Autoregressive Decoding

Streaming automatic speech recognition (ASR) must be judged jointly on what it transcribes and on how quickly it commits each word. Delayed streams modeling (DSM) has become the dominant paradigm for streaming large audio-language models, exposing a structural delay $τ$ that bounds the decoder's lookahead. We show that $τ$ is a poor proxy for user-perceived latency, and that the alignment-based supervision of DSM leaves latency on the table: the same forced-aligned transcript is used at every $τ$, forcing the model to withhold words it could already commit. We introduce AWED, a word-level emission-delay metric defined relative to the acoustic end of each word, and post-train a DSM recognizer with GRPO under a reward that scores transcription accuracy and measured delay jointly. Trained at a single operating point ($τ=6$ frames), our model dominates both its supervised fine-tuning initialization and the Voxtral Realtime backbone across all evaluated lookahead budgets: it cuts WER by 30.8% relative at an 80,ms structural delay, and by 5.7% relative at 480,ms while lowering median AWED from 1.17,s to 1.04,s. Latency-rewarded post-training thus advances the accuracy--latency Pareto frontier of streaming ASR without architectural change.

Open Thai automatic speech recognition (ASR) is dominated by offline, Whisper-based models that read the whole utterance before transcribing, ruling out low-latency uses such as live captioning and voice agents. We present a deployable system for streaming Thai ASR that lets a user steer its vocabulary at decode time, without retraining. A widely used open Thai model, trained with full context, collapses when run as a true stream; we restore streaming with a cache-aware encoder, by converting it or adapting a natively streaming one, and add a shallow-fusion layer that re-ranks candidates inside the streaming decoder with a GPU n-gram language model and phrase boosting. Across two Thai benchmarks and two model sizes, the streaming models stay usable where the full-context model fails, cutting character error rate 4.3-4.5x at a one-second look-ahead while running faster than real time. Decode-time steering then lifts keyword recall from 16.6% to 20.7% at no accuracy cost and negligible overhead; most of the gain comes from an n-gram over ordinary training transcripts, which resolves the written form of code-switched words the model hears but spells inconsistently, with phrase boosting adding targeted control over rare domain terms.

Vision-language models (VLMs) are increasingly evaluated on complex image and video understanding tasks, yet conventional metrics primarily assess final-answer quality and reveal little about how different information sources shape the generation process. We propose a causal and temporal evaluation framework that traces the evolving roles of visual input, question text, and generated prefixes during autoregressive decoding. Grounded in a Structural Causal Model, we use interventions and backdoor adjustment to derive three step-indexed causal-drive metrics---Visual Causal Drive (VCD), Question Causal Drive (QCD), and Prefix Causal Drive (PCD)---for characterizing source-specific generation patterns without requiring reference answers. Experiments on Qwen3-VL-8B-Instruct across MAVIS, LLaVA-Video-178K, and MiraData, together with cross-model validation on InternVL2-8B, reveal a consistent transition from stronger early question and visual guidance toward increasing reliance on generated prefixes. Randomized-intervention validation shows that QCD and PCD reduce recovery error over observational PMI baselines by 34.8% and 47.1%, respectively. On VLMBias, the prefix--visual imbalance score achieves 0.767 AUROC and 0.873 AUPRC for distinguishing prior-driven from visually grounded generations. These results show that causal-drive trajectories provide complementary source-level diagnostics for multimodal generation.

Vision-Language-Action (VLA) models are reshaping autonomous driving (AD) by unifying perception, reasoning, and control through language, enabling semantic grounding, interpretable decisions, and better long-tail generalization. But language is expensive onboard: latency and memory budgets are tight, and autoregressive decoding is inherently sequential. This work reframes the central question as when and where language should act at inference, since inference cost recurs at every deployed frame while training cost is paid once. We introduce the Language Residue taxonomy to organize methods by their inference-time use of language: train-time-only supervision (L1), latent non-textual reasoning (L2), conditional invocation (L3), and full per-frame generation (L4). We review representative methods and tag each across five deployment axes (latency, parameters, memory, FLOPs, tokens), analyzing them on major open- and closed-loop driving benchmarks (e.g., nuScenes, NAVSIM, Bench2Drive). We further trace how efficient methods from NLP/LLM are adapted in AD, identifying the constraints and motivations driving these adaptations. A continuously updated repository will be available at Github.

Edge inference is a promising paradigm to provide large language model (LLM) inference services in next-generation mobile networks. LLM inference mainly relies on two approaches: Autoregressive decoding (AD) generates output tokens sequentially, resulting in long latency; Speculative decoding (SD) accelerates inference by using a small language model (SLM) to generate multiple draft tokens for LLM verification, but incurs extra memory costs. Due to this latency-memory tradeoff, neither approach alone can efficiently serve users with heterogeneous demands under limited edge computing resources. To address this challenge, we propose a hybrid autoregressive-speculative inference (BALANCE) framework for edge LLM inference. In BALANCE, an edge server hosts both an SLM and an LLM, assigns each user to AD or SD, and performs the two modes simultaneously. To maximize the number of served users, we formulate a task throughput maximization problem to jointly determine user scheduling and computing resource allocation between AD and SD under user latency requirements and server memory constraints. Since the problem is NP-hard, we develop a polynomial-time algorithm that transforms the original problem into two sub-problems and obtains a sub-optimal solution with a constant approximation guarantee. Experiments demonstrate that BALANCE consistently outperforms conventional AD and SD and significantly improves task throughput.

Masked diffusion language models (dLLMs) can commit tokens in any order -- a freedom marketed as their core advantage over autoregressive decoding. We show that on reasoning tasks this freedom is instead the axis of failure. Logging every commitment during decoding of LLaDA-8B on GSM8K, we find that unconstrained (pure) decoding commits the final answer at 15-24% of the trajectory while half the reasoning region is still masked, and collapses to answer-only outputs on up to 90% of problems as the canvas grows. The cause is not the model's termination beliefs -- EOS "pressure" is nearly identical across decoders -- but reachability: whether the sampler may act on those beliefs at distant positions. A 2x2 prompt-decoder design shows that chain-of-thought helps only under ordered commitment (interaction +34.8 percentage points, 95% CI [26.8, 42.8]; without reasoning text the decoders are indistinguishable), an interaction we decompose into a collapse channel and an order channel and replicate on Dream-7B and MATH-500. A single-knob intervention -- frontier-gated commitment -- causally recovers the full gap (0.528 to 0.852) while preserving up to 4x parallel decoding, along a measured frontier whose optimal window flips from w=1 at full refinement to unconstrained at 8 tokens/step. Our results reframe existing window-style samplers, previously motivated by efficiency, as the minimal fix for a reasoning pathology they were never designed to address.

Vision-Language-Action (VLA) models have emerged as a prominent paradigm for end-to-end autonomous driving; however, their efficient deployment is severely constrained by high computational latency and exposure bias arising from sequential autoregressive decoding. Conversely, while specialized diffusion policies enable low-latency, parallel execution, training them from scratch typically yields narrow, single-task architectures that lack holistic visual-linguistic reasoning. Successfully transforming pre-trained autoregressive generalists into parallel diffusion models could combine multi-task cognitive intelligence with execution efficiency, yet this transition presents a formidable architectural challenge due to mismatched attention patterns (causal versus bidirectional) and divergent optimization objectives. To bridge this divide, we introduce WAM-Diff2, a multi-task discrete diffusion VLA framework powered by a three-stage hierarchical distillation strategy. By structuring the architectural shift through progressive block-wise adaptation, block-wise distillation, and model-wise cross-scale distillation, WAM-Diff2 preserves the underlying semantic foundations of the base model while accelerating inference. Extensive evaluations across driving understanding, perception, and planning benchmarks demonstrate that WAM-Diff2 effectively mitigates exposure bias and achieves performance parity with autoregressive baselines. Crucially, the autoregressive-to-diffusion transition yields a 2.8x decoding speedup, which scales to an ultimate 15.1x acceleration when combined with system-level optimizations including FlashInfer and CUDA Graphs.

Audio-encoder-LLM-decoder architectures have become the dominant paradigm for modern automatic speech recognition (ASR), improving transcription quality through large-scale language modeling. However, the cost of autoregressive decoding scales with decoder size, creating a fundamental trade-off between recognition quality and serving latency. We argue this trade-off is not inherent: unlike open-ended text generation, ASR outputs are strongly anchored to the input speech signal, providing a natural inductive bias toward high-parallelism decoding. Building on this, we introduce ParaASR, an ASR system that leverages Multi-Token Prediction (MTP) to let a 4B LLM decoder emit multiple tokens per forward step. Starting from a publicly available audio-language foundation, the model first establishes a robust autoregressive recognizer and then aligns five future-token branches through a staged optimization recipe. At inference, it proposes a six-token continuation per step and admits only the verified prefix into the transcript, preserving the safety of standard autoregressive decoding. The average accepted length reaches 5.0 out of 6 proposed tokens, confirming that the deterministic structure of speech makes ASR an especially natural setting for multi-token decoding. ParaASR further retains a native 32K-context window and transcribes up to 30 minutes of audio in a single pass. Across diverse benchmarks, it attains average error rates of 2.97%, 3.68%, and 3.70% on Chinese, English, and long-form evaluations, respectively, while reaching a real-time factor (RTF) as low as 0.0053. These results show that decoder scaling, low-latency inference, and long-context transcription need not be competing goals when future-token proposals are anchored by the acoustic signal and guarded by autoregressive verification.

Diffusion language models enable flexible arbitrary-order generation, but existing sampling methods are mostly designed for early masked diffusion models (MDMs). In this work, we study sampling for recent block diffusion language models (BDLMs). We show empirically and analytically that these models are naturally more aligned with left-to-right decoding than MDMs. Based on this observation, we propose Parallel Autoregressive Decoding (PARD), a simple training-free sampling method that preserves left-to-right unmasking structure while allowing parallel token commitment. Extensive experiments show that PARD consistently outperforms existing parallel samplers in generation quality, while achieving substantial speedups over pure AR decoding with only a small quality gap.

The increasing deployment of large language models (LLMs) has magnified the computational and memory bottlenecks of autoregressive decoding, where low compute intensity and bandwidth-bound kernels dominate inference cost. Weight pruning offers a promising remedy, but existing methods remain confined to either static pruning (SP), which permanently removes redundant weights but lacks adaptivity, or dynamic pruning (DP), which adapts to input sparsity but introduces runtime irregularity. This paper presents SPDP, a unified sparse-inference framework that integrates unstructured SP with input-adaptive DP for efficient LLM inference on GPUs. SPDP co-designs a new Tiled-Column-wise Bitmap Compressed (Tiled-CBC) format and two complementary GPU kernels: (1) a CUDA-core spMspV kernel featuring Hybrid Activation-aware Dynamic Shared-Memory Bitmap Decoding (HAD-SMBD) for fine-grained, runtime activation skipping, and (2) a Tensor-Core SpMM kernel optimized for prefill computation. This joint format-kernel design harmonizes static and dynamic sparsity, maintaining bandwidth-efficient memory access and high compute intensity under both phases of LLM inference. Comprehensive evaluations on inference-optimized GPUs demonstrate that SPDP achieves 1.24x-1.37x average speedup (up to 2.51x) over state-of-the-art sparse frameworks such as SpInfer, while matching perplexity with up to 25% higher sparsity. SPDP advances the inference efficiency-quality Pareto frontier, showing that unified static-dynamic pruning can deliver substantial throughput and performance-per-watt improvements in large-scale LLM serving.

Single-stream autoregressive decoding of large language models is bound by memory bandwidth: each generated token requires one full forward pass through the target model, and successive passes cannot be parallelized. Speculative decoding restructures this computation: a small draft model proposes $K$ tokens autoregressively, the target model scores all of them in one batched pass, and a rejection-sampling rule provably preserves the target model's output distribution. We present a from-scratch, device-agnostic (CUDA/MPS/CPU) implementation and an empirical study across five draft/target backend configurations on a consumer Apple-silicon laptop. Distribution equivalence is verified at three levels, culminating in a two-sample test over roughly 9,200 real-model tokens per method ($χ^2 = 162.5$, dof $= 200$, $p = 0.976$) and exact greedy-sequence agreement. The best configuration reaches a measured $1.61\times$ wall-clock speedup at $K=6$, on an acceptance profile declining from 69.7% at $K=1$ to 37.8% at the optimum, while three of five configurations decelerate, either because the draft fails to out-speed a small target or because the quantized Metal backend executes "parallel" verification serially, an effect we isolate and quantify. The failures are as instructive as the successes: speculative decoding pays off only when verification is genuinely batch-parallel and the draft/target latency gap is real.

Masked diffusion language models (DLMs) enable parallel text generation by iteratively refining masked tokens, offering a promising alternative to autoregressive decoding. Recent lookahead-based decoding methods improve the accuracy--efficiency trade-off by exploring future decoding states before committing token updates. However, existing approaches mainly rely on shallow one-step lookahead, which optimizes immediate information gain but can be suboptimal for longer-horizon decoding trajectories. Meanwhile, we find that a naive extension for deeper lookahead is also ineffective, as fixed-depth rollout introduces additional computation and cannot adapt to heterogeneous intermediate decoding states. Thus, in this work, we propose AdaLook, an adaptive lookahead framework for DLM decoding. AdaLook dynamically determines whether to continue rollout based on candidate-score variance and further enables branch expansion when intermediate rollout states require additional exploration. This design avoids unnecessary deep rollout while allowing the decoder to re-trigger lookahead from informative intermediate states. Experiments on various benchmarks and models demonstrate that AdaLook achieves a better accuracy--decoding steps trade-off than existing one-step lookahead decoding methods.

Transformer reasoning is limited by autoregressive decoding, which repeat edly compresses rich hidden computation through token space and makes it difficult for intermediate reasoning states to persist across time. We in troduce Transformers with Temporal Middle-Layer Recurrence (T2MLR), a transformers-based latent reasoning architecture that fuses a cached middle layer representation from the previous token directly into an earlier layer of the current token position, enabling abstract intermediate computation to persist across decoding steps with little inference overhead. Across natural-language pretraining and multi-hop reasoning finetuning, T2MLR consistently outperforms data- and parameter-matched Transformer base lines. Moreover, applying recurrence to only a localized middle-layer block (as little as 20% of the network) often outperforms full-layer recurrence. Im portantly, T2MLR does not require pretraining from scratch: retrofitting the recurrent pathway into an existing pretrained 1.7B Transformer and briefly finetuning substantially improves math reasoning, lowering the barrier to practical adoption. These results suggest that effective latent reasoning in Transformers does not require looping over all layers as in previous works, but can instead emerge more strongly from targeted middle-layer recurrence.

Despite the rapid advancements of large language models (LLMs), LLM serving systems remain memory-intensive and costly. The key-value (KV) cache, which stores KV tensors during autoregressive decoding, is crucial for enabling low-latency, high-throughput LLM inference serving. In this survey, we focus on system-aware KV infrastructure for serving LLMs (abbreviated as sKis). We revisit recent work from a system behavior perspective, organizing existing efforts into three dimensions: execution and scheduling (temporal), placement and migration (spatial), and representation and retention (structural). Furthermore, we analyze cross-behavior co-design affinity and behavior-objective links, highlighting future opportunities. Our work systematizes a rapidly evolving area, providing a foundation for understanding and innovating KV cache designs in modern LLM serving infrastructure.

Speech large language models (Speech LLMs) typically encode speech into sequences far longer than text, creating a major efficiency bottleneck during autoregressive decoding. A common remedy is to compress the speech sequence at the adapter level to remove temporal redundancy before it enters the LLM; however, such early downsampling risks discarding fine-grained information that cannot be recovered. We propose SpeechKV, which applies a learned pooling to the KV cache of speech tokens inside the LLM. This design allows the LLM to fuse speech and text internally while directly accelerating decoding. Trained on 71K hours of speech data, SpeechKV compresses the speech to approximately text-level granularity yet maintains performance on par with or even slightly better than the uncompressed baseline, with relative gains of 6.6% on out-of-domain entity recognition and 2.3% on OpenASR, while delivering at least 1.49 times decoding speedup that scales with audio length.

Speculative decoding greatly increases the interactivity of autoregressive language models by trading off computation for extra tokens generated in a single forward pass. Factorized draft models are especially efficient because they predict future-token marginals in parallel, but their independence assumption causes acceptance rates to degrade sharply as the speculative budget grows. We analyze this limitation and introduce Weaver, a lightweight autoregressive adapter that constructs proposal trees from the top-K marginals of a factorized drafter. Weaver restores conditional dependencies between proposed tokens while avoiding a full-vocabulary projection. To support fast verification for models with Gated Delta Net layers, we derive a rollback-free tree-verification algorithm and implement optimized CUDA kernels in SGLang. By combining these model and systems contributions we achieve a 4.37-fold speedup over autoregressive decoding, and outperform a highly optimized DFlash baseline by 24.7%.

Key-value (KV) cache growth is a major bottleneck in autoregressive decoding, as memory and bandwidth scale linearly with context length. Existing KV eviction methods often rely on static heuristics or proxy scores, which poorly track future token utility and cause brittle eviction as relevance shifts. To address this, we introduce KVpop, which learns a fixed-budget KV eviction policy by directly supervising the keep-or-drop decision. The scorer is trained against a novel future-attention target, computed efficiently without materializing dense attention maps. We further introduce a delayed memory-based scorer that, uniquely among learned eviction methods, defers scoring for a fixed number of steps to exploit near-future context. On AIME and HMMT mathematical reasoning, KVpop retains 98% of full-attention performance on Qwen3-4B at 75% KV cache compression and 97% at 88% compression, consistently outperforming established eviction baselines. Qwen3-8B shows even stronger results, reaching near-full teacher performance. These results show that supervising eviction with future-attention signals cuts memory costs while maintaining quality.

Large language models (LLMs) generate fluent outputs that can be wrong. Unlike humans, who often exhibit cues when providing false information, LLMs produce errors that are difficult to detect because autoregressive decoding provides no mechanism for verifying intermediate reasoning before state progression. We introduce Heaviside Continuity of Rolling Coefficients (HCRC), a verification-first execution framework that reformulates inference as predicate-gated state transitions governed by a Heaviside Gate. HCRC combines model confidence with independent verification signals from a parallel worker architecture, allowing execution to advance only when predefined correctness predicates are satisfied. This prevents invalid intermediate states from propagating, reducing epistemic entropy without modifying the underlying model. We evaluate HCRC on software-engineering and reasoning tasks across thirteen proposers from four providers. On capable proposers, the gate reduces the false-completion rate (FCR) from 4--7% to 0% while remaining latency-competitive and, in some settings, faster than the unwrapped model. On weaker proposers, it converts false completions into honest halts instead of corrupting downstream state. Beyond benchmarking, HCRC has operated for months as the production control plane of an agentic coding environment, authorizing file mutations, verification-driven progress reporting, and memory compaction. These results establish HCRC as a general framework for verification-driven LLM execution, showing that reliable reasoning can be achieved through principled execution control rather than model scale alone.

Dense video captioning aims to generate temporally grounded descriptions of video events, benefiting both event-level video understanding and generation. In this domain, autoregressive video large language models have emerged as a prevalent paradigm due to their strong generative and cross-modal modeling capacity. However, generating dense captions under the token-by-token paradigm severely limits inference efficiency and hinders scalability as video length and event density increase. In this work, we propose a parallelized autoregressive framework that not only improves generation efficiency but also enhances temporally grounded captioning performance. Our key insight is to exploit the weak local dependencies across temporally distinct events to restructure the causal dependency graph, thereby enabling lossless parallel generation. Specifically, tokens with weak cross-event dependencies can be decoded in parallel, while tightly coupled tokens within each event retain sequential decoding to preserve local semantic coherence. To realize this insight, we introduce two key components for lossless parallel decoding: (1) a latent global planning mechanism that automatically learns the event-level structure and produces compact tokens encoding global inter-event causality while adaptively aggregating event-level audio-visual semantics, guiding subsequent dependency restructuring and parallel decoding; and (2) an event-factorized parallel decoding mechanism that effectively balances local focus with global inter-event awareness. Experiments on various benchmarks demonstrate the clear advantage of our approach in both efficiency and performance in omni-modal event grounding and captioning. Project website: https://github.com/showlab/PadCaptioner.

Low-bit quantization shrinks language models but treats precision as a single global hyper-parameter: every weight uses the same bit-width. We introduce Variable Bit-width Quantization (VBQ), a training-time method in which each contiguous group of 64 weights learns its own resolution from {1,2,4,8} bits via a Gumbel-Softmax relaxation, trained jointly by an alternating optimization that gives the precision logits a clean, task-aligned signal. VBQ discovers a consistent, strongly heterogeneous allocation within individual projection types, not merely across layers, impossible to express with per-layer methods: 69% of groups collapse to 1 bit, the LM head averages 1.09 bits, while the first MLP block keeps ~2.5 bits. This pattern is stable enough to freeze into a fixed recipe and reuse without further search. The recipe yields a "bigger-but-smaller" regime: a 131M model at 1.82 mean bits reaches perplexity 4.2 on TinyStories, beating a 55M FP16 model (PPL 4.4) at 3.8x less storage, and lets a 1.46B model on FineWeb-Edu match a 593M FP16 control at ~3.7x less storage with 2.5x more parameters. As quality-per-byte, VBQ is 3.9-8.4x more efficient than FP16. The recipe maps directly to packed low-bit storage, so it also accelerates inference: with custom fused dequantize-and-multiply kernels, memory-bandwidth-bound autoregressive decode is faster at equal output, and the speedup grows with scale (parity at 131M, 1.9x at 1.0B, 4.7x at 9B on Apple silicon). A distributional analysis (KL divergence and argmax-flip rate) reveals a striking mechanism: deeper layers progressively self-heal the quantization error injected by early layers. The win is a from-scratch, train-time phenomenon; scaling the search economically beyond 1.5B parameters remains open. VBQ reframes precision as a learnable, non-uniform resource and shows that spending a fixed bit budget unevenly beats spending it uniformly.

Interleaved multimodal reasoning improves visual grounding by revisiting visual evidence during multi-step generation, yet existing methods typically rely on token replay, repeatedly forwarding selected visual tokens. A natural shortcut is to reuse the historical visual key-value (KV) cache directly. However, we identify a critical failure mode of this strategy: cached visual keys are already bound to their original positional context. Such stale positional binding distorts attention under later decoding contexts and can trigger severe autoregressive decoding collapse. This failure suggests that effective cache reuse requires reconstructing visual evidence under positions compatible with the current decoding state, rather than directly copying position-bound historical cache entries. To this end, we propose Position Rebinding Cache Reuse (PRCR), a cache-level framework for replay-free visual revisiting. PRCR stores raw visual KV cache together with their original spatial coordinates, then reassigns position-compatible coordinates to select entries and rebinds their keys before injecting the reconstructed cache into the active decoder cache. This design reuses historical visual evidence while preserving textual positional continuity and relative visual structure. Experiments across multiple multimodal reasoning benchmarks show that PRCR achieves replay-level or better performance, improving average accuracy by 5 percent and reducing visual-revisiting computation by up to tens of thousands of times.

Masked diffusion language models decode by iteratively unmasking tokens, where the unmasking order defines an "order of thought" that strongly influences generation quality yet is typically chosen heuristically. We derive a tractable upper bound on the sequential decoding mismatch, measured by the Kullback-Leibler divergence and expressed in terms of the model's pathwise log-likelihood, with tightness under sufficient model expressivity. This bound induces a dense self-aware reward over ordered trajectories, casting order selection as a principled policy optimization problem with a frozen denoiser. We instantiate this idea as Self-Aware Scheduling (SAS), which learns a lightweight order policy using Group Relative Policy Optimization and applies seamlessly to both any-order and semi-autoregressive decoding. On Sudoku with 1B MDM, SAS improves puzzle accuracy from 82.0% (best heuristic schedule) to 91.8%, and reaches 97.5% with second-stage fine-tuning along learned trajectories. On mathematical reasoning with LLaDA-8B, SAS improves pass@1 on GSM8K from 64% to 76% and on MBPP from 39.5% to 41%, consistently matching or exceeding heuristic schedules across generation lengths and block sizes. Project page: https://jimmyxu123.github.io/SAS

In speech processing, most state-of-the-art sequence prediction models rely on auto-regressive (AR) strategies to generate output sequences based on the raw predictions of the model. Despite their crucial role in the inference process, a comprehensive overview of AR strategies as a unified field is lacking, due largely to implicit and multiple definitions of next-token decoding. This context complicates the choice, comparison, and evaluation of strategies, while creating inconsistencies in the characterization of approaches as auto-regressive or not. We begin by setting explicit inclusion criteria for the field of AR search in speech processing, and derive a generalized theoretical framework to categorize and report on search strategies for neural models. We show the capabilities of this formalism in simplifying the design of benchmarks centered around the decoding process, allowing for ablation studies that are focused on search strategies.

Non-autoregressive (NAR) decoding generates output tokens in parallel, making speech recognition faster than autoregressive decoding, which generates them sequentially from left to right. However, the recognition performance is degraded because NAR decoding cannot resolve uncertainty by conditioning on previously generated tokens. To address this issue, we propose a novel NAR decoding framework based on minimum Bayes' risk (MBR) decoding, termed NAR-MBR decoding, that maximizes the expected utility calculated from samples drawn from the output probability of an NAR model rather than maximizing the output probability. Notably, by leveraging the nature of NAR models, multiple samples are obtained efficiently with a single forward computation. Our experiments across LibriSpeech, Switchboard, AMI, and web presentation corpus demonstrated that our NAR-MBR decoding outperformed previous NAR decoding and ran faster than AR decoding.

Low frame rates in neural audio codecs are attractive for autoregressive speech synthesis, where the generation cost scales linearly with the sequence length. Recent work has demonstrated that codecs can operate at 12.5 Hz and below, but the mechanisms underlying low frame rate degradation remain insufficiently understood. We investigate these mechanisms through a controlled frame rate ablation. We reproduce a quality cliff at 6.25 Hz reported in previous works and evaluate candidate explanations: phonemic collisions and codebook saturation, neither of which shows evidence of a fundamental barrier. The cliff is instead caused by suboptimal training configuration: fixed clip duration during training yields too few tokens at low frame rates, starving the decoder of inter-token context. Once corrected, WER degrades smoothly with phonemic load down to 3.1 Hz and 1.6 Hz, suggesting the inference-time efficiency gains of low frame rate codecs are more accessible than previously assumed.

Large reasoning models (LRMs) improve complex problem-solving by generating long intermediate reasoning traces, but this substantially increases inference costs. NVFP4 inference offers a promising approach to reduce both computational and memory costs through hardware-supported low-precision execution. However, directly applying NVFP4 to LRMs introduces two practical limitations: reasoning accuracy degrades under quantization, and existing NVFP4 kernels do not fully realize latency benefits in small-batch autoregressive decoding. In this work, we analyze the effect of NVFP4 quantization on token-level uncertainty during reasoning. We show that quantization increases incorrect sampling at low-entropy symbolic tokens, while causing over-concentration on a small set of tokens in high-uncertainty reasoning steps. Based on this observation, we propose \textbf{ReSET}, a reasoning-step entropy-based temperature-scaling method that estimates step-level uncertainty online and adapts the decoding temperature using both token-level and step-level entropy signals. To address the latency gap, we further design a CUDA-core small-$M$ NVFP4 kernel for latency-critical autoregressive decoding. Across reasoning benchmarks and model scales, ReSET improves NVFP4 reasoning accuracy by up to $\sim\!$2 points over the NVFP4 baseline. Our CUDA-core small-$M$ kernel further improves latency-critical decoding, delivering up to $2.5\!\times$ kernel-level speedup over NVFP4 vLLM and approximately $2\!\times$ end-to-end decoding speedup over BF16. Code is available at https://github.com/aiha-lab/ReSET.

De novo peptide sequencing from tandem mass spectrometry is pivotal in proteomics, enabling identification of novel peptides without reference databases. While recent Transformer-based encoder-decoder models have achieved remarkable performance, we uncover a critical pathology in their inference dynamics. Through comprehensive feature scaling experiments, we demonstrate that existing auto-regressive peptide decoders tend to over-rely on generated-sequence priors while progressively under-utilizing fine-grained physical evidence from the input mass spectrum. This phenomenon leads to suboptimal results, where generated peptide sequences are biologically plausible yet not faithful to the input spectrum. To rectify this, we propose MemNovo, a training-free and plug-and-play mechanism that re-balances peptide and spectral contributions at inference time. MemNovo alleviates the information bottleneck by establishing a persistent spectral memory bank and injecting retrieved features directly into the final decoding stage via an ultra-conservative residual connection. Theoretical analysis confirms that this mechanism restores the mutual information between the decoder state and the raw spectrum. Extensive experiments on the Nine Species benchmark with two representative baselines, Casanovo and InstaNovo, demonstrate that MemNovo consistently improves both amino acid precision and peptide precision, achieving up to 39.1% relative improvement in peptide precision for Casanovo and up to 3.9% for InstaNovo, with negligible computational overhead.

Large language model inference is bottlenecked by autoregressive decoding, where each token requires a full forward pass. Multi-token prediction (MTP) offers a promising acceleration path, but existing approaches suffer from a fundamental architectural flaw: the MTP head for the first token competes with the backbone's own language model (LM) head, leading to severe quality degradation when predictions are accepted. We identify this head-backbone competition as the root cause of repetitive and incoherent outputs in prior MTP-based acceleration methods. To address this, we propose Backbone-as-Architect, a design principle where the backbone LM head always generates the first token, and MTP heads are responsible only for subsequent tokens. Building on this principle, we introduce CLP (Collocation-Length Predictor), a lightweight span-level decision layer that predicts how many additional tokens can be safely accepted at each decoding step. CLP uses only a single linear layer (4.6K--7.7K parameters), replacing the over-engineered 1M-parameter gate networks used in prior work. Experiments on Qwen2.5 models (0.5B, 1.5B, 7B) show that CLP achieves 1.20x--1.29x speedup on 1.5B and 1.14x--1.20x on 7B, with zero quality degradation (repetition ratio < 0.02), while gate-based approaches fail to accelerate (1.07x) or produce severely degraded outputs (repetition ratio > 0.5%). We further demonstrate that shorter prediction horizons (k=2) recover 24% higher MTP head accuracy on large models, establishing a scaling-aware design principle. We identify MTP head prediction accuracy as the binding constraint on acceleration and establish a clear roadmap for future improvements.

While speech quality is typically assessed on complete utterances, streaming and generative systems require incremental estimation from partial audio. Existing predictors assume full context, degrading on prefix-constrained inputs. Extending ARECHO, we propose ANCHOR, reformulating incremental assessment as a multi-resolution autoregressive task. It models chunk- and utterance-level quality within a single decoder using dual-resolution tokens and a resolution-aware hierarchy for coarse-to-fine refinement. Experiments show substantial robustness under partial input, including a 48% PLCMOS error reduction on 2-second prefixes. Convergence analysis reveals a 4-6 s effective perceptual context horizon. A stress test further isolates structured extrapolation biases under localized corruption. Results demonstrate that hierarchical supervision improves incremental prediction and elucidates how perceptual quality accumulates over time.

Block-wise semi-autoregressive decoding is the standard inference paradigm for diffusion large language models (DLMs), but it imposes a strict dependency between blocks: the next block cannot begin until the current block is fully decoded or its denoising budget is exhausted. We observe that once a block exposes a reliable delimiter boundary or stable semantic prefix, continuation generation need not wait for every residual token to be resolved. We propose AsyncLane, a training-free decoding scheduler that decouples refinement from advancement. AsyncLane forks a generate lane at observed delimiter boundaries into a refine lane and a continuation generate lane: the prefix remains editable, while the continuation advances before prefix refinement finishes. The resulting lane tree records decoding dependencies and output order, while execution proceeds over the active lane set. To make this asynchronous schedule efficient under bidirectional attention, AsyncLane combines shared-prefix lane batching, lookahead draft reuse, cascading termination, and compact cache refresh with refresh-logit reuse, preventing model-call cost from scaling directly with the number of lanes. AsyncLane is a drop-in replacement for block-wise DLM samplers and requires no retraining. Experiments on mathematical reasoning and code generation show that AsyncLane consistently improves throughput while maintaining competitive quality. Across LLaDA and Dream backbones, AsyncLane achieves the highest TPS in all evaluated benchmark-length settings; relative to the fastest competing baseline, it reaches peak speedups of 2.95x on LLaDA and 3.04x on Dream, with especially large gains under longer generation budgets.

Neural symbolic regression models improve inference efficiency by shifting structural search to pretraining, but their one-pass autoregressive decoding is prone to error accumulation, which may lead to generating structurally incorrect expressions, especially in complex expression generation scenarios. Existing rectification strategies can alleviate this issue, but they often depend on restarting global search, thereby weakening the efficiency advantage of neural models, and remain susceptible to error accumulation. In this paper, we propose EditSR, a two-layer framework that combines a neural symbolic regression model in the first layer with an edit-based Rectifier in the second layer to achieve efficient prediction and post-hoc rectification. Instead of restarting the global search, we maintain rectification efficiency by pretraining the Rectifier. Specifically, we formulate the rectification process as a step-by-step state-transition chain starting from an incorrect expression, and develop a state-transition algorithm to construct supervised rectification chains for training the Rectifier. To ensure syntactic validity throughout rectification, each edit action is restricted to a syntactically valid space so that every edited expression remains parseable. In addition, because each edit decision is conditioned on the current state rather than the history, the Rectifier allows errors made in earlier steps to be rectified by subsequent edits, thereby reducing the risk of error accumulation. Extensive experiments and ablation studies show that EditSR substantially improves symbolic structure recovery with limited extra cost, with more pronounced gains on complex expressions, where one-pass autoregressive decoding is more susceptible to error accumulation.

Many modern vision-language models (VLMs) build on autoregressive decoding of discrete tokens. While text-based output interfaces enable scalable pretraining and strong zero-shot generalization across diverse tasks, they are poorly suited for problems that require precise continuous outputs, such as localizing temporal boundaries of events or generating robotic control actions. To address this challenge, we propose DRIFT, a general framework for adapting pretrained VLMs to continuous decoding tasks. DRIFT combines a base predictor, which provides a coarse estimate of the target output, with a generative refinement module based on flow matching that iteratively improves the prediction. This residual formulation transforms the generative modeling problem from learning a global output distribution to modeling a localized residual distribution around a strong prior, substantially simplifying optimization. We evaluate DRIFT on both perception and planning tasks, including visual grounding and robotic control. Across multiple tasks and architectures spanning MLLMs, VLAs, and WAMs, DRIFT consistently outperforms a strong set of regression- and generative-based solutions.

Test-time scaling is a powerful approach to obtain better reasoning in large language models, but it becomes memory-bottlenecked during long-horizon decoding, as the KV-cache grows. KV-cache quantization can help improve this, but current methods are evaluated under prefill-like settings and errors behave differently under autoregressive decoding. We show that in the latter regime, quantization errors accumulate across timesteps, driven primarily by incorrect token scales. We introduce KVarN, a calibration-free KV-cache quantizer that applies a Hadamard rotation followed by a dual-scaling variance normalization across both axes of the K and V matrices. We find that this combination fixes outlying token-scale errors and substantially reduces error accumulation over existing baselines. KVarN establishes a new state-of-theart for KV-cache quantization on generative benchmarks, including MATH500, AIME24 and HumanEval, at 2-bit precision. A vLLM implementation of the KVarN method is available at https://github.com/huawei-csl/KVarN

While LLM-based Automatic Speech Recognition (ASR) achieves high accuracy, its speed is limited by sequential autoregressive decoding. Diffusion Language Models (DLMs) offer a parallel alternative, yet their decoding strategies remain under-explored in ASR contexts. This paper analyzes three decoding schemes for DLM-based ASR: fixed-number, static confidence threshold, and dynamic confidence threshold. We introduce a round-wise analysis of decoding progress using Negative Log-Likelihood-based uncertainty as a proxy for prediction reliability. Our results show that both threshold-based strategies provide a better accuracy-speed trade-off than fixed-number schemes. This behavior is associated with the more concentrated confidence distribution observed in the evaluated ASR settings: many tokens reach high confidence early, enabling multiple tokens to be committed in early decoding rounds while lower-confidence tokens are deferred to later rounds. The static-threshold strategy achieves accuracy close to autoregressive decoding at lower decoding cost.

Vision-Language Models (VLMs) have emerged as a critical and fast-growing extension of Large Language Models (LLMs) that enable multimodal reasoning through both text and image inputs. Although VLMs enrich the capabilities of language models, they also inherit and amplify key computational bottlenecks: the memory overhead caused by the large key-value (KV) cache during autoregressive decoding. This challenge is particularly severe in VLMs, where images produce longer token sequences and denser feature representations compared to text. Moreover, the spatial and information-rich nature of vision tokens introduces structured attention patterns that make many LLM-oriented KV cache compression techniques ineffective when applied directly to VLMs. In this work, we conduct a detailed empirical analysis of the behavior of vision tokens, highlighting the critical differences from purely text-based models. Based on these insights, we propose KVCapsule, a novel KV cache compression framework for vision tokens. KVCapsule keeps the pretrained VLM backbone frozen, requires no modification to the attention computation modules, and can be integrated into existing VLMs through lightweight compression and reconstruction components. We evaluate KVCapsule on multiple VLMs and benchmark tasks, demonstrating up to 2x improvement in TPS and 2.4x reduction in KV cache memory at a 60% compression ratio, with negligible degradation in accuracy or response quality. Our findings offer practical pathways to scale VLM inference under constrained memory budgets and inspire further research into structure-aware cache compression for multimodal models.

Speculative decoding accelerates LLM inference by drafting future tokens with a small model, but drafter models degrade sharply under template perturbation and long-context inputs. We identify a previously-unreported phenomenon we call \textbf{attention drift}: as the drafter generates successive tokens within a speculation chain, attention progressively moves from the prompt onto its own recently-generated tokens. We observe this across both \emph{EAGLE3} drafters and \emph{MTP heads}, suggesting drift is a property of drafter designs. We trace this to the un-normalized residual path between chain steps: the drafter's hidden state magnitude grows monotonically with chain depth, which exhibits dynamics consistent with additional pre-norm transformer layers stacked on the target rather than as a standalone autoregressive predictor. In order to limit the growth, we propose two architectural changes: Post-norm on the drafter hidden states and per-hidden-state RMSNorm after capturing target hidden states. Our interventions improve acceptance length over the current leading model, pre-norm EAGLE3, by up to $2\times$ under template perturbation, $1.18\times$ on long-context tasks, and $1.10\times$ on seven standard benchmarks spanning multi-turn chat, math, and coding. Our changes also allow shorter train-time-test depths to generalize over longer drafting sequences.

Large language models (LLMs) often suffer from hallucinations due to error accumulation in autoregressive decoding, where suboptimal early token choices misguide subsequent generation. Although multi-path decoding can improve robustness by exploring alternative trajectories, existing methods lack principled strategies for determining when to branch and how to regulate inter-path interactions. We propose Adaptive Path-Contrastive Decoding (APCD), a multi-path decoding framework that improves output reliability through adaptive exploration and controlled path interaction. APCD consists of two components: (1) Entropy-Driven Path Expansion, which delays branching until predictive uncertainty - measured by Shannon entropy over top candidate tokens - indicates multiple plausible continuations; and (2) Divergence-Aware Path Contrast, which encourages diverse reasoning trajectories while dynamically attenuating inter-path influence as prediction distributions diverge. Experiments on eight benchmarks demonstrate improved factual accuracy while maintaining decoding efficiency. Our code is available at https://github.com/zty-king/APCD.

Autoregressive inference is typically assumed to scale predictably with decoding length, with latency increasing smoothly as generated sequence length grows. In this work, we identify unexpected non-monotonic latency behavior in the Apple MPS backend, where latency changes abruptly across nearby decoding configurations during transformer decoding. Using multiple model families (GPT-2, BLOOM, and OPT), we observe latency spikes of up to 21x within specific decoding-budget intervals, followed by recovery at neighboring configurations. Controlled experiments show that these anomalies originate primarily during the decode phase rather than prefill, are not explained by memory pressure alone, and remain absent on CPU and NVIDIA CUDA backends under identical conditions. We further show that key-value (KV) cache interacts strongly with these pathological execution regimes: KV caching remains beneficial overall, but its practical speedup collapses sharply within anomalous configurations, while cache-disabled decoding still exhibits residual non-monotonic behavior. These findings suggest that autoregressive decoding on MPS enters discrete execution regimes that are not captured by coarse-grained benchmarking, highlighting the importance of hardware-aware evaluation for long-context inference.

MiniMind-O is an open 0.1B-scale omni model built on the MiniMind language model. It accepts text, speech, and image inputs, and returns both text and streaming speech. The release includes model code, checkpoints, and the main Parquet training datasets for text-to-audio, image-to-text, and audio-to-audio training, making the complete interaction loop directly inspectable. The model uses a full MiniMind backbone as the Thinker and an independent four-layer Talker made from MiniMind blocks. Frozen SenseVoice-Small and SigLIP2 encoders provide speech and image features, which are mapped by lightweight MLP projectors and injected at modality-placeholder positions. The Talker reads a middle-layer Thinker state together with an autoregressive eight-layer Mimi-code buffer. Speaker control is handled by a dedicated speaker token, right-aligned reference codec prompts, and precomputed CAM++ speaker embeddings, so voice conditioning remains part of the audio-code context rather than a separate TTS module. With a 768-dimensional Talker, the dense and MoE variants reach average CERs of 0.0897 and 0.0900 in Thinker--Talker consistency evaluation, with overall voice-cloning similarities of 0.5995 and 0.5937. Beyond reporting a working system, the paper identifies three scale-critical design choices for small omni models: middle-layer semantic bridging, a released multimodal sequence format, and a parameter-efficient eight-codebook interface.

Autoregressive decoding becomes bandwidth-limited at long contexts, as generating each token requires reading all $n_k$ key and value vectors from KV cache. We present Stochastic Additive No-mulT Attention (SANTA), a method that sparsifies value-cache access by sampling $S \ll n_k$ indices from the post-softmax distribution and aggregates only those value rows. This yields an unbiased estimator of the post-softmax value aggregation while replacing value-stage multiply-accumulates with gather-and-add. We introduce stratified and systematic sampling to design variance-reduced, GPU-friendly variants. Evaluated on Llama-3.1-8B-Instruct at 32k-token contexts, S$^2$ANTA matches baseline accuracy while achieving up to $1.5\times$ decode-step attention-kernel speedup over FlashInfer and FlashDecoding on an NVIDIA RTX 6000 Ada. In batched long-context generation, these kernel gains translate to up to $1.25\times$ end-to-end decode-latency speedup. Finally, we propose Bernoulli $qK^\mathsf{T}$ sampling as a complementary technique to sparsify the score stage, reducing key-feature access through stochastic ternary queries. Both methods are complementary to upstream quantization, low-rank projection, KV-cache compression, and KV-cache selection methods. Together, they point toward sparse, multiplier-free, and energy-efficient inference. We open-source our kernels at: https://github.com/OPUSLab/SANTA.git

Tree-based speculative decoding accelerates autoregressive generation by verifying multiple draft candidates in parallel, but this advantage weakens for sparse Mixture-of-Experts (MoE) models. As the draft tree grows, different branches activate different experts, expanding the union of activated experts and substantially increasing target-side verification cost. We propose EVICT, a training-free, hyperparameter-free, and lossless adaptive verification method for MoE speculative decoding. EVICT makes every verified token count by truncating the draft tree before target verification and retaining only the cost-effective prefix. It leverages fine-grained drafter signals to estimate candidate benefit, combines them with offline-profiled verification cost, and remains highly compatible with the high-performance graph-based serving framework SGLang. Extensive experiments on diverse MoE backbones and benchmarks show that EVICT achieves up to 2.35x speedup over autoregressive decoding and an average 1.21x speedup over the state-of-the-art baseline EAGLE-3, while significantly reducing unnecessary expert activations during verification.

Real-time automatic speech recognition (ASR) systems face a fundamental trade-off between transcription accuracy and computational efficiency, particularly when deploying large-scale transformer models like Whisper. Existing streaming approaches either sacrifice accuracy through aggressive chunking or incur prohibitive memory costs through unbounded context accumulation. We present WhisperPipe, a novel streaming architecture that achieves bounded memory consumption while maintaining transcription quality through three key innovations a hybrid Voice Activity Detection (VAD) pipeline combining Silero VAD with energy-based filtering to reduce false activations by 34%, a dynamic buffering mechanism with overlapping context windows that prevents information loss at segment boundaries, and an adaptive processing strategy that balances latency and accuracy based on speech characteristics. Evaluated on 2.5 hours of diverse audio data, WhisperPipe demonstrates a median end-to-end latency of 89ms (90th percentile: 142ms) while consuming 48% less peak GPU memory and 80.9% lower average GPU utilization compared to baseline Whisper implementations. The system maintains stable memory usage over extended sessions, with zero growth rate across 150-minute continuous operation. Comparative analysis against related work shows that WhisperPipe achieves competitive accuracy (WER within 2% of offline Whisper) while operating at 3-5x lower latency than existing streaming solutions. The architecture's modular design enables deployment across resource-constrained environments, from edge devices to cloud infrastructure. Our results demonstrate that careful architectural design can reconcile the competing demands of real-time responsiveness and model sophistication in production ASR systems.

KV caching is a ubiquitous optimization in autoregressive transformer inference, long presumed to be numerically equivalent to cache-free computation. This assumption fails under standard FP16 precision: cache-ON and cache-OFF execution paths employ different floating-point accumulation orderings which, due to FP16 non-associativity, produce a deterministic divergence in decoded token sequences. Across three open-weight models (LLaMA-2-7B, Mistral-7B-v0.3, Gemma-2-2B) evaluated on GSM8K, we observe a 100% token divergence rate across all sampling strategies, including greedy decoding, which rules out sampling randomness as a cause, and also with cache-ON yielding higher accuracy in 8 of 9 conditions, where the accuracy difference serves as an indicator that the divergence direction is systematic rather than random. Controlled FP32 falsification reduces divergence by eight orders of magnitude, eliminates token flips, and drops the flip rate to exactly 0.0%, confirming FP16 non-associativity as the sole causal driver. Layer-wise drift profiling reveals architecturally predictable propagation patterns: models using Grouped-Query Attention exhibit sharp divergence at the first layer, while Gemma's larger head dimension and sliding window attention produce uniform accumulation across all layers. Finally, activation patching of the entire residual stream fails to recover the cache-free trajectory, localizing the causal variable to the stateful KV cache. These findings establish that FP16 KV cache inference is fundamentally non-equivalent to recomputation and provide a mechanistic framework for understanding numerical instability in modern LLM inference systems.

Autoregressive decoding in Large Language Models (LLMs) generates one token per step, causing high inference latency. Speculative decoding (SD) mitigates this through a guess-and-verify strategy, but existing training-free variants face trade-offs: retrieval-based drafts break when no exact match exists, while logits-based drafts lack structural guidance. We propose $\textbf{RACER}$ ($\textbf{R}$etrieval-$\textbf{A}$ugmented $\textbf{C}$ont$\textbf{e}$xtual $\textbf{R}$apid Speculative Decoding), a lightweight and training-free method that integrates retrieved exact patterns with logit-driven future cues. This unification supplies both reliable anchors and flexible extrapolation, yielding richer speculative drafts. Experiments on Spec-Bench, HumanEval, and MGSM-ZH demonstrate that RACER consistently accelerates inference, achieving more than $2\times$ speedup over autoregressive decoding, and outperforms prior training-free methods, offering a scalable, plug-and-play solution for efficient LLM decoding. Our source code is available at $\href{https://github.com/hkr04/RACER}{https://github.com/hkr04/RACER}$.

Neural audio codecs have largely promoted the application of language models (LMs) for speech applications. However, the effectiveness of autoregressive LM-based models in unifying speech enhancement (SE) tasks remains underexplored. In this work, we propose UniSE, a unified decoder-only LM-based framework to handle different SE tasks including speech restoration, target speaker extraction, and speech separation. Conditioned on input speech features, it autoregressively generates target discrete tokens, facilitating compatibility between distinct learning patterns of multiple tasks. To further optimize speech quality, we introduce a progressive reinforcement learning strategy with multiple assessment criteria. Experiments on several benchmarks show that UniSE achieves competitive performance compared to discriminative and generative baselines, demonstrating the capacity of LMs in unifying SE tasks. The code and demo are available at: https://github.com/alibaba/unified-audio/tree/main/QuarkAudio-UniSE.