Language Model Decoding

Follow this topic Ask about this topic

Momentum

18 papers in the last four weeks, up 200% on the four weeks before. 0.2% of all new papers.

Jul 13Week of Sep 28

Latest papers 108

Jul 27, 2026cs.AI

Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation

Semantic IDs (SIDs) are now a central component of generative recommendation. Current SID-based systems assign three roles to the same token sequence. Shared prefixes are intended to organize related items, the complete SID identifies an individual item, and each generated token narrows the items that can still be returned. We systematically investigate SIDs from item encoding and SID construction to autoregressive generation and final recommendation. We examine how SID construction changes item representations and how those changes affect generation. Across three Amazon domains and eight SID constructions, SID neighborhoods recover only 32.2% of the encoder's ten nearest neighbors on average. Alternative item descriptions still retrieve the corresponding item first in 99.57% of controlled cases, yet change 38.4% of exact SIDs. These results show that SIDs retain broad organization but lose much of the encoder's fine local structure, while their exact tokens are not determined by item meaning alone. This loss becomes consequential during generation. After the final semantic token, TIGER retains only 29.9% of held-out targets that were plausible recommendations before SID filtering. Motivated by these findings, we propose Item-Supported Decoding (ISD), a lightweight inference-time method that allows a user-specific item ranking to support corresponding SID prefixes before beam search discards them. The same ranking then orders the generated items. ISD requires no additional parameters or retraining of the SID constructor or decoder. We empirically show that ISD improves NDCG@10 over the corresponding SID backbone in every evaluated setting, with relative gains of up to 31.2%. Our results show that SIDs provide useful coarse item organization, but their fine boundaries should not alone determine which items remain available during generation.

Junting Wang, Xinrui He, Yunzhe Li +1

University of Illinois Urbana-Champaign Urbana, Illinois, USA

Semantic ID Recommendation Language Model Decoding Generative Recommendation

Jul 27, 2026cs.CL

SyRuP: Enhancing System-Prompt Following via Reward-Guided Prediction in LLM Decoding

Large Language Models (LLMs) are increasingly controlled through system prompts that specify roles, formats, and safety requirements. However, models follow these prompts only implicitly through in-context learning, which can be insufficient for complex or compositional prompts. Existing approaches often require model tuning or response-level reranking, limiting their practicality for lightweight inference-time control. We introduce SyRuP, a decoding-time framework for improving system-prompt adherence while keeping the base LM frozen. SyRuP trains a cross-attention reward head from system-prompt-conditioned preference pairs, treating the system prompt as a separate memory to produce token-level adherence scores. At inference, SyRuP reranks the base LM's top-k candidates by combining base logits with both the learned reward signal and a contrastive signal that captures system-induced logit shifts. Experiments on system-prompt following benchmarks show that SyRuP consistently outperforms prompting and decoding-time baselines with moderate inference overhead. These results suggest that explicit token-level guidance is an effective and practical mechanism for reliable system-prompt following.

Seoyeon Kim, Minjae Kang, Jaehyung Kim

Yonsei University

Language Model Decoding Language Model Steering Instruction Following

Jul 22, 2026cs.AI

Refusal-Gated Decoding: Preserving Refusal Behavior Under High-Temperature Sampling

Recent advances in truncation-based sampling have helped mitigate drawbacks of high-temperature sampling such as neural text degeneration, thereby enabling greater diversity without sacrificing coherence. However, increasing the entropy of the token probability distribution via high temperatures has also been shown to weaken the model's refusal response. Existing solutions for maintaining the refusal behavior of LLMs either replace the model's own refusal decision with a separate safety classifier or alter its output distribution for every prompt. To address this gap, we propose refusal-gated decoding (RGD): an efficient sequential decoding approach which preserves a model's greedy decoding refusal response at high temperatures and samples all other prompts from its exact direct high-temperature distribution, while incurring minimal additional latency. RGD runs a short greedy probe that reuses the prompt's KV cache and exits as soon as it becomes incompatible with a learned set of refusal prefixes; it returns the greedy response if the probe remains compatible and otherwise discards the probe and samples from the original prompt. Across seven models and three benchmark datasets at T=2.0, RGD raises greedy-refusal preservation from 91.9% under direct sampling to 98.3% on average while adding only 2.2-4.3% to the median per-request latency of non-refusals across temperatures. Unlike prompt-screening baselines which route many greedy non-refusals to greedy decoding, RGD keeps at least 98.1% of greedy non-refusals on unchanged high-temperature sampling, thereby preserving the model's natural high-temperature sampling behavior. We also propose a residual-stream variant of our method which lowers this latency overhead to at most 0.5% with comparable prompt routing accuracy. Our work shows that unlocking greater diversity via high-temperature sampling need not erode a model's refusal behavior.

Phillip Howard, Xin Su, Allen Roush +3

Thoughtworks

Language Model Decoding LLM Refusal Behavior Refusal Behavior in Language Models

Jul 21, 2026cs.CL

Speculative Correction: Draft-then-Refine Decoding for Diffusion Language Models

Diffusion language models (DLMs) can revise tokens bidirectionally, but standard decoding procedures often adapt them to left-to-right generation by producing text block by block. We study a simple plug-and-play inference pattern: first generate a complete draft, then refine the full response using bidirectional diffusion. Using LLaDA2.1-Flash and LLaDA2.1-Mini, we evaluate two configurations. In Flash-Flash, the same Flash model serves as both drafter and refiner, testing whether an existing model can improve its own block-autoregressive output through global refinement. In Mini-Flash, inspired by speculative decoding, we introduce speculative correction: Mini drafts a full response, and Flash revises it as an editable initialization. Flash-Flash improves GSM8K-384 accuracy from 0.848 to 0.899 while running 1.20 times faster than the selected Flash block-autoregressive baseline, and improves MBPP-384 from 0.545 to 0.693. Latency-window-matched Flash-only controls indicate that these gains persist after targeted tuning of block-autoregressive decoding. Causal ablations indicate that completed drafts provide useful initializations: refinement from a fully masked span performs poorly, full global refinement provides a clear additional gain on GSM8K, and local refinement captures much of the gain on MBPP and MATH. Mini-Flash provides useful quality-latency trade-offs, including MATH-384 performance of 0.294 versus 0.300 for Flash while running 2.17 times faster. These results support a Pareto-frontier interpretation rather than the claim that the heterogeneous cascade uniformly matches Flash quality. Overall, same-model draft-and-refine provides evidence that bidirectional refinement is a useful decoding primitive for DLMs, while speculative correction demonstrates a training-free route to fast DLM generation.

Brian K Chen, Chong Wu, Kenji Kawaguchi

National University of Singapore · City University of Hong Kong

Language Model Decoding Speculative Decoding for Diffusion Language Models Diffusion Language Model Decoding Diffusion Language Models Diffusion Model Inference Acceleration

Jul 17, 2026cs.AI

Bayesian Repetition Penalty: A Principled Adjacent-Conditional Framework for Reversing Attention Collapse in Autoregressive Language Models

Attention collapse in autoregressive language models -- manifested as repetitive token loops where the model becomes trapped in self-reinforcing attractors -- is a persistent pathology that existing decoding-time heuristics fail to address at its root cause. We present a principled framework that penalises or compensates anomalous confidence arising from collapsed generation patterns, by comparing a token's observed frequency against its corpus prior through an adjacent-conditional probability construction. The resulting self-normalising penalty ratio

R=f(m,n,p)/f(np,n,p)

requires no ad hoc standardisation and admits a closed-form logit offset with zero approximation error. The correction is isolated from the loss gradient and accumulated into a frozen output-layer bias via exponential moving average, enabling deployment as a repair mechanism for models that have already collapsed without requiring intrusive modifications to standard training pipelines. Experimental validation on a 1.5B-parameter model demonstrates that the frozen-bias mechanism can rescue a model already trapped in a collapsed attractor, reducing 2-gram repetition from 0.073 to near 0 while preserving generation quality.

Wenjie Fan, Bin Ma, Dong Li

Yotta Labs

Language Model Decoding Autoregressive Language Modeling Text Generation

Jul 17, 2026cs.CL

Adaptive Multi-Step Lookahead Decoding for Diffusion Language Models

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.

Yingqian Cui, Wei Deng, Lantao Mei +4

Michigan State University · Morgan Stanley · IBM T.J. Watson Research Center

Masked Diffusion Models Language Model Decoding Diffusion Language Model Decoding Diffusion Model Inference Acceleration

Jul 13, 2026cs.AI

Calibrated e-CUSUM Decoding for Quantized Reasoning Models: Why Token Log-Probability Is the Wrong Observable for Decoding Monitors

Low-bit quantization makes small reasoning models inexpensive to deploy but can degrade their chains of thought. This motivates decoder-side monitors that intervene when generation becomes unreliable. We show that a natural candidate, the centered token log-probability increment

\log p(w_t)+H_t

, is the wrong observable for this purpose. Under the model's own sampling law it is a mean-zero martingale by construction, so it measures sampling self-consistency rather than trajectory health and is nearly silent during confident repetition, where both

\log p(w_t)

and entropy are close to zero. We introduce a training-free decoding controller that combines (i) a degeneration-aware alarm score fusing token uncertainty with explicit verbatim repetition and (ii) a calibrated e-process-inspired sequential detector. The raw product process is Ville-valid under a conditional-mean null, while the deployed CUSUM-floored statistic is treated as an empirical change detector because the score is history-dependent and autocorrelated. On GSM8K with DeepSeek-R1-Distill-Qwen-1.5B in FP16 and INT4, calibration turns a monitor that fires on 93--95% of generations into a selective detector of failing traces (

φ\approx 0.3

, precision

\approx 0.6

against a 0.38 base rate). In this pilot, the controller reduces measured verbatim-degeneration signals and yields a positive but statistically inconclusive INT4 accuracy change from 63% to 69% (paired McNemar

p=0.18

,

n=100

), at a 28% token-budget cost. We also find that non-termination, rather than looping, is the dominant failure mode on GSM8K. The main contribution is methodological: an explanation of why centered token log-probability is inadequate for decoder monitoring and a calibrated, cautiously evaluated replacement.

El Hassane Ettifouri, Ayoub Belfatmi, Mahaman Sanoussi Yahaya Alassan +1

Novelis Research, Paris, France

Language Model Decoding Change-Point Detection

Jul 9, 2026cs.LG

Gauge dependence and structured-output corruption in sign-branched repetition penalties: measurements across models, inference stacks, and alternative repetition controls

The multiplicative repetition penalty shipped across the LLM inference ecosystem (HuggingFace, vLLM, llama

.

cpp, and a dozen further engines) branches on the sign of each raw logit (divide positives by theta, multiply negatives). But the softmax is unchanged by adding a constant to every logit, so a model's logit zero-point is arbitrary (a gauge choice), and the sign-branch reads it. Two measurable consequences follow. (1) The penalty is not well-defined: re-centering a model's logits by a constant is a provable no-op at theta=1, yet at a routine theta=1.3 it changes 58-96% of greedy tokens, while subtractive and normalized penalties change none; real checkpoints sit at widely different zero-points, so a fixed repetition_penalty is a different operation on every model. (2) It corrupts structured output: on 200 real-world JSON schemas, theta=1.3 drops the rate of valid, schema-conformant output from 97% to 23%. Applying the penalty to normalized log-probabilities instead of raw logits removes the gauge dependence by construction; HuggingFace's beam search has applied its processor chain, penalty included, to log-probabilities since at least v4.0.0, so repetition_penalty already names two different operators depending on decoding strategy. Because equal theta is not equal strength across the two operators, we also compare them at matched suppression, calibrated per model by search: there the normalized operator is statistically no worse on any quality metric measured, but on four of seven models it cannot match the raw operator's suppression at theta >= 1.15, and on six of seven at theta=1.3, the setting where the corruption was measured. This note gives the mechanism, the measurements (five models up to 7B; two code models; both effects replicated inside vLLM and llama

.

cpp through their own samplers), the per-model calibration map, and the normalized variant.

Peter Hollows

Language Model Decoding LLM Inference LLM Reliability

Jul 6, 2026cs.LG

Noisy-Channel Minimum Bayes Risk Decoding

Minimum Bayes Risk (MBR) decoding yields more robust and higher-quality text generation than maximum a posteriori (MAP) decoding by selecting hypotheses that maximize expected utility over sampled pseudo-references. However, there exists a discrepancy in the design: hypothesis selection calculates expected utility scores conditioned on given pseudo-references, while commonly used evaluation metrics, e.g., BLEU and COMET, are asymmetric. Therefore, it is important to consider both hypothesis-to-reference and reference-to-hypothesis directional effects. In this study, we introduce a noisy channel decomposition of MBR decoding that naturally incorporates bidirectional effects to account for these asymmetries. We decompose MBR decoding into four interacting components: hypothesis-to-reference likelihood, reference-to-hypothesis likelihood, hypothesis prior, and reference prior. This decomposition provides a unified interpretation of existing MBR variants and enables metric- and task-specific interpretability by isolating the contribution of each channel. Our comprehensive analysis reveals that channel-wise contributions exhibit distinct characteristics across metrics while remaining consistent across tasks, and suggests that appropriate channel weighting may lead to improvements over original MBR decoding.

Yusuke Sakai, Hidetaka Kamigaito, Taro Watanabe

Nara Institute of Science and Technology, Nara, Japan.

Language Model Decoding MBR Decoding

Jul 3, 2026cs.LG

What Does a Routing Oracle Measure Under Stochastic Decoding? Coupling, Scorer Choice, and Single-Commit Ceilings

Routing benchmarks often compare a policy that commits to one model before seeing its response with a hindsight oracle that credits any correct recorded output. Under stochastic decoding these are different decision classes. For success marginals

p_{im}

under a frozen query--model--decoder--scorer protocol, we distinguish the clairvoyant single-commit ceiling

R_i=\max_m p_{im}

, product-coupling union

U_i^\perp=1-\prod_m(1-p_{im})

, and premium

Δ_i^\perp=U_i^\perp-R_i

. The marginals alone identify exactly the union interval

[R_i,\min\{1,\sum_m p_{im}\}]

: zero premium is attainable over compatible couplings, not established for an actual deployment. If generation is independent across models, product is the specified protocol's union probability. We audit frozen correctness tensors from 11 open models and 30 archived responses per query--model cell on GSM8K, MATH-500, and GPQA-Diamond. Full-pool display-channel product premiums are 0.371, 3.542, and 5.100 percentage points; the premium intervals from empirical marginals are

[0,0.473]

,

[0,4.787]

, and

[0,7.744]

. Their zero lower endpoints are algebraic. All retained scorers, eight finite-draw paths, and all 2,047 nonempty subpools expose scorer, estimator, and pool sensitivity, not confidence intervals. The GSM8K and GPQA display scorers were developed after limited output inspection. A separate retrospective held-out policy illustration instantiates the policy-specific gap decomposition without establishing new-data generalization. A limited reference-based human check supports scorer agreement only on definite-consensus subsets. Hash-bound evidence supports number checks, not end-to-end reproduction. The contribution is a measurement contract for interpreting oracle gaps conditional on coupling, decision class, scorer, pool, and finite draws, not a population effect or an equal-cost routing gain.

Teng-Ruei Chen

Institute of Bioinformatics and Systems Biology, National Yang Ming Chiao Tung University, Hsinchu 300, Taiwan · Krixvon AI, Taiwan

LLM Evaluation Language Model Decoding LLM Routing

Jul 2, 2026cs.LG

Set Diffusion: Interpolating Token Orderings Between Autoregression and Diffusion for Fast and Flexible Decoding

Discrete diffusion models have steadily improved in quality relative to autoregressive (AR) models. However, these models are normally constrained to fixed-length generation and do not support key-value (KV) caching. Block diffusion partially bridges diffusion and AR by generating token blocks left-to-right, but its fixed-size sequential blocks limit decoding flexibility and parallelism. Here, we present a new class of language models, set diffusion, comprised of (i) a likelihood parameterization that factorizes over flexible-position, flexible-length token sets and (ii) a set-causal diffusion architecture that supports KV cache updates after every inference step. By factorizing over token sets instead of fixed-size blocks, tokens can be decoded in arbitrarily-ordered sets, including sliding-window sets, enabling faster inference and support for any-order decoding. Set diffusion achieves better speed-quality tradeoffs on mathematical reasoning, summarization, and unconditional generation compared to prior diffusion language models while offering stronger infilling performance than block diffusion. We provide the code, along with the model weights and blog post on the project page: https://m-arriola.com/setdlms/

Marianne Arriola, Volodymyr Kuleshov

Department of Computer Science, Cornell University, Ithaca, NY, USA.

Language Model Decoding Diffusion Language Model Decoding Diffusion Model Caching Diffusion Language Models Diffusion Model Inference Acceleration Diffusion Models

Jun 28, 2026cs.LG

Depth Exploration for LLM Decoding

Autoregressive LLM decoding evaluates every generated token through the full layer stack, even though many tokens become predictable at intermediate depths. Existing lossless depth-adaptive methods exploit this redundancy by choosing a single non-final exit depth and verifying its prediction with the final-depth model. However, our measurements show that this selection-based strategy leaves substantial headroom: choosing an exit too late wastes computation, while choosing one too early triggers fallback and discards dependent drafts. We propose Depth Exploration Decoding (DEX), a lossless decoding algorithm that replaces single-depth selection with parallel exploration over multiple candidate depths. At each commit position, DEX validates candidates against the final-depth reference, commits exactly the final-depth token, and collapses the exploration lattice to retain only reusable branch states. This expand--commit--collapse procedure preserves equivalence to standard autoregressive decoding while reducing the cost of committing each token. Across early-exit-trained and standard LLMs, DEX outperforms representative depth-selection baselines and achieves competitive end-to-end throughput against speculative and distributed decoding methods. Moreover, DEX improves as the explored depths become finer, showing that parallel depth exploration provides a scalable way to exploit the underused depth axis of LLM decoding.

Weisi Yang, Zipeng Sun, Stephen Xia

Northwestern University · MILA – Quebec AI Institute · McGill University

Language Model Decoding LLM Inference Acceleration Autoregressive Decoding

Jun 25, 2026stat.ML

When are likely answers right? On Sequence Probability and Correctness in LLMs

Many decoding methods for large language models can be understood as shifting probability mass toward outputs that are more likely under the model, either locally at the token level or globally at the sequence level. Therefore, their success depends on a fundamental question: when does sequence probability, that is, the conditional probability of a continuation given a prompt, actually align with correctness? In this paper, we set out to quantify this relationship across decoding methods, models, and benchmarks at four levels: across decoding methods, across hyperparameters within a method, across prompt-answer pairs within a dataset, and across repeated responses to the same prompt. We find that higher sequence probability is often predictive of correctness across prompt-answer pairs within a fixed dataset. However, this relationship does not generally transfer to decoding decisions: increasing sequence probability by changing hyperparameters or methods does not reliably improve accuracy. Further, sequence probability is not a good indicator of correctness for responses to the same prompt. These findings clarify when decoding can and cannot be expected to improve correctness, and provide practical guidance for decoding, self-consistency, and verifier-free self-improvement.

Johannes Zenn, Jonas Geiping

Max Planck Institute for Intelligent Systems · ELLIS Institute Tübingen · AI Center Tübingen +2

LLM Evaluation Language Model Decoding LLM Reliability

Jun 22, 2026cs.LG

Scheduling Thoughts: Learning the Order of Thought in Diffusion Language Models

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

Jiawei Xu, Minghui Liu, Aakriti Agrawal +2

University of Maryland, College Park · University of California, Los Angeles.

Masked Diffusion Models Language Model Decoding LLM Inference Scheduling Diffusion Language Model Decoding RL for Diffusion Models Diffusion Language Models+1

Jun 22, 2026cs.CL

The Anatomy of the CTC Oracle Gap: Acoustic Exhaustion and Linguistic Recovery

We study the limits of CTC-internal scoring for N-best hypothesis selection and locate the information bottleneck separating acoustic confidence from linguistic plausibility. Eleven CTC-internal and acoustic-feature scoring strategies produce no statistically significant WER improvement over greedy decoding on LibriSpeech dev-other at G=16 (all p > 0.05). The exhaustion is systematic: CTC's Spearman

ρ

between hypothesis score and per-utterance WER degrades from -0.574 at G=4 to -0.270 at G=128, a 53% loss driven by blank-path proliferation. This establishes that the discriminative capacity of CTC-internal representations is saturated: no recombination of acoustic signals can close the oracle gap. Confirming that the bottleneck is linguistic, not acoustic, external linguistic information introduced via MBR decoding breaks through it. MBR-CER decoding with a RoBERTa pseudo-log-likelihood (PLL) posterior (

τ

=10, G=128) achieves 5.42% WER on held-out LibriSpeech test-other (greedy 5.96%,

Δ

=-0.535 pp, p<0.0001, 9.0% relative). RoBERTa PLL

ρ

degrades only 21% over the same range, retaining discriminating power where CTC loses it. Applied without retuning across two Zipformer architectures, three domains (LibriSpeech, TED-LIUM 3, VoxPopuli), and four MUSAN noise levels, the recipe gives significant gains in 11 of 13 conditions. On the training side, standard MWER training via the CTC forward-backward algorithm implements Rao-Blackwellized REINFORCE at the output projection (variance about 3x below Viterbi). Yet sequence-level fine-tuning fails at near-converged checkpoints: all four MWER configurations on CR-CTC collapse (+6.18 to +8.90 pp WER), as a training oracle gap of 0.007 pp provides no usable reward signal.

Ivan Novosad

Faculty of Computer Science, HSE University, Moscow, Russia.

ASR Evaluation Language Model Decoding MBR Decoding

Jun 21, 2026cs.CL

Breaking the Likelihood Trap: Variance-Calibrated Modulation for Large Language Model Decoding

In open-ended generation, LLMs frequently fall into the "likelihood trap", characterized by repetitive degeneration and vocabulary dullness, resulting in a discrepancy between machine-generated and human-written text. While post-hoc tail truncation (e.g., Top-p, Min-p) avoids sampling from the unreliable tail, it can misalign generation with human lexical preferences by over-sampling from the uncalibrated head; fixed scalar repetition penalties, in turn, ignore how the scale of the logit distribution varies across inference steps, which can disrupt semantic coherence. To address both shortcomings, we propose Variance-Calibrated Modulation (VCM), a training-free pre-decoding intervention. VCM directly reshapes the probability distribution prior to truncation via two dynamic mechanisms: (1) Contextual Searchlight via PMI, which naturally suppresses global stopwords and elevates context-evoked tokens, and (2) Adaptive Self-Debiasing, which utilizes real-time logit standard deviation to provide scale-invariant penalization. In experiments across open-ended generation, factual QA, and mathematical reasoning, we show that VCM consistently mitigates the likelihood trap. With negligible computational overhead, VCM integrates with existing decoding strategies, improving diversity and coherence and, particularly at higher decoding temperatures, reasoning accuracy. Our code is publicly available on GitHub: https://github.com/AetherDing/VCM

Yuanhao Ding, Meimingwei Li, Esteban Garces Arias +3

School of Computer and Information Engineering, Henan University · Department of Statistics, LMU Munich · 3Munich Center for Machine Learning (MCML)

Open-Ended Generation Language Model Decoding Text Generation

Jun 20, 2026cs.CL

Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding

Autoregressive generation in large language models (LLMs) conventionally decodes from the final layer, assuming that deeper representations yield more reliable next-token predictions. We revisit this assumption by revealing a recurring Guess-Refine-Perturb dynamic: early layers form coarse guesses, intermediate layers refine reasoning-relevant semantics, and final layers can perturb these refined predictions toward generic or alignment-preferred tokens. We introduce Confident Decoding, a training-free decoding strategy that dynamically selects the most reliable near-final layer through entropy-guided conservative backward search. We further provide a theoretical formulation of layer selection as an optimal stopping problem, showing that under bounded projection noise and dominant late-stage alignment perturbation, our search rule filters perturbation while bounding the loss relative to the oracle refinement layer. Experiments across dense and Mixture-of-Experts LLMs demonstrate consistent gains on challenging reasoning benchmarks, including GPQA-Diamond, Omni-MATH, and HLE, with zero memory overhead and less than 2% latency increase. These results suggest dynamically bypassing final-layer perturbations can unlock stronger reasoning behavior from aligned LLMs.

Xuanming Zhang, Sining Zhoubian, Yuxuan Chen +8

Qwen Team, Alibaba Inc. · Tsinghua University · Nanyang Technological University

LLM Alignment Language Model Decoding Optimal Stopping LLM Reasoning

Jun 15, 2026cs.CL

From Awareness to Adherence: Bridging the Context Gap in Spoken Dialogue Systems via Context-Aware Decoding

Despite the success of end-to-end (E2E) spoken dialogue systems, maintaining strict context adherence in multi-round conversations remains a challenge. While prior works attribute these failures to models forgetting dialogue history, we highlight an equally critical but overlooked bottleneck: a gap between latent context awareness and active adherence. Although models internally recognize relevant past utterances, strong parametric priors often overshadow these signals during decoding. To bridge this gap, we propose an audio-adapted Context-Aware Decoding (CAD) approach. By leveraging internal attention mechanisms to isolate key historical rounds, our approach contrasts output distributions with and without this key context during inference, directly amplifying multimodal contextual signals. Evaluations on the Audio MultiChallenge benchmark demonstrate significant improvements in Semantic Memory and Self Coherence subtasks, successfully enforcing strict, context-faithful adherence.

Che Hyun Lee, Heeseung Kim, Sungroh Yoon

ECE and 2 IPAI, Seoul National University, Seoul 08826, Korea · Department of AI, University of Seoul, Seoul 02504, Korea

Language Model Decoding Spoken Dialogue Systems

Jun 15, 2026cs.CL

Tyler: Typed Latent Reasoning for Language Models -- When to Think, What to Compute, and How Much to Allocate

Chain-of-thought (CoT) prompting improves reasoning in large language models (LLMs) by externalizing intermediate computation as discrete text tokens, but this textual interface also introduces redundancy and inference overhead. Latent reasoning offers a promising alternative by carrying part of the computation in continuous representations. However, existing methods typically predefine when latent computation is invoked and how it is allocated during decoding, leaving a key problem unresolved: when to invoke latent computation, what type of computation to perform, and how much budget to allocate. We propose \textbf{Ty}ped \textbf{L}at\textbf{e}nt \textbf{R}easoning (Tyler), a typed and budget-aware framework for latent reasoning during autoregressive decoding. Tyler learns a policy that, at each decoding step, chooses between emitting a text token and switching to a latent computation module specialized for a particular reasoning function. Once invoked, an operator maps the current reasoning state into latent tokens that support global planning, local state updates, or reusable procedural abstraction. Across extensive experiments on three backbone LLMs, Tyler improves accuracy by up to 14.49 points over CoT and by up to 4.30 points over the strongest competing baseline. It further generalizes across diverse reasoning domains and achieves the best final-stage performance with the lowest forgetting.

Hanyu Lin, Min Cai, Jiawei Wen +1

1Shenzhen University · University of Alberta

Language Model Decoding Efficient Language Model Reasoning Continuous Latent Reasoning LLM Reasoning Adaptive Computation Latent Reasoning

Jun 12, 2026stat.ML

Adaptive Nucleus Truncation for Long-Form Reasoning

Sampling plays an important role in long-form language-model reasoning. Over thousands of decoding steps, small changes in the candidate token set can compound into different reasoning trajectories, stability profiles, and final answers. Existing truncation methods such as top-

p

, min-

p

, and fixed top-

nσ

sampling improve over unrestricted sampling, but they rely on fixed thresholds that cannot adapt to changes in entropy, task difficulty, training stage, or generation budget. We introduce Adaptive Nucleus Truncation Sampling (ANTS), which extends top-

nσ

sampling from a fixed decoding rule into an adaptive rollout-control mechanism for long-form generation. ANTS selects standardized neighborhoods around the maximum logit before temperature scaling, adapts the truncation width using an entropy-conditioned controller, and retains a no-truncation fallback arm to stabilize training when truncation becomes unsafe. On a 33B-total / 4B-active sparse Mixture-of-Experts reasoning model, ANTS improves average performance over percentage-based benchmarks by +1.9, +3.8, and +5.2 points at 8K, 16K, and 32K generation budgets, respectively. The strongest gains appear on instruction following and mathematical reasoning, with IFBench improving by more than 10 points at 32K and AIME 2025 improving by 7 points. Code generation reveals an important budget interaction. On Codeforces, ANTS trails the baseline at 8K, but reverses this gap and substantially improves ELO at 16K and 32K. These results suggest that sampler design should be treated not just as a decoding hyperparameter, but as part of how we stabilize and scale long-budget reasoning.

Ousmane Amadou Dia

Language Model Decoding Adaptive Sampling LLM Reasoning Large Reasoning Models

Jun 11, 2026cs.LG

ReSET: Accurate Latency-Critical NVFP4 Reasoning via Step-Aware Temperature Scaling

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.

Sihwa Lee, Janghwan Lee, Donghoon Yoo +4

Hanyang University, Seoul, Republic of Korea · Xenoscube Korean Inc., Seoul, Republic of Korea

LLM Quantization Language Model Decoding GPU Kernel Optimization LLM Inference Acceleration Large Reasoning Models

Jun 10, 2026cs.CL

A Controlled Study of Decoding-Time Truthfulness Methods on Instruction-Tuned LLMs

Decoding-time truthfulness methods -- layer-contrast decoding, inference-time intervention, and learned logit adapters -- have demonstrated 10-30 point gains on TruthfulQA when applied to base language models. However, modern instruction-tuned LLMs already achieve substantially higher baselines (61-76%), raising the question of whether these methods remain effective in practice. We design a six-control evaluation framework -- out-of-distribution training, multi-judge validation, simple decoding baselines, confound controls, bootstrap confidence intervals, and seed variance -- and apply it across 5 models (1B-70B), 3 benchmarks, and 15 methods. We find that previously reported gains shrink substantially under strict controls: on the full TruthfulQA benchmark (N=817), no token-level method achieves statistically significant improvement, and the best learned adapter scores -2.0 points below greedy (p=.23). We identify five evaluation sensitivities -- contamination, judge choice, missing baselines, confounds, and statistical noise -- that individually or jointly account for these discrepancies. Cross-benchmark validation on HaluEval QA and TriviaQA confirms that these patterns extend beyond TruthfulQA. Deliberative prompting methods (chain-of-thought, self-critique) appear more robust in the evaluated regime, with CoT achieving +5.6-19pp across benchmarks as a training-free, single-pass method. We release a seven-point evaluation checklist and discuss implications for future truthfulness research.

Ao Sun

Independent Researcher

LLM Evaluation Language Model Decoding Question Answering

Jun 9, 2026cs.LG

CLP: Collocation-Length Prediction for Zero-Loss Adaptive Multi-Token Inference

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.

Xuezhen Xie, Zhiqiang Zhou

Multi-Token Prediction Language Model Decoding Next-Token Prediction LLM Inference Acceleration

Jun 9, 2026cs.LG

K-Forcing: Joint Next-K-Token Decoding via Push-Forward Language Modeling

Autoregressive (AR) language modeling is the dominant paradigm for text generation, yet its sequential token-by-token decoding makes inference memory-bound and inefficient. Existing acceleration approaches, such as speculative decoding and diffusion language models, can yield speedups under certain conditions but do not directly address high-load batch serving--the scenario most critical for industrial-scale deployment. We introduce K-Forcing, a push-forward language modeling paradigm for joint next-k-token decoding. K-Forcing distills an existing AR model into a conditional push-forward mapping--one that transforms independent uniform noise variables into a joint sample of multiple future tokens in a single forward pass. This design preserves fixed-length outputs, reuses the AR teacher backbone, and remains compatible with standard AR serving infrastructure. We train this mapping via progressive self-forcing distillation, which gradually expands the prediction window while enabling the student to closely match the sequence distribution of the AR teacher. We evaluate K-Forcing on LM1B and OpenWebText using a standard causal Transformer backbone. When aggressively configured to generate k = 4 tokens per forward pass, K-Forcing delivers approximately 2.4-3.5x speedup across different batch sizes, while incurring modest quality degradation relative to its AR teacher. As inference increasingly dominates the lifetime compute cost of modern LLMs, K-Forcing offers a promising route toward accelerating AR generation under real-world high-load deployment.

Zhiwei Tang, Yuanyu He, Yizheng Han +4

1DAMO Academy, Alibaba Group · 2Hupan Lab · 3Zhejiang University +1

LLM Serving Language Model Decoding Autoregressive Language Modeling LLM Inference Acceleration Language Model Distillation

Jun 9, 2026cs.IR

STORM: Stepwise Token Optimization with Reward-Guided Beam Search

Modern retrieval increasingly relies on dense and learned-sparse neural models that are effective but require encoding the entire corpus into a specialized index, rebuilt whenever the model changes. Lexical retrievers like BM25 stay efficient and transparent on a standard inverted index that need not change as models evolve, but suffer from vocabulary mismatch. LLM query rewriting can help, yet prompted rewriters emit well-formed but retrieval-ineffective or harmful-terms, and training against a retrieval reward gives only delayed, sequence-level supervision that obscures which terms helped. We introduce STORM (Stepwise Token Optimization with Reward-guided beaM search), a self-supervised framework for lexical query expansion. STORM trains the rewriter through generation guided by retrieval metrics: at each step, candidate expansions are scored against the BM25 index and low-reward continuations pruned, turning the retrieval reward into a token-level signal that concentrates exploration on retrieval-effective vocabulary. Across TREC DL and BEIR, STORM lets 0.6B-8B backbones match or surpass competitive LLM rewriters while retrieving as fast as plain BM25; at 8B it rivals far larger proprietary rewriters. It further transfers zero-shot to 18 languages (MIRACL), beating dedicated multilingual dense retrievers on average, making STORM a competitive, infrastructure-light alternative to dense neural retrieval.

Arthur Satouf, Giulio D'Erasmo, Yuxuan Zong +3

MILA – Quebec AI Institute & ILLS, Canada · 5Sapienza, University of Rome, Italy · 4Sorbonne Université & ISIR & CNRS, France +2

Language Model Decoding Lexical Retrieval Information Retrieval Query Expansion

Jun 8, 2026cs.CL

Gradient-Guided Reward Optimization for Inference-time Alignment

Ensuring the reliability of Large Language Models (LLMs) under distribution drift requires inference-time adaptation. While inference-time alignment methods such as Best-of-

N

and rejection sampling are widely used, they frame the task as a sampling-intensive, reward-guided search, leading to two key limitations: their performance is bounded by the base model's generation quality, and their reliance on imperfect reward models makes them vulnerable to reward hacking. To address these challenges, we introduce Gradient-Guided Reward Optimization (GGRO), a lightweight inference-time method that performs targeted, minimal intervention during decoding via gradient guidance. Specifically, GGRO monitors token-level entropy to identify high-uncertainty regions indicative of drift or misalignment. Upon detection, it responds by injecting nudging tokens, generated using gradient signals from an off-the-shelf reward model, to steer the generation trajectory rather than merely re-ranking samples. Experiments show that GGRO consistently improves inference-time alignment across safety, helpfulness, and reasoning benchmarks. It also increases coverage of high-quality responses and robustness to reward hacking, with minimal computational overhead. Code is available at https://github.com/lhk2004/GGRO.

Hankun Lin, Ruqi Zhang

Department of Computer Science, Purdue University

Reward Hacking LLM Alignment Language Model Decoding Inference-Time Language Model Alignment Inference-Time Guidance

Jun 8, 2026cs.CR

Steganography Without Modification: Hidden Communication via LLM Seeds

We demonstrate that widely deployed Large Language Model (LLM) inference stacks harbor a steganographic channel that requires no modification to model weights, sampling code, or output distributions. The channel exploits a structural property of deterministic decoding: pseudo-random number generators (PRNGs) used in inverse-transform sampling produce a seed-dependent sequence of token-level probability intervals that can be reconstructed from the generated text alone. A sender encodes a secret message in the PRNG seed before generation; a receiver reconstructs the intervals and recovers the seed, and thus the hidden payload, by exhaustive search over the seed space. We formalize two operational modes. In the known-prompt setting, sender and receiver share the prompt, enabling exact interval reconstruction and perfect seed recovery via forced alignment. In the unknown-prompt setting, only the generated text is available; approximate interval reconstruction combined with a maximum-hit-count scoring strategy still permits reliable recovery from sufficiently long outputs. Extensive experiments across six model families and five heterogeneous text domains show that, in the known-prompt setting, full 32-bit seed recovery from the complete 2^32 candidate space achieves up to 100% accuracy, depending on model and text domain, within 300 tokens and under 35 seconds on a single GPU. In the unknown-prompt setting, recovery reaches near-perfect accuracy at 600-800 tokens in about 12 seconds. We further analyze the influence of prompting strategies, tokenization ambiguities, and sampling hyperparameters on channel reliability. Moreover, we discuss several applications of our results: First, it allows for the steganographic transmission of 32 bits, but also shows that ignorance of the prompt is not a valid security assumption.

Felix Mächtle, Jonas Sander, Sebastian Berndt +3

Institute for IT Security, University of Lübeck, Lübeck, Germany · Technische Hochschule Lübeck, Lübeck, Germany

Language Model Decoding LLM Inference LLM Security

Jun 8, 2026cs.AI

The Token Not Taken: Sampling, State, and the Stochasticity of AI Agents

Agentic AI systems can behave differently across runs: the same request may produce a different plan, a different tool call, a different code edit, or a different final answer. Such variability arises from several layers that are often conflated. At the core of many current agents is a foundation model, a large pretrained model adaptable to many downstream tasks, embedded in an orchestration loop that plans, calls tools, observes results, and updates state. One explicit intrinsic source of variability in such systems is token generation: the model computes scores over possible next tokens, the scores are converted into probabilities, and a decoder may sample tokens using a pseudo-random number generator. A small sampled token difference can then cascade downstream into a different tool call, code path, search query, or agent state. Other sources of variability are extrinsic to token sampling, including changing environments, live data, serving infrastructure, batch effects, and numerical details. By separating these layers, this tutorial clarifies what it means to call agentic AI systems stochastic, when such variability can be reproduced under matched conditions, and why deterministic execution need not imply identical behavior in deployed settings.

Muhammad Zia Hydari, Raja Iqbal

University of Pittsburgh · Ejento AI Inc.

Language Model Decoding LLM Agents LLM Agent Reliability

Jun 7, 2026cs.CL

Inside the LLM Word Factory

Transformer language models process input provided as subword fragments, but natural language semantics usually rely on word-level concepts. Detokenization is the process where models reconcile these two facts, aggregating subwords into word-level representations through their computation. Prior work has found that this takes place mostly in early-to-middle layers, but so far the exact mechanics of the process have not been pinned down. We venture deep into detokenization using activation patching in controlled paired experiments that isolate the contribution of different model components, localizing English detokenization in Llama2-7B to a two-stage process at Layer 1. Attention transmits a token-specific signal from nonfinal subwords, using sequential relays if necessary, while the MLP composes it with the local embedding. This two-stage structure generalizes to twelve models from eight families, but the depth over which it takes place depends on the flavor of positional encoding: RoPE-based models detokenize over 1 to 5 layers, while learned-absolute models take 5 to 10. Finally, we provide a probe for determining the success of the detokenization process based on early-layer activations alone, performing at 0.94-0.97 AUROC depending on the amount of context.

Benzi Busigin, Yuval Pinter

Stein Faculty of Computer and Information Science Ben-Gurion University of the Negev Beer Sheva, Israel

Language Model Decoding LLM Interpretability Subword Tokenization

Jun 5, 2026cs.CV

Don't Pause: Streaming Video-Language Synchrony for Online Video Understanding

Online Video Large Language Models (Video-LLMs) have advanced toward seamless human-AI interaction through frame-by-frame processing and proactive responding. However, a critical challenge remains in streaming scenarios: existing models typically pause video perception while generating responses, breaking real-time video-language synchrony and causing stutters. To address this, we introduce a novel paradigm for online video understanding: Streaming Video-Language Synchrony (SVLS), and present LyraV, a live streaming assistant built upon a hierarchical control framework with two core innovations. First, the Frame-Driven Transition Controller (FDTC), a training-free verification-based finite-state machine, makes high-level semantic decisions on when to continue speaking, start a new response, or stay silent. Second, the Streaming Token Pacer (SToP), a plug-and-play lightweight predictive module, dynamically adapts the language generation rate to match the pace of the visual content. Concretely, LyraV performs \emph{per-frame incremental, sub-budget decoding}: within each frame interval it emits only a small chunk of tokens that fits the real-time budget, so perception is never blocked for a full sentence. Together, these components enable LyraV to seamlessly interleave incoming video frames with generated word tokens, achieving a fine-grained synchrony. Extensive experiments conducted on five online and three offline benchmarks demonstrate that LyraV preserves the backbone's general understanding ability while substantially improving streaming synchrony and narrative fluency, delivering a 98.29% synchrony with video playback and a real-time processing speed of 3.89 FPS. Interestingly, we observe an empirical capability in LyraV: dynamic reasoning over streaming tokens, enabling continuous interpretation and "thinking" alongside visual input.

Zhenyu Yang, Kairui Zhang, Shengsheng Qian +2

Streaming Video Understanding Language Model Decoding VLM Inference