two components: (i) a verification head, a lightweight jointly trained binary classifier that supervises the draft model on whether each position survives sequential verification; (ii) verification-adaptive weighting, which replaces the fixed weighting schedule by keeping full weight up to each sample's first rejection point and re-anchoring the decay to start there. VAT modifies only the training objective, so it can be layered on top of existing methods without changing the draft architecture, the target model, or the inference procedure. Applied to EAGLE-3 and DFlash on Qwen3-4B, Qwen3-8B, and LLaMA-3.1-8B, VAT improves average acceptance length by up to 11.4% and wall-clock speedup by up to 8.7%, with consistent gains across math, code, and chat benchmarks. Code will be available at https://github.com/naver-ai/vat

naver-ai NAVER AI Lab
·
Aug 30 2

Scheduling Recursive Reasoning in Looped Transformers

Recurrent reasoning models have attracted growing attention for scaling test-time computation, typically by iteratively refining latent states with shared parameters. However, these models apply each learned update with a fixed unit scale, which can be conservative when updates make persistent progress and overly aggressive when they fluctuate, limiting the benefit of additional loops. To understand how the scale should vary along the trajectory, we first analyze the sensitivity of terminal loss to recurrent update scale. We show that its temporal average admits an exact decomposition into persistent-progress and centered-fluctuation contributions. Based on this, we introduce the Trajectory Adaptive Progress-Fluctuation Scheduler (TAPS), which tracks their balance across recurrent updates and adapts the step size online. Theoretically, we establish sufficient conditions under which TAPS reduces expected terminal loss and reaches a target quality in fewer recurrent loops. Empirically, we show that TAPS improves terminal accuracy across structured reasoning tasks without retraining. By further incorporating the progress-fluctuation principle into training, TAPS yields additional accuracy gains with up to 1.56 times wall-clock speedup at matched baseline accuracy. The broad applicability of TAPS is supported by its effectiveness across diverse recurrent architectures and inference strategies. Together, these results establish update scale as complementary control axis of recurrent inference alongside architecture and depth.

DLoop: Looped Speculative Decoding

Speculative decoding accelerates autoregressive generation in large language models. In each drafting stage, a lightweight draft model proposes tokens that the target model subsequently verifies. With increasingly capable draft models, we find that the target model frequently accepts all tokens produced in a drafting stage. A verification nevertheless follows each drafting stage, resulting in unnecessary target-model forward passes even when drafting could have continued. Adaptive draft length methods decide during decoding how many draft tokens precede a verification, but they raise the speedup only for autoregressive draft models. For a parallel draft model, drafting further requires target-model hidden states for draft tokens that have not been verified. We propose DLoop, a looped form of speculative decoding that adaptively performs multiple drafting stages before verification. DLoop continues drafting while the draft model remains confident and verifies all accumulated draft tokens together. Loop-aware training keeps the draft model reliable in the additional drafting stages by exposing it to its own hidden states for unverified draft tokens. By spending additional draft-model forward passes, DLoop reduces the number of target-model forward passes required for verification. Across diverse speculative decoding methods including EAGLE-3, DFlash, Domino, DSpark, and multi-token prediction modules, DLoop improves the wall-clock speedup by 5 to 41 percent while preserving lossless decoding. Code will be available at https://github.com/naver-ai/DLoop.

naver-ai NAVER AI Lab
·
Oct 5 2

Prefix-Guided On-Policy Distillation: Mining Golden Trajectories from Rollouts

On-policy distillation (OPD) improves reasoning models by applying dense teacher supervision on student-sampled trajectories. However, scaling OPD to long-horizon reasoning exposes a reliability and efficiency problem: standard OPD assigns every candidate the same long rollout budget, even though some trajectories may quickly become weakly aligned with the teacher and provide less useful supervision. Prior analyses suggest that teacher--student compatibility is important for OPD success, motivating early-prefix top-k overlap as a proxy for continuation value. Continuing low-overlap trajectories to the full rollout length may cause them to drift further from the teacher as generation proceeds, increasing computational cost while providing limited distillation benefit. To address this, we introduce Prefix-Guided On-Policy Distillation (PG-OPD), a rollout-allocation framework that estimates continuation value from fixed-length prefixes. PG-OPD computes teacher-student top-k overlap in an early probe window and allocates long rollouts only to high-overlap candidates, while stopping the rest at the prefix length. Across teacher--student combinations on AMC, AIME, and HMMT benchmarks, PG-OPD achieves up to a 4.80-point accuracy gain and up to a 2.46x wall-clock speedup across configurations. These results demonstrate that using early-prefix compatibility to guide candidate pruning can improve both training efficiency and reasoning performance.

  • 5 authors
·
Jun 20

Flux Attention: Context-Aware Hybrid Attention for Efficient LLMs Inference

The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8timesA800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to 2.8times and 2.0times in the prefill and decode stages.

DUET: Optimize Token-Budget Allocation for Reinforcement Learning with Verifiable Rewards

Reinforcement learning with verifiable rewards (RLVR) generates hundreds of thousands of tokens per training step, with rollout generation dominating the computational cost. The overall token budget can be controlled along two main dimensions: (i) deciding which prompts to allocate rollouts to, and (ii) deciding how long each rollout should be. Prior work has generally controlled only one of these dimensions at a time. We show that jointly tuning both decisions under a shared compute budget improves both reasoning quality and wall-clock training time. We instantiate this view as DUal-controlled tokEn allocaTion (DUET), a computationally efficient layer over GRPO that uses a lightweight pre-rollout surrogate of prompt informativeness to set how many rollouts each prompt receives, and a marker-gated abort rule with importance reweighting to set when to stop them. On Qwen3-1.7B trained on MATH, DUET outperforms full-budget GRPO and the other three budget-aware baseline methods. DUET's advantage further generalizes to other benchmarks across math and coding, and is on par with the best baseline on the scientific Q\&A domain, while also achieving a 1.62times wall-clock speedup. More notably, using only 50\% of the token budget, DUET still outperforms all baseline methods at their full budget, achieving an even higher 2.51times speedup over full-budget GRPO. We verify the high performance of DUET on other backbone LLMs, including Qwen3-4B and Llama-3.2-3B-Instruct. Notably, the gap between DUET and the strongest baseline widens as the budget tightens, contrary to the usual pattern in which efficient methods trade off quality as compute decreases. More broadly, these results suggest that DUET budget-aware control strategies are valuable not only for accelerating training, but also for improving the quality of the learning signal.

  • 4 authors
·
May 7

DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding

Block diffusion speculative decoding accelerates LLM inference by predicting all tokens within a block simultaneously for the target model to verify in parallel. Predicting an entire block at once requires a sufficiently capable draft model and effective utilization of the target model's internal knowledge. However, the state-of-the-art method DFlash constrains all draft layers to share a single fused representation derived from only a few target layers, limiting per-layer expressiveness and hindering further scaling of draft capacity. In this paper, we present \modelname, which flares out the narrow conditioning bottleneck of DFlash through a lightweight layer-wise fusion mechanism: each draft layer attends to its own learnable combination of a broad set of target layers at negligible overhead, simultaneously injecting richer target knowledge and providing every draft layer with a distinct input. This enhanced per-layer expressiveness enables scaling the draft model to deeper architectures with consistent gains. We further scale training data from 800K to 2.4M samples to fully exploit the enlarged capacity. On six benchmarks spanning mathematical reasoning, code generation, and conversation, \modelname attains average wall-clock speedups of 5.52x on Qwen3-4B, 5.46x on Qwen3-8B, and 3.91x on GPT-OSS-20B, improving over DFlash by roughly 11\%, 8\%, and 5\% respectively. Our code is available at https://github.com/Tencent/AngelSlim.

  • 12 authors
·
Jun 1

CURE: Local Uncertainty Repair for Block-Parallel Speculative Decoding

Speculative decoding mitigates the latency of sequential generation in autoregressive Large Language Models (LLMs) by interleaving draft generation with target verification. However, existing parallel drafting backends often suffer from rapid accuracy degradation over long horizons, leading to high rejection rates during verification and suboptimal wall-clock speedups. We observe that drafting errors are not uniformly distributed but typically stem from localized high-uncertainty tokens that destabilize downstream generation trajectories. Motivated by this token error pattern, we propose CURE, a budget-aware dynamic repair tree designed to repair errors at uncertainty focal points without incurring prohibitive tree-verification overheads. Specifically, our method uses predictive confidence margins to dynamically locate candidate error tokens within a block-parallel draft, expands bounded repair paths only at these fragile nodes, and employs a novel repair resynchronization mechanism to realign draft states post-verification. Evaluations on code-generation benchmarks (HumanEval, MBPP, and LiveCodeBench-lite) and mathematical reasoning benchmark (GSM8K) demonstrate that CURE increases the average accepted length by 4.2-7.5% over parallel baselines without repair, translating to an end-to-end speedup of 2.66-3.49times over target-only decoding. Furthermore, we provide a plug-and-play repair module compatible with standard parallel drafting frameworks. We also characterize the trade-off between draft compute and verification efficiency.

  • 4 authors
·
Jul 31

FastDraft: How to Train Your Draft

Speculative Decoding has gained popularity as an effective technique for accelerating the auto-regressive inference process of Large Language Models (LLMs). However, Speculative Decoding entirely relies on the availability of efficient draft models, which are often lacking for many existing language models due to a stringent constraint of vocabulary incompatibility. In this work we introduce FastDraft, a novel and efficient approach for pre-training and aligning a draft model to any large language model by incorporating efficient pre-training, followed by fine-tuning over synthetic datasets generated by the target model. We demonstrate FastDraft by training two highly parameter efficient drafts for the popular Phi-3-mini and Llama-3.1-8B models. Using FastDraft, we were able to produce a draft with approximately 10 billion tokens on a single server with 8 Intel^circledR Gaudi^circledR 2 accelerators in under 24 hours. Our results show that the draft model achieves impressive results in key metrics of acceptance rate, block efficiency and up to 3x memory bound speed up when evaluated on code completion and up to 2x in summarization, text completion and instruction tasks. We validate our theoretical findings through benchmarking on the latest Intel^circledR Core^{tiny TM} Ultra, achieving a wall-clock time speedup of up to 2x, indicating a significant reduction in runtime. Due to its high quality, FastDraft unlocks large language models inference on AI-PC and other edge-devices.

  • 4 authors
·
Nov 17, 2024

Learning to Accelerate Vision-Language-Action Models through Adaptive Visual Token Caching

Vision-Language-Action (VLA) models have demonstrated remarkable generalization capabilities in robotic manipulation tasks, yet their substantial computational overhead remains a critical obstacle to real-world deployment. Improving inference efficiency is therefore essential for practical robotic applications. Existing acceleration methods often rely on heuristic or static strategies--such as rule-based token caching or pruning--that are decoupled from task objectives and fail to adapt to dynamic scene changes. In this work, we reformulate inference acceleration as a learnable policy optimization problem and propose a novel framework that integrates a dynamic, task-aware decision-making process directly into the VLA model. At its core are two lightweight, cooperative modules: a Cached Token Selector, which determines which tokens should be reused, and a Cache Ratio Predictor, which controls how many tokens to reuse. Training these modules is non-trivial due to their discrete decisions. We address this by adopting a differentiable relaxation that allows gradient-based end-to-end optimization. Extensive experiments on the LIBERO and SIMPLER benchmarks, as well as real-robot evaluations, show that our method achieves a 1.76x wall-clock inference speedup while simultaneously improving the average success rate by 1.9 percentage points (from 75.0% to 76.9%) on LIBERO and by 5.0 percentage points on real-world tasks, significantly outperforming existing baselines. This work highlights the potential of learning task-aware computational allocation policies, paving the way for VLA models that are both powerful and efficient.

  • 8 authors
·
Jan 31

Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model

Standard vision-language models (VLMs) suffer from Moravec's paradox: they excel at complex offline visual reasoning but struggle with simple streaming perception tasks and process them inefficiently. We present Mage-VL, an efficient codec-native streaming foundation model for real-time multimodal understanding and interaction. At its core, our custom tokenizer, Mage-ViT, replaces uniform frame sampling by selectively encoding dynamic, entropy-rich regions using motion vectors and residual energy across sparse anchor (I) and predicted (P) frames. Operating at a 16 x 16 patch level, this reduces visual token consumption by over 75% while preserving spatiotemporal context. Trained from scratch on approximately 560M unlabeled images and 100M unlabeled video frames, Mage-ViT matches or outperforms flagship encoders trained on billions of image-text pairs. We establish AI4AI data pipelines encompassing prompt-code joint optimization for multimodal captioning and AI-driven performance diagnosis to guide training recipes. Furthermore, through a bio-inspired dual-system architecture - a lightweight System 1 event gate and a causal System 2 decoder - Mage-VL enables proactive streaming perception. Extensive evaluations show that Mage-VL-4B matches Qwen3-VL-4B on static tasks while achieving strong gains in video understanding and 2D/3D spatial reasoning, with up to a 3.5x wall-clock inference speedup, and comprehensively surpasses the 15B Phi-4-reasoning-vision baseline. Beyond model artifacts, we deliver seven key empirical findings covering pre-training data efficiency, variable-resolution scaling, codec system acceleration, VideoQA SFT redundancy, motion-spatial synergy, AI4AI data pipelines, and Zero-Vision SFT for multimodal RL.

microsoft Microsoft
·
Jul 26 2

SparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference

The memory-bound nature of the decoding stage of large language model (LLM) inference incurs significant latency. Layer-wise training-free network pruning approaches guided by the Hessian have been a prominent solution to this problem, as pruning reduces the number of nonzero parameters read from memory during decoding. Nevertheless, typical methods in this line compute the Hessian using pre-collected natural sequences, whereas the model is fed self-generated tokens during decoding, creating a distribution shift between the two sequences. The Hessian calculated on the natural sequence is different from that calculated on the generated sequence. We observe that this discrepancy causes the activation distribution during generation to deviate from that used for pruning, further hurting the pruned model performance. Moreover, most existing LLM pruning methods that bring actual speedup primarily target the sparse matrix-matrix (SpMM) multiplication, providing limited support for the sparse matrix-vector (SpMV) operations, which dominate decoding. To solve these problems, we introduce SparseDecoding, a principled decoding-aware pruning framework tailored for accurate and efficient LLM decoding. Specifically, at the algorithmic axis, SparseDecoding constructs calibration matrices from layer-wise activations collected during the dense-model autoregressive generation, excluding prefill, thereby aligning the pruning objective with the decoding activations. At the system axis, we develop an optimized N:M sparse matrix-vector kernel with bitmask indexing and fixed-step traversal. Substantial empirical results on representative LLMs (Llama-3.1-8B, Llama-3.3-70B, Qwen3-14B / 32B) demonstrate that our method consistently outperforms standard fixed-text calibration on the long-form generation benchmarks while achieving up to 1.48x end-to-end wall-clock decoding speedup on A100 GPUs.

CATS: Contextually-Aware Thresholding for Sparsity in Large Language Models

Large Language Models (LLMs) have dramatically advanced AI applications, yet their deployment remains challenging due to their immense inference costs. Recent studies ameliorate the computational costs of LLMs by increasing their activation sparsity but suffer from significant performance degradation on downstream tasks. In this work, we introduce a new framework for sparsifying the activations of base LLMs and reducing inference costs, dubbed Contextually Aware Thresholding for Sparsity (CATS). CATS is relatively simple, easy to implement, and highly effective. At the heart of our framework is a new non-linear activation function. We demonstrate that CATS can be applied to various base models, including Mistral-7B and Llama2-7B, and outperforms existing sparsification techniques in downstream task performance. More precisely, CATS-based models often achieve downstream task performance within 1-2% of their base models without any fine-tuning and even at activation sparsity levels of 50%. Furthermore, CATS-based models converge faster and display better task performance than competing techniques when fine-tuning is applied. Finally, we develop a custom GPU kernel for efficient implementation of CATS that translates the activation of sparsity of CATS to real wall-clock time speedups. Our custom kernel implementation of CATS results in a ~15% improvement in wall-clock inference latency of token generation on both Llama-7B and Mistral-7B.

  • 5 authors
·
Apr 12, 2024

Act First, Reason Later: Accelerating On-Policy Distillation for Multi-Turn Agents via Reference-Conditioned Inverse Dynamics

On-policy distillation (OPD) trains multi-turn language agents with dense teacher supervision on student-generated responses. However, standard think-then-act rollouts require lengthy reasoning before each short action, delaying environment transitions and experience collection. Generating actions directly reduces this delay but can degrade rollout quality. To address this, we propose ActFirst-OPD, an act-first, reason-later training framework that decouples environment interaction from full-response generation. The student infers and executes actions through reference-conditioned inverse dynamics using its current interaction context and a reference next observation, and switches to autonomous next-action prediction when the resulting transition deviates from the reference trajectory. From the collected interaction contexts, the student asynchronously generates full think-then-act responses for token-level teacher supervision. Experiments across 0.6B-, 1.7B-, and 4B-parameter Qwen3 students show that ActFirst-OPD achieves average wall-clock training speedups of 2.3times on ALFWorld, 1.8times on WebShop, and 4.9times on ScienceWorld over Vanilla OPD. It matches or exceeds all compared OPD baselines in mean task success rate across eight of nine benchmark-model settings. These results demonstrate that reasoning need not block acting during multi-turn agent distillation.

One Epoch Is All You Need

In unsupervised learning, collecting more data is not always a costly process unlike the training. For example, it is not hard to enlarge the 40GB WebText used for training GPT-2 by modifying its sampling methodology considering how many webpages there are in the Internet. On the other hand, given that training on this dataset already costs tens of thousands of dollars, training on a larger dataset naively is not cost-wise feasible. In this paper, we suggest to train on a larger dataset for only one epoch unlike the current practice, in which the unsupervised models are trained for from tens to hundreds of epochs. Furthermore, we suggest to adjust the model size and the number of iterations to be performed appropriately. We show that the performance of Transformer language model becomes dramatically improved in this way, especially if the original number of epochs is greater. For example, by replacing the training for 10 epochs with the one epoch training, this translates to 1.9-3.3x speedup in wall-clock time in our settings and more if the original number of epochs is greater. Under one epoch training, no overfitting occurs, and regularization method does nothing but slows down the training. Also, the curve of test loss over iterations follows power-law extensively. We compare the wall-clock time of the training of models with different parameter budget under one epoch training, and we show that size/iteration adjustment based on our proposed heuristics leads to 1-2.7x speedup in our cases. With the two methods combined, we achieve 3.3-5.1x speedup. Finally, we speculate various implications of one epoch training and size/iteration adjustment. In particular, based on our analysis we believe that we can reduce the cost to train the state-of-the-art models as BERT and GPT-2 dramatically, maybe even by the factor of 10.

  • 1 authors
·
Jun 16, 2019

A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts

Training and serving long-context large language models (LLMs) incurs substantial overhead. To address this, two critical steps are often required: a pretrained LLM typically undergoes a separate stage for context length extension by training on long-context data, followed by architectural modifications to reduce the overhead of KV cache during serving. This paper argues that integrating length extension with a GPU-friendly KV cache reduction architecture not only reduces training overhead during length extension, but also achieves better long-context performance. This leads to our proposed LongGen, which finetunes a pretrained LLM into an efficient architecture during length extension. LongGen builds on three key insights: (1) Sparse attention patterns, such as window attention (attending to recent tokens), attention sink (initial ones), and blockwise sparse attention (strided token blocks) are well-suited for building efficient long-context models, primarily due to their GPU-friendly memory access patterns, enabling efficiency gains not just theoretically but in practice as well. (2) It is essential for the model to have direct access to all tokens. A hybrid architecture with 1/3 full attention layers and 2/3 efficient ones achieves a balanced trade-off between efficiency and long-context performance. (3) Lightweight training on 5B long-context data is sufficient to extend the hybrid model's context length from 4K to 128K. We evaluate LongGen on both Llama-2 7B and Llama-2 70B, demonstrating its effectiveness across different scales. During training with 128K-long contexts, LongGen achieves 1.55x training speedup and reduces wall-clock time by 36%, compared to a full-attention baseline. During inference, LongGen reduces KV cache memory by 62%, achieving 1.67x prefilling speedup and 1.41x decoding speedup.

  • 5 authors
·
Oct 2, 2024

SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores

The ever-growing complexity of reinforcement learning (RL) tasks demands a distributed RL system to efficiently generate and process a massive amount of data to train intelligent agents. However, existing open-source libraries suffer from various limitations, which impede their practical use in challenging scenarios where large-scale training is necessary. While industrial systems from OpenAI and DeepMind have achieved successful large-scale RL training, their system architecture and implementation details remain undisclosed to the community. In this paper, we present a novel abstraction on the dataflows of RL training, which unifies practical RL training across diverse applications into a general framework and enables fine-grained optimizations. Following this abstraction, we develop a scalable, efficient, and extensible distributed RL system called ReaLly Scalable RL (SRL). The system architecture of SRL separates major RL computation components and allows massively parallelized training. Moreover, SRL offers user-friendly and extensible interfaces for customized algorithms. Our evaluation shows that SRL outperforms existing academic libraries in both a single machine and a medium-sized cluster. In a large-scale cluster, the novel architecture of SRL leads to up to 3.7x speedup compared to the design choices adopted by the existing libraries. We also conduct a direct benchmark comparison to OpenAI's industrial system, Rapid, in the challenging hide-and-seek environment. SRL reproduces the same solution as reported by OpenAI with up to 5x speedup in wall-clock time. Furthermore, we also examine the performance of SRL in a much harder variant of the hide-and-seek environment and achieve substantial learning speedup by scaling SRL to over 15k CPU cores and 32 A100 GPUs. Notably, SRL is the first in the academic community to perform RL experiments at such a large scale.

  • 5 authors
·
Jun 29, 2023

LongSpec: Long-Context Lossless Speculative Decoding with Efficient Drafting and Verification

As Large Language Models (LLMs) can now process extremely long contexts, efficient inference over these extended inputs has become increasingly important, especially for emerging applications like LLM agents that highly depend on this capability. Speculative decoding (SD) offers a promising lossless acceleration technique compared to lossy alternatives such as quantization and model cascades. However, most state-of-the-art SD methods are trained on short texts (typically fewer than 4k tokens), making them unsuitable for long-context scenarios. Specifically, adapting these methods to long contexts presents three key challenges: (1) the excessive memory demands posed by draft models due to large Key-Value (KV) cache; (2) performance degradation resulting from the mismatch between short-context training and long-context inference; and (3) inefficiencies in tree attention mechanisms when managing long token sequences. This work introduces LongSpec, a framework that addresses these challenges through three core innovations: a memory-efficient draft model with a constant-sized KV cache; novel position indices that mitigate the training-inference mismatch; and an attention aggregation strategy that combines fast prefix computation with standard tree attention to enable efficient decoding. Experimental results confirm the effectiveness of LongSpec, achieving up to a 3.26x speedup over strong Flash Attention baselines across five long-context understanding datasets, as well as a 2.25x reduction in wall-clock time on the AIME24 long reasoning task with the QwQ model, demonstrating significant latency improvements for long-context applications. The code is available at https://github.com/sail-sg/LongSpec.

  • 7 authors
·
Feb 24, 2025

Activation Steering for Chain-of-Thought Compression

Large language models (LLMs) excel at complex reasoning when they include intermediate steps, known as "chains of thought" (CoTs). However, these rationales are often overly verbose, even for simple problems, leading to wasted context, increased latency, and higher energy consumption. We observe that verbose, English-heavy CoTs and concise, math-centric CoTs occupy distinct regions in the model's residual-stream activation space. By extracting and injecting a "steering vector" to transition between these modes, we can reliably shift generation toward more concise reasoning, effectively compressing CoTs without retraining. We formalize this approach as Activation-Steered Compression (ASC), an inference-time technique that shortens reasoning traces by directly modifying hidden representations. In addition, we provide a theoretical analysis of the impact of ASC on the output distribution, derived from a closed-form KL-divergence-bounded constraint to regulate steering strength. Using only 100 paired verbose and concise examples, ASC achieves up to 67.43% reduction in CoT length on MATH500 and GSM8K datasets, while maintaining accuracy across 7B, 8B, and 32B parameter models. As a training-free method, ASC introduces negligible runtime overhead and, on MATH500, delivers an average 2.73x speedup in end-to-end reasoning wall-clock time on an 8B model. This makes ASC a practical and efficient tool for streamlining the deployment of reasoning-capable LLMs in latency- or cost-sensitive settings. The code is available at: https://github.com/ArminAzizi98/ASC

  • 3 authors
·
Jul 7, 2025 1

Three-Phase Transformer

We present Three-Phase Transformer (3PT), a residual-stream structural prior for decoder-only Transformers on a standard SwiGLU + RMSNorm + RoPE + GQA backbone. The hidden vector is partitioned into N equally-sized cyclic channels, each maintained by phase-respecting ops: a per-channel RMSNorm, a 2D Givens rotation between attention and FFN that rotates each channel by theta + i*(2*pi/N), and a head-count constraint aligning GQA heads with the partition. The architecture is a self-stabilizing equilibrium between scrambling and re-imposition, not a bolted-on module. The partition carves out a one-dimensional DC subspace orthogonal to the channels, into which we inject a fixed Gabriel's horn profile r(p) = 1/(p+1) as an absolute-position side-channel composing orthogonally with RoPE's relative-position rotation. The canonical N=3 borrows its metaphor from balanced three-phase AC, where three sinusoids 120 degrees apart sum to zero with no anti-correlated pair. At 123M parameters on WikiText-103, 3PT achieves -7.20% perplexity (-2.62% bits-per-byte) over a matched RoPE-Only baseline at +1,536 parameters (0.00124% of total), with 1.93x step-count convergence speedup (1.64x wall-clock). N behaves as a parameter-sharing knob rather than a unique optimum: at 5.5M an N-sweep over {1,2,3,4,6,8,12} is near-monotone with N=1 winning; at 123M a three-seed sweep finds N=3 and N=1 statistically indistinguishable. The load-bearing mechanism is the channel-partitioned residual stream, per-block rotation, per-phase normalization, and horn DC injection. We characterize (a) self-stabilization of the geometry without explicit enforcement, a novel instance of the conservation-law framework for neural networks; (b) a U-shaped depth profile of rotation-angle drift at 12 layers; (c) orthogonal composition with RoPE, attention, and FFN.

BrainsBuild BrainsBuild
·
Apr 14 5

Where the Model Changes Its Mind: Hindsight-Divergence Localization for Efficient Reinforcement Learning with Verifiable Rewards

Group-relative methods for reinforcement learning with verifiable rewards (RLVR) learn from differences in rollout outcomes. Independently sampling complete trajectories is costly and does not explicitly explore the decision space at critical positions. Feedback on a completed trajectory can reveal which earlier choices the policy reconsiders, suggesting where to sample alternative continuations. We introduce Hindsight-Divergence Localization (HDL), which uses hindsight-induced changes in token log-likelihoods to select branch points. HDL generates a small number of complete root trajectories and fills each training group with continuations from the selected positions under the original task context. Each continuation reuses its root prefix and contributes policy updates only through its newly generated suffix, reducing generation cost while focusing additional exploration and learning on decisions after branching. Experiments with three models across math, code, and agent tasks show gains in both rollout efficiency and task performance. Compared with GRPO at matched group sizes and training steps, HDL yields up to a 2.5times reduction in generated tokens and a 1.8times speedup in rollout wall-clock time. Despite this reduced generation budget, HDL improves performance across all three domains, with gains of up to 12.5 points on agent tasks.

nvidia NVIDIA
·
Sep 28 1

Efficient Graph Generation via Direct Prediction and Flow Matching

Generative modeling of graph-structured data is crucial for tasks ranging from drug discovery to social network simulation. Among these models, denoising diffusion models have achieved great success in graph generation by learning to progressively reverse a process that adds noise to the original graph. However, the standard noise-prediction approach of diffusion models is suboptimal for graph data. The goal for a graph generative model is to learn the clean graphs' topological properties, such as connectivity and degree distribution. Because a diffusion model that predicts noise does not explicitly learn these topological properties, it is challenging for the model to output graphs with the desired structural statistics. To address this challenge, we introduce Direct Graph Flow Matching (DiGFM), a novel graph transformer model guided by two goals: predict clean graphs and improve sampling efficiency. Distinct from the prevailing diffusion approach, DiGFM employs a continuous flow-matching paradigm and integrates direct graph prediction. Specifically, DiGFM maps the prior noise distribution to the clean graph distribution via a multi-step process: the model repeatedly predicts the underlying clean graph, and a transformation is employed to convert the model output to the velocity vector that points in the direction toward the clean graph distribution. This design enables DiGFM to generate high-quality samples using only 2.5% to 15.6% of the steps required by diffusion-based models, which leads to a 5.3x to 257x speedup in wall-clock inference time. Experiments demonstrate that DiGFM outperforms or matches prior state-of-the-art models across general graph benchmarks and molecular datasets, generating graphs with strong adherence to ground-truth structural statistics at significantly faster inference speeds.

  • 1 authors
·
Oct 4

Ladder-residual: parallelism-aware architecture for accelerating large model inference with communication overlapping

Large language model inference is both memory-intensive and time-consuming, often requiring distributed algorithms to efficiently scale. Various model parallelism strategies are used in multi-gpu training and inference to partition computation across multiple devices, reducing memory load and computation time. However, using model parallelism necessitates communication of information between GPUs, which has been a major bottleneck and limits the gains obtained by scaling up the number of devices. We introduce Ladder Residual, a simple architectural modification applicable to all residual-based models that enables straightforward overlapping that effectively hides the latency of communication. Our insight is that in addition to systems optimization, one can also redesign the model architecture to decouple communication from computation. While Ladder Residual can allow communication-computation decoupling in conventional parallelism patterns, we focus on Tensor Parallelism in this paper, which is particularly bottlenecked by its heavy communication. For a Transformer model with 70B parameters, applying Ladder Residual to all its layers can achieve 30% end-to-end wall clock speed up at inference time with TP sharding over 8 devices. We refer the resulting Transformer model as the Ladder Transformer. We train a 1B and 3B Ladder Transformer from scratch and observe comparable performance to a standard dense transformer baseline. We also show that it is possible to convert parts of the Llama-3.1 8B model to our Ladder Residual architecture with minimal accuracy degradation by only retraining for 3B tokens.

  • 10 authors
·
Jan 11, 2025

Efficient LLM Training and Serving with Heterogeneous Context Sharding among Attention Heads

Existing LLM training and inference frameworks struggle in boosting efficiency with sparsity while maintaining the integrity of context and model architecture. Inspired by the sharding concept in database and the fact that attention parallelizes over heads on accelerators, we propose Sparsely-Sharded (S2) Attention, an attention algorithm that allocates heterogeneous context partitions for different attention heads to divide and conquer. S2-Attention enforces each attention head to only attend to a partition of contexts following a strided sparsity pattern, while the full context is preserved as the union of all the shards. As attention heads are processed in separate thread blocks, the context reduction for each head can thus produce end-to-end speed-up and memory reduction. At inference, LLMs trained with S2-Attention can then take the KV cache reduction as free meals with guaranteed model quality preserve. In experiments, we show S2-Attentioncan provide as much as (1) 25.3X wall-clock attention speed-up over FlashAttention-2, resulting in 6X reduction in end-to-end training time and 10X inference latency, (2) on-par model training quality compared to default attention, (3)perfect needle retrieval accuracy over 32K context window. On top of the algorithm, we build DKernel, an LLM training and inference kernel library that allows users to customize sparsity patterns for their own models. We open-sourced DKerneland make it compatible with Megatron, Pytorch, and vLLM.

  • 7 authors
·
Jul 24, 2024 2

Sophia: A Scalable Stochastic Second-order Optimizer for Language Model Pre-training

Given the massive cost of language model pre-training, a non-trivial improvement of the optimization algorithm would lead to a material reduction on the time and cost of training. Adam and its variants have been state-of-the-art for years, and more sophisticated second-order (Hessian-based) optimizers often incur too much per-step overhead. In this paper, we propose Sophia, Second-order Clipped Stochastic Optimization, a simple scalable second-order optimizer that uses a light-weight estimate of the diagonal Hessian as the pre-conditioner. The update is the moving average of the gradients divided by the moving average of the estimated Hessian, followed by element-wise clipping. The clipping controls the worst-case update size and tames the negative impact of non-convexity and rapid change of Hessian along the trajectory. Sophia only estimates the diagonal Hessian every handful of iterations, which has negligible average per-step time and memory overhead. On language modeling with GPT-2 models of sizes ranging from 125M to 770M, Sophia achieves a 2x speed-up compared with Adam in the number of steps, total compute, and wall-clock time. Theoretically, we show that Sophia adapts to the curvature in different components of the parameters, which can be highly heterogeneous for language modeling tasks. Our run-time bound does not depend on the condition number of the loss.

  • 5 authors
·
May 23, 2023 1

NOBLE: Accelerating Transformers with Nonlinear Low-Rank Branches

We introduce NOBLE (Nonlinear lOw-rank Branch for Linear Enhancement), an architectural augmentation that adds nonlinear low-rank branches to transformer linear layers. Unlike LoRA and other parameter-efficient fine-tuning (PEFT) methods, NOBLE is designed for pretraining from scratch. The branch is a permanent part of the architecture as opposed to an adapter for finetuning on top of frozen weights. The branch computes σ(xWdown)Wup where σ is a learnable nonlinearity. We evaluate several activation functions and find that CosNet, a two-layer cosine nonlinearity with learnable frequency and phase with a linear projection in between them in the bottleneck space, performs best. NOBLE achieves substantial improvements with minimal overhead: up to 1.47x step speedup to reach baseline eval loss (up to 32% fewer training steps), with as low as 4% additional parameters and 7% step time overhead, resulting in up to 1.22x net wallclock speedup. Experiments on LLMs (250M and 1.5B parameters), BERT, VQGAN, and ViT consistently show improved training efficiency. We identify one caveat: Mixup/CutMix augmentation interferes with NOBLE's benefits in Imagenet classification along with other stochastic augmentations, but when disabled, ViT also improves. This discrepancy is possibly explained by regularization techniques that encourage smoother fits to the target function while NOBLE may specialize more in sharper aspects of the target function.

  • 1 authors
·
Mar 6

Tempo Adaptation in Non-stationary Reinforcement Learning

We first raise and tackle a ``time synchronization'' issue between the agent and the environment in non-stationary reinforcement learning (RL), a crucial factor hindering its real-world applications. In reality, environmental changes occur over wall-clock time (t) rather than episode progress (k), where wall-clock time signifies the actual elapsed time within the fixed duration t in [0, T]. In existing works, at episode k, the agent rolls a trajectory and trains a policy before transitioning to episode k+1. In the context of the time-desynchronized environment, however, the agent at time t_{k} allocates Δt for trajectory generation and training, subsequently moves to the next episode at t_{k+1}=t_{k}+Δt. Despite a fixed total number of episodes (K), the agent accumulates different trajectories influenced by the choice of interaction times (t_1,t_2,...,t_K), significantly impacting the suboptimality gap of the policy. We propose a Proactively Synchronizing Tempo (ProST) framework that computes a suboptimal sequence {t_1,t_2,...,t_K} (= { t_{1:K}}) by minimizing an upper bound on its performance measure, i.e., the dynamic regret. Our main contribution is that we show that a suboptimal {t_{1:K}} trades-off between the policy training time (agent tempo) and how fast the environment changes (environment tempo). Theoretically, this work develops a suboptimal {t_{1:K}} as a function of the degree of the environment's non-stationarity while also achieving a sublinear dynamic regret. Our experimental evaluation on various high-dimensional non-stationary environments shows that the ProST framework achieves a higher online return at suboptimal {t_{1:K}} than the existing methods.

  • 6 authors
·
Oct 26, 2023

JaxMARL: Multi-Agent RL Environments in JAX

Benchmarks play an important role in the development of machine learning algorithms. For example, research in reinforcement learning (RL) has been heavily influenced by available environments and benchmarks. However, RL environments are traditionally run on the CPU, limiting their scalability with typical academic compute. Recent advancements in JAX have enabled the wider use of hardware acceleration to overcome these computational hurdles, enabling massively parallel RL training pipelines and environments. This is particularly useful for multi-agent reinforcement learning (MARL) research. First of all, multiple agents must be considered at each environment step, adding computational burden, and secondly, the sample complexity is increased due to non-stationarity, decentralised partial observability, or other MARL challenges. In this paper, we present JaxMARL, the first open-source code base that combines ease-of-use with GPU enabled efficiency, and supports a large number of commonly used MARL environments as well as popular baseline algorithms. When considering wall clock time, our experiments show that per-run our JAX-based training pipeline is up to 12500x faster than existing approaches. This enables efficient and thorough evaluations, with the potential to alleviate the evaluation crisis of the field. We also introduce and benchmark SMAX, a vectorised, simplified version of the popular StarCraft Multi-Agent Challenge, which removes the need to run the StarCraft II game engine. This not only enables GPU acceleration, but also provides a more flexible MARL environment, unlocking the potential for self-play, meta-learning, and other future applications in MARL. We provide code at https://github.com/flairox/jaxmarl.

  • 20 authors
·
Nov 16, 2023

Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models

Inference-time scaling for text-to-image generation has progressed from simple Best-of-N (BoN) sampling to guided search methods that verify and steer candidate trajectories at intermediate denoising steps. These approaches focus on when and how often to verify during denoising but largely treat the cost of generation itself as fixed. Moreover, the standard practice of comparing methods by number of function evaluations (NFEs) counts only denoising forward passes and ignores verifier overhead, which can distort efficiency rankings. We show that under wall-clock evaluation, simple BoN already matches or outperforms several guided search techniques, suggesting that compute is better spent on broader exploration than on repeated intermediate verification. This motivates Flash-BoN, which generates a large pool of inexpensive draft candidates by combining three complementary acceleration knobs: timestep truncation, layer skipping, and activation proxies into a single configuration optimized once per model. An efficient multi-stage verification procedure then identifies the most promising draft, which is refined at full quality. Across three benchmarks and three model scales, Flash-BoN consistently outperforms all baselines under fixed wall-clock budgets, with gains that grow at larger model scales (+8% AUC). We further show that our strategy combines well and improves existing orthogonal techniques such as reflection-based prompt optimization (+16% AUC). The gains correlate with increased candidate diversity, which also enables draft-guided selection to accelerate RL post-training convergence.

The Stability-Efficiency Dilemma: Investigating Sequence Length Warmup for Training GPT Models

Recent works have demonstrated great success in pre-training large-scale autoregressive language models on massive GPUs. To reduce the wall-clock training time, a common practice is to increase the batch size and learning rate. However, such practice is often brittle and leads to a so-called stability-efficiency dilemma: increasing the batch sizes and learning rates leads to better training efficiency but can also result in training instability, leading to poor generalization accuracy or failed runs. To better understand this phenomenon, we conduct an in-depth analysis on large-scale pre-training experiments replicating the GPT-2 model. We find that there is a strong correlation between training instability and extreme values of gradient variance, and that samples with long sequence lengths contribute to these extreme gradient variance values, especially at the beginning of the training, indicating that long sequence length can be a main source of training instability. Based on the analysis, we present a Sequence Length Warmup method that aims to solve the training stability-efficiency dilemma. Experiments replicating GPT-2 models show that our approach enables stable training with 8x larger batch size and 4x larger learning rate, whereas the baseline approach struggles with training instability. To achieve the same or better zero-shot evaluation results, our method reduces the required number of training tokens and wall clock time by up to 2.2x and 3.7x, respectively. Experiments replicating GPT-3 model (125M) show that our approach enables stable training with 8x larger batch size and 40x larger learning rate, and retains 99% of the zero-shot accuracy on 11 tasks using 10x less data and 17x less time compared to the original GPT-3 training recipe, while the baseline diverges under the same settings and only retain 95% of accuracy under lower learning rate.

  • 3 authors
·
Oct 15, 2022

Closed-Form Spectral Regularization for Multi-Task Model Merging

Model merging combines several independently fine-tuned experts into a single multi-task model without any training data, reducing the storage, serving, and decentralized-development costs of large foundation models. State-of-the-art merging methods formulate merging as a layer-wise quadratic interference minimization problem. Although this problem admits an exact closed-form pseudoinverse solution, that solution underperforms hundreds of iterations of gradient descent in practice. The iterative loop dominates the cost of the pipeline, yet its effectiveness has remained unexplained. We revisit this regime and show that the iterative solver does not primarily act as an optimizer; rather, it serves as an implicit spectral regularizer for an ill-posed normal equation, where small-eigenvalue directions of the per-layer interference operator amplify proxy noise. Building on this finding, we formalize multi-task model merging as a noisy linear inverse problem and propose a spectral filtering estimator parameterized by a per-direction filter. We instantiate this estimator with SWUDI, a closed-form method that combines a soft exponential filter, which matches the gradient-flow trajectory of iterative descent, with a hard top-K truncation that suppresses noise-amplifying small-eigenvalue directions. Furthermore, we propose SWUDI-A, an adaptive variant that replaces the global rank hyperparameter with per-layer rank rules, further improving robustness across architectures. Both variants share a single symmetric eigendecomposition per linear layer and require no training data or optimizer state. Across four general benchmarks and a multimodal merging benchmark spanning VQA, Geometry, Chart, OCR, Grounding, and modality merging, our proposed spectral solvers match or outperform state-of-the-art merging methods. Crucially, they reduce wall-clock time by 28-72x and peak GPU memory by up to 50%.

  • 7 authors
·
Jun 4