Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but how these gains relate to inference-time decoding and search remains unclear. Does RL create reasoning the base model lacks, or shift the rollout distribution toward trajectories it can already reach but rarely samples? We study this behaviorally with a Unified Decoding Framework (UDF), which expresses token-level sampling, beam-like search, tree search, and sequence-level resampling as executable policies over a shared budgeted operating space, scored post hoc with pass@$k$, self-consistency, best-of-$N$, and first-finish success. Using paired Base/RL checkpoints from SimpleRL-Zoo, we ask whether an RL default-policy curve can be approximated by a structured path of Base operating points. On Math500, AIME, GPQA, and IFEval, the pass@$k$ recovery path follows a Budgeted Operating-Point Transition Rule (BOPTR), $N_{\mathrm{Base}} \approx αN_{\mathrm{RL}}^β$, with benchmark-conditioned exponents. On Qwen2.5-7B, BOPTR gives the lowest transfer error among the non-oracle rules we test, 3.41 pp (95% CI [2.32, 5.53]); a three-seed replication gives 3.07 $\pm$ 0.39 pp. The rule extends to ten models across four families (3.28 to 4.87 pp on checkpoints added after fitting), to four benchmarks it was never fitted on (5.03 pp vs. 4.44 pp in fit), and holds without an RL checkpoint for the target model (4.19 pp) or without RL supervision of any kind (5.08 pp). These results support a qualified internalized-search reading: under the recipe we test, much of the measured RL gain corresponds to a change in sampling efficiency toward operating points the base model can already reach under search. We treat the scaling patterns as descriptive of this recipe and cohort, report where they break down, and use UDF and BOPTR as behavioral diagnostics rather than evidence of parameter-level equivalence.
As one of the most critical challenges in large language models, contextual faithfulness directly determines their reliability in knowledge-intensive applications. This task is particularly challenging as it requires balancing factual consistency with generation efficiency. Contrastive decoding methods require dual forward passes (with and without context) to compare model outputs, doubling inference computational overhead, while post-training alignment demands extensive reinforcement learning with substantial computational overhead. To address this challenge, we present \textbf{SFAD}, a speculative decoding framework that enhances contextual faithfulness without inference degradation. We first construct \textbf{ConFide}, a preference dataset with fine-grained atomic perturbations, to train a context-faithful draft model via Direct Preference Optimization. During inference, Epistemic Friction detects potential hallucinations by quantifying distributional tension weighted by specialist certainty. When friction exceeds the threshold, Asymmetric Logit Steering refines the target distribution through residual-based logit injection; otherwise, standard speculation proceeds. Extensive experiments demonstrate that SFAD substantially improves faithfulness while achieving $2.48\times$ speedup, offering a practical solution for efficient LLMs.
Philipp Emanuel Weidmann, Allen Roush, Judah Goldfeder +2cs.CL cs.AI cs.LG
Large Language Models generate text autoregressively, but open-ended generation is prone to verbatim looping, in which models repeat spans already present in context. Standard defenses such as repetition, presence, and frequency penalties and n-gram blocking act on token recurrence rather than the sequential structure of a loop, and often suppress looping only at strengths that also degrade formatting or fluency. We propose Don't Repeat Yourself (DRY), a sampling-time logit adjustment that penalizes a candidate token only when generating it would extend the current suffix into an exact continuation of a span seen earlier in the context. Sequence breakers protect chat templates and formatting tokens. Across models from 1.5B to 120B parameters, nine prompt families, and a 600-pair human study, DRY reduces suffix-extension rate by 47% while improving lexical diversity. An intervention-matched placebo produces no comparable reduction, identifying suffix matching as the operative mechanism. On AWQ-quantized 70B and 120B models, DRY reduces loop rate by roughly half while preserving MT-Bench, MMLU, and GSM8K performance, whereas standard alternatives lose measurable ground. DRY has been adopted by popular open-source LLM inference frameworks including llama.cpp, ExLlamaV2, and text-generation-webui, highlighting its practical impact on text generation.
Philipp Emanuel Weidmann, Allen Roush, Judah Goldfeder +2cs.CL cs.AI cs.LG
Standard decoding rules for autoregressive language models promote diversity by rescaling the full next-token distribution or truncating its low-probability tail. These strategies overlook a common regime of open-ended generation in which several continuations are plausible but too much probability mass remains concentrated on the most generic choice. We introduce XTC (Exclude Top Choices), a lightweight head-aware decoding operator that targets this regime directly. XTC identifies tokens whose probabilities exceed an absolute plausibility threshold $τ$: when at least two qualify, it removes the dominant eligible choices with probability $ρ$ and retains only the weakest plausible alternative before renormalization. Across 60 experiments on Gemma 3 27B Q4, Gemma 3 12B Q6, and DeepSeek R1 14B Q6, with scaling validation on Llama 3.3 70B Q4, XTC improves the diversity-repetition Pareto frontier. On creative generation, Distinct-2 increases by 11--15% and repeat trigrams decrease by 27--47% across the four models. Combined with temperature scaling, gains reach 38% in Distinct-2 and 71% in repeat-trigram reduction over baseline. A blinded Amazon Mechanical Turk study with 150 Master raters yields a 62.3% creativity preference for XTC ($p<10^{-4}$) without reduced fluency, while a GPT-4o control judge reproduces the Anthropic-judge direction on every measure. On IFEval with Llama 3.3 70B Q4, XTC preserves prompt-level strict accuracy within 1.7 percentage points of baseline while recovering most of the diversity gain; a temperature setting matched on Distinct-2 reduces IFEval by 8.8 points. The effect is additive with temperature and repetition penalties, robust across quantization levels and model families, and consistent across twelve prompt genres. XTC has been adopted by llama.cpp, ExLlamaV2, and text-generation-webui.
Diffusion multimodal large language models (dMLLMs) frequently produce long-form outputs marred by semantic drift and repetition, with quality generally degrading as output length increases. We identify two structural deficiencies in existing decoding methods as primary drivers of these failures: confidence-based scoring ignores decoded-neighbor support, and block partitioning prevents access to high-readiness semantic anchors, together causing tokens to be committed before their local context is sufficiently established. We propose \ours{} (\textbf{C}ontext-\textbf{A}ware \textbf{C}luster \textbf{D}ecoding), a training-free decoding method that scores each masked position by a multiplicative composite of softmax confidence and neighbor proximity, promoting contextually ready tokens above isolated candidates while suppressing low-confidence positional noise, operating block-free to keep high-readiness anchors globally accessible. \ours{} further applies architecture-aware calibration to handle confidence heterogeneity induced by diverse visual integration strategies. Experiments on three dMLLMs across four benchmarks demonstrate consistent quality gains and hallucination reduction over Original, with larger gains in several longer generation settings, highlighting the importance of neighbor support and visual integration strategy for future dMLLM decoding method design. Our code is openly available at https://github.com/zhaoyk-sysu/CACD-dMLLM.
Mixture-of-Experts (MoE) models have been widely adopted in real-time interactive applications such as coding assistants, real-time audio-video interaction systems. To meet the extremely low response latency requirements of these scenarios, practitioners commonly employ small-batch decoding, under which MoE inference becomes memory-bound and is severely bottlenecked by expert weight loading. However, this bottleneck has received limited attention, and existing solutions such as post-training weight compression or fine-grained expert design during pre-training either degrade model accuracy or introduce additional computation and communication overhead. To tackle this issue, we propose DeaMoE, a decoding-efficient MoE architecture, in which the experts are grouped into several departments, and the experts belonging to the same department share most parameters since they come from the same professional field, and additionally each expert contains a few private parameters to reflect its uniqueness. Moreover, we design customized two-stage routing strategy for DeaMoE to avoid redundant loading, under which DeaMoE greatly improves the efficiency during LLM decoding. Compared with vanilla MoE, DeaMoE reduces per-step loaded weights by up to 50.9% and achieves up to 1.33 end-to-end TPOT speedup for the pre-trained 7B model on A40, and up to 2.00x and 1.97x peak speedup for DeepSeek-V3 on A40 and H100 in microbenchmarks.
Large Vision-Language Models (LVLMs) have achieved remarkable performance across a wide range of tasks; however, they often inherit social biases from their training data, resulting in biased behavior when processing portraits from different social groups. Existing debiasing approaches typically compare token probabilities between the original and biased generations during decoding, but they are fundamentally limited by their reliance on a single, stereotyped viewpoint and fail to account for the diversity of social perspectives. Inspired by the social science principle that diversity fosters fairness, we propose Counterfactual Ensemble Decoding (CED), a novel framework that constructs multi-group counterfactual perspectives within the visual representation space and integrates them during decoding to promote equitable model behavior. CED first performs counterfactual steering in the visual space by identifying semantic directions associated with each social group and generating counterfactual representations along these directions, thereby offering diverse perspectives that disrupt stereotypical narratives. During decoding, CED locates the decoder layer exhibiting the greatest divergence among these perspectives and ensembles their token distributions using uncertainty-aware weights, prioritizing high-confidence tokens from different groups to yield a more balanced probability distribution that guides fairer generation. Extensive experiments on three social bias evaluation benchmarks demonstrate that \tool achieves substantial improvements over leading baselines, reducing bias by up to 47.97% across scenarios involving occupations, descriptors, and persona traits. Moreover, CED also preserves the core capabilities of the original model with minimal degradation.
Classifier-free guidance (CFG) is usually kept on throughout masked diffusion language model decoding, although its benefit varies across prompts and over time. We study when CFG is actually needed by comparing, from any partial output, the probability of eventual constraint satisfaction under continued CFG and under base-only continuation. Their difference defines the remaining value of guidance. Guidance dependence is highly prompt-specific. Many prompts already succeed without CFG, while for others it provides no measurable benefit or can be harmful. For prompts that do benefit, the gain is often concentrated early. We define the commitment horizon $\astar$ as the earliest point from which switching all remaining decoding to the base model reduces final success by no more than a chosen tolerance. Under the base model, the corresponding success probability, or committor, is a martingale. To first order, CFG's per-step effect is governed by the covariance between the guidance logit direction and the successor committor. This gives a local account of when guidance can help, but it does not by itself locate the horizon. Among prompts with an observed preterminal horizon, $\astar$ is usually early and varies more within constraint families than between them. Freezing each prompt at its own cross-fitted horizon is noninferior to full CFG on all 13 subtasks at the prespecified margin, even while many tokens remain masked. This separates commitment from realization. The boundary also identifies a later region in which higher parallelism adds only a small cost in constraint success, although fluency still degrades with parallel width. For failed trajectories, reopening committed positions improves recovery in both failure modes.
Multimodal Large Reasoning Models (MLRMs) have achieved strong performance on tasks requiring visual understanding and multi-step inference. However, as reasoning trajectories grow, models may become less effective at using information established earlier in the context, increasing the risk of reasoning errors. Existing approaches primarily address this problem by sustaining visual grounding throughout reasoning. However, reasoning also transforms visual observations into task-specific relations, constraints, and intermediate conclusions whose influence may weaken over long trajectories. Our attribution analysis suggests that correctness is not consistently separated by image attribution alone, but is more closely associated with whether trajectories retain and integrate such reasoning-derived information across stages. Motivated by this, we introduce TRAM (TRajectory-derived Auxiliary Memory), a training-free method that augments standard decoding with an auxiliary memory pathway derived from the model's own reasoning trajectory. TRAM consolidates completed reasoning into a compact latent memory, updates it online through fast and slow recurrent streams, and feeds it back into selected decoder layers through a lightweight residual pathway. Experiments across four MLRM variants on eight benchmarks show that TRAM improves performance over vanilla decoding on mathematical, scientific, and general visual reasoning tasks without additional training.
Diffusion language models (DLMs) update many tokens in parallel, yet practical decoders often use a fixed denoising horizon. Many predictions stabilize early, but blockwise decoding continues until all positions are resolved, causing repeated dense forward passes. Existing accelerators often rely on learned filters, modified scores, dependency models, or cache-specific mechanisms. We ask whether native trajectory signals can identify residual positions likely to match the deterministic dense endpoint. We propose CORA-Diff, a training-free method that preserves the original transfer rule and applies confidence-and-persistence gating only to positions that rule leaves unresolved. Accepted tokens remain visible as context, and the block terminates once all positions are resolved. This requires no backbone change, learned acceptance model, or logit modification. Our theory explains why high-confidence, persistent predictions are more likely to match the fixed-horizon dense endpoint, and paired post-intervention trajectories provide direct empirical support. We select one operating point on a separate GSM8K calibration subset and freeze it for all evaluations. Under a matched Learn2PD-style LLaDA protocol, CORA-Diff has the lowest measured runtime in all eight task-length settings. Task scores match or exceed dense decoding in five settings, and the largest observed drop is 1.22 points. Its incremental speedups over EOS-aware dense decoding are 2.70x and 3.32x on GSM8K and HumanEval. It also reaches 13.14x under the fixed-horizon 1024/1024 mechanism-isolation protocol and transfers to Dream without retuning at 3.18x-3.53x. These results show that native confidence and persistence enable reliable residual acceptance, reducing repeated denoising computation while preserving task quality.
Prediction uncertainty is a widely adopted metric for quantifying model confidence, with downstream applications spanning model explanation, data selection, and prediction rollback. Despite its demonstrated utility, the potential of uncertainty quantification to enhance code generation in large language models (LLMs) remains largely underexplored, raising a critical question: to what extent can uncertainty serve as an effective signal for improving LLM-based code generation? To answer this question, we study uncertainty-aware rollback decoding, an inference-time strategy that uses uncertainty signals to identify unreliable generation regions and roll back to earlier valid prefixes without retraining the model. We evaluate this framework on seven code LLMs, five code generation benchmarks, and eight token-level uncertainty signals under a unified decoding setup. Our results show that the complete rollback framework improves over equal-budget restart across the evaluated benchmarks and model settings, with gains of up to 0.26 in pass@1 and 0.35 in AvgTestPassRate on functional code generation benchmarks, and an absolute improvement of up to 6.4\% in Patch-Aligned Safe Rate on Dsec-Python. Among the evaluated signals, information-theoretic measures such as token entropy and negative log-likelihood show the most favorable overall trend, frequently achieving the best or near-best results on standard benchmarks. A component-controlled ablation further shows that feedback-guided rollback provides the main improvement, while uncertainty localization provides an additional gain when checking, budget, rollback, and branch decay are held fixed.
Large vision-language models (LVLMs) often hallucinate objects that are absent from an image. Despite recent progress, existing mitigation methods still lack reliable object-level grounding diagnostics and therefore tend to apply coarse-grained interventions, which can impair visual understanding, shorten responses, and reduce coverage of genuinely grounded objects. The key challenge is thus to detect, during generation, whether each emerging object mention is supported by reliable visual evidence, so that hallucination can be mitigated selectively. Yet output confidence reflects next-token plausibility rather than visual support, allowing language priors to make absent objects appear certain. We show that the missing diagnostic evidence is encoded in an Intrinsic Grounding Signature (IGS), a distributed signed attention pattern that remains informative for such confident hallucinations. Based on IGS, we propose Verifier-Guided Decoding (VGD), a decoding framework in which a lightweight verifier examines each emerging object mention, rolls back the KV cache when the mention is identified as high risk, suppresses the object and its synonyms, and regenerates the affected continuation. Because VGD intervenes only on object mentions identified as high risk, it reduces object hallucination while preserving the model's original visual understanding and grounded object coverage. Experiments on CHAIR and AMBER-G show that VGD achieves state-of-the-art object hallucination reduction: at @rec90, it cuts AMBER-G CHAIR by 43.6\% while retaining 99.6\% of grounded-object coverage, and reduces CHAIR-MSCOCO CHAIR$_i$/CHAIR$_s$ by 37.0\%/30.4\% without shortening captions.
Large Language Models (LLMs) are increasingly controlled through system prompts that specify roles, styles, 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 the learned reward signal and an optional contrastive signal capturing 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.
Recent work shows that fine-tuning language models on even a small amount of poisoned data can install targeted misbehavior, and ostensibly benign data can transmit hidden preferences that generalize broadly. Standard defenses, such as data filtering, mixing in harmless data, and regularization, attenuate these effects but do not eliminate them. We instead pursue robustness through redundancy: collecting multiple datasets from different sources and only learning what is common between them. Thus, if only a subset of sources are malicious, the misbehavior will be blocked. In order to implement this defense strategy, we fine-tune a separate reference model on each source's dataset and aggregate their next-token distributions at decoding time. We introduce two consensus decoders: a token-wise minimum, which caps each token at the lowest probability any source assigns, and a base-relative variant, which reverts to the base probability on any token the sources move in opposing directions. We further relax exact agreement to tolerate partial support across sources and different surface expressions of the same intention. Across controlled poisoning tasks, subliminal learning, and emergent misalignment, consensus decoding suppresses source-specific misbehavior while preserving shared desirable behavior, including cases where union training and weight averaging retain the unwanted behavior.
Brian K Chen, Chong Wu, Kenji Kawaguchics.CL cs.AI
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.
In this work, we address diacritic restoration for Arabic speech transcripts. Most speech data are undiacritized, limiting the ability of modeling fine-grained phonological distinctions. The speech modality has recently been explored as a way to complement text-based diacritic restoration efforts. We propose an efficient non-autoregressive approach for speech-to-text diacritization based on Connectionist Temporal Classification (CTC). Our method incorporates hard constraints during decoding by constructing a character-level diacritization lattice from an undiacritized transcript and restricting hypotheses to valid diacritized realizations. We evaluate on Classical Arabic and Modern Standard Arabic test sets (namely, ArVoice and ClArTTS) against a more computationally-complex multi-modal diacritic restoration baseline, and show statistically significant reductions in diacritic error rates in both, demonstrating that the proposed approach offers both performance and efficiency gains.
Cross-domain sequential recommendation (CDSR) aims to model users' dynamic interest transitions and sequential patterns across multiple domains. Recently, generative recommendation (GR) has emerged. It first learns semantic identifiers (SIDs) from item semantics and formulates recommendation as autoregressive generation. However, existing methods face two critical issues: (1) they ignore collaborative correlations across domains during tokenization, and (2) they adopt inefficient decoding strategies, such as beam search, during generation, which hinders real-time deployment. To address these limitations, we propose GenCDSR, an effective and efficient generative framework for CDSR. Specifically, we design a cross-domain hybrid tokenization mechanism with a multi-tower architecture to jointly capture cross-domain commonalities and domain-specific distinctions through hierarchical shared-specific and fine-grained codebooks. Furthermore, we develop a cross-domain serial-parallel decoding strategy that leverages the hierarchical SID structure to partially parallelize generation, significantly reducing inference latency while preserving generation consistency. Experiments on three public datasets show that GenCDSR achieves an average accuracy improvement of 1.5 percent and an average inference latency reduction of 85.1 percent compared with state-of-the-art baselines. The implementation code and datasets are available online: https://github.com/Applied-Machine-Learning-Lab/RecSys2026_GenCDSR.
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.
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.
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.
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.
Diffusion language models (dLLMs) commit multiple tokens per denoising step by decoding each selected position independently from a shared context. When these positions are dependent, this factorization introduces an error captured by conditional total correlation, which confidence-based selection cannot infer from marginal probabilities alone. We propose CoCommit, a marker-gated coordination pass that delays commitment. After the usual bundle selection, a learned marker identifies the commit set, and the backbone's last n layers are re-applied to coordinate the marked positions before greedy argmax writes the tokens. This approximates joint-mode decoding while reusing existing weights, requiring only one partial forward pass and no auxiliary model. On LLaDA 2.1 with LoRA adapters and greedy inference, joint commitment improves five of the seven evaluated benchmarks over the released factorized decoder. The largest gains occur on code and reasoning tasks, while the remaining tasks are near parity.
Diffusion language models (DLLMs) generate text by iteratively denoising masked positions, exposing a trajectory of predictive distributions rather than a single instantaneous belief. Most existing decoders ignore this trajectory and commit tokens from the current snapshot alone, conflating confidence with commitment readiness: a transient top-1 peak under incomplete context can be locked in, while candidates with consistent cross-step support are delayed. We propose Trajectory-Aware Commit Gating (TACG), a training-free gate-level decoder that anchors token identities to the base posterior and uses trajectory-aware signals only to decide whether the current proposal is ready to commit. TACG combines Temporal Implicit Logits Guidance (TILG), which keeps an exponential moving average of past logits as a self-reference and contrasts the current logits against this reference in natural-parameter space, with a History Gate (HG) that enforces short-term proposal persistence before commitment. Together with a capped extra-promotion budget, these components yield a stability-constrained commit rule without auxiliary networks or extra forward passes. We evaluate TACG on LLaDA, Dream, and LLaDA2-Mini across code (HumanEval, MBPP) and math (GSM8K, MATH500) benchmarks; it typically improves or preserves accuracy while reducing denoising steps and increasing tokens per forward (TPF). The code is publicly available at https://github.com/Clarence-CV/TACG-DLLM.
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/
Large output embedding matrices create a significant memory bandwidth bottleneck during autoregressive decoding, especially for compact LLMs with large multilingual vocabularies. We reformulate the output projection followed by top-k token selection as a maximum inner product search over token embeddings and replace the dense vocabulary projection with an HNSW-based vector index. The resulting output head retrieves only a small candidate set of high-scoring tokens and can be integrated into existing decoding pipelines by scattering retrieved logits into a sparse full-vocabulary tensor. On CPU inference with Gemma 3, Llama 3.2, and Qwen 3 models, our method substantially accelerates the output projection and improves end-to-end batch-size-one decoding throughput by up to 82% for Gemma 3 270M, while preserving generation quality under AlpacaEval evaluation. These results suggest approximate retrieval is a practical alternative to dense output projections in latency-sensitive small-batch decoding.
Raymond Li, Md Tawkat Islam Khondaker, Amirhossein Abaskohi +3cs.CL
Retrieval-augmented generation (RAG) increasingly requires models to answer questions from multiple retrieved documents, where only some sources are relevant and the retrieved bundle may contain stale, noisy, or conflicting evidence. Existing contrastive decoding methods primarily focus on resolving conflicts between the model's internal memory and the retrieved context. In contrast, we study the complementary problem of intra-context conflict in multi-document RAG. To evaluate this setting, we introduce DRQA, a factual-conflict question answering benchmark derived from enterprise deep-research scenarios, where answers are grounded in synthetic enterprise-specific facts that are designed not to be recoverable from the model's internal memory. We further propose Dual-Confidence Contrastive Decoding (DCCD), a training-free decoding method that combines document-level confidence, which estimates whether a document appears sufficient for answering the question, with token-level confidence, which estimates whether that document supports a confident next-token prediction. DCCD selects positive and negative document-conditioned streams using these dual-confidence signals and scales a document-level contrast by their confidence margin. Across DRQA and standard multi-document QA benchmarks, DCCD achieves the best average performance among full-context and contrastive decoding baselines, with the largest gains on DRQA. These results highlight the importance of source-aware, confidence-gated decoding when retrieved evidence is internally conflicting.
Test-time scaling improves language-model reasoning, but existing approaches often face a difficult trade-off: long chain-of-thought sampling remains single-threaded, while sentence- or solution-level search can be computationally expensive and hard to train end-to-end. We introduce Local Branch Routing (LBR), a token-level test-time scaling framework that expands a small local lookahead tree, forwards all sampled branches through the language model, and uses a lightweight router to select the depth-1 subtree to commit. By routing over the hidden states of candidate local futures, LBR allows each token decision to use evidence beyond the root next-token distribution while avoiding full solution-level search. The resulting prune-shift-grow decoding process preserves discrete branch identities and defines a tractable tree-trajectory likelihood: newly grown nodes are counted when first sampled, and router decisions are assigned explicit probabilities. This enables end-to-end reinforcement learning with verifiable rewards, jointly optimizing the base model and router under the same likelihood-ratio principle as discrete-token RLVR. On synthetic hierarchical-planning tasks, LBR shows that post-candidate hidden states provide useful routing evidence. On mathematical reasoning benchmarks, LBR improves both Pass@1 and Pass@32 over discrete chain-of-thought, vanilla discrete-token RLVR, and RL-compatible soft-token branching baselines. These results suggest that lightweight local branching offers an efficient, trainable, and discrete form of language-model test-time scaling.
In sparse Mixture-of-Experts language models, does the same token id imply the same router state and the same experts producing it? Holding the emitted token id fixed at repeated anchors, we find it does not: the experts that produce it still separate task context, trajectory history, and reasoning-effort mode. This residual structure supports test-time control: near \emph{boundary} anchors (the final-response transition) and \emph{delimiter} anchors (which open the answer, e.g.\ \texttt{\textbackslash boxed\{} or code fences), routing neighborhoods already align with final-answer basins at a marker-only readout and strongest when the routing is read at the answer opening. We operationalize this as \textbf{RAD} (Routing Agreement Decoding), an answer-string-free multi-rollout selector: it locates a fixed anchor, represents each rollout by its anchor-window MoE routing states, and returns the densest Weighted-Jaccard $K$-NN route-basin center, without parsing, normalizing, executing, or voting over answer strings. Across 10 sparse-MoE configurations (gpt-oss, Qwen3-MoE) and 6 datasets spanning math, GPQA, and code, RAD is on par with Majority where string voting is well-posed, with small positive paired deltas (RAD $73.9$ / RAD+DC $74.2$ vs.\ Majority $73.6$). Like majority voting, RAD is not a verifier: a dense \emph{wrong} basin can still win. Its value is the interface: the same selector gives direct pass@1 on code, where exact-string voting is ill-defined, and the same routing-density principle, re-anchored to the agentic boundary, improves best-of-16 patch selection on SWE-bench Verified over random, where patches have no answer string to vote on.
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
Amr Mohamed, Guokan Shang, Michalis Vazirgianniscs.CL
Diffusion large language models (dLLMs) offer a promising alternative to autoregressive decoding by iteratively refining masked sequences, enabling parallel token updates and bidirectional conditioning. Their practical efficiency, however, is limited by sampling procedures that execute a fixed number of reverse denoising steps selected before decoding, spending computation on already-stable positions and sometimes committing unstable ones too early. We present \textsc{LESS}, a training-free, model-agnostic adaptive sampler that treats token commitment as an online stopping problem. \textsc{LESS} implements mutual-stability sampling through a joint stability rule that makes a masked position eligible for unmasking only when its top-1 prediction has high confidence, its top-1 token persists across recent reverse steps, and its predictive distribution is stable under top-$K$ inter-step Jensen--Shannon divergence. We evaluate \textsc{LESS} on Dream-7B, LLaDA-8B, and LLaDA-1.5-8B, covering full-sequence diffusion and semi-autoregressive blockwise sampling regimes, across seven benchmarks spanning general knowledge, math, and code. \textsc{LESS} improves average accuracy over strong training-free adaptive samplers while using $72.1\%$ fewer reverse steps than fixed-budget decoding. Since each reverse step requires a Transformer forward pass, these step-count reductions translate into fewer forward evaluations, lower measured wall-clock latency, and lower estimated inference compute.