NeurIPS 2026 论文列表

Detecting LLM reasoning failures at inference time without ground-truth labels motivates a family of confidence baselines, including self-consistency, semantic entropy, and $P(\text{True})$, built on within-question sampling and self-evaluation. Operad theory, the formalism for systems built by iterated substitution, suggests a complementary diagnostic: a model's direct answer to a compositional query should agree with the answer it produces by composing a stated decomposition of the same query. We instantiate this idea as *operadic consistency* (OC), a per-question signal. Across twelve instruction-tuned LLMs (4B--671B parameters, open-weights and closed-source) on four multi-hop QA datasets, OC is strongly correlated with accuracy on every dataset (Pearson $r \in [{+}0.86, {+}0.94]$, all $p \leq 0.0004$), making it a substantially stronger population-level accuracy proxy at roughly one-third the inference cost of the strongest sample-based baseline (whose own cross-model rates never exceed $r{=}{+}0.80$ on any dataset). At the per-question level, OC contributes information beyond self-consistency, semantic entropy, $P(\text{True})$, and constructed decomposition-aware baselines on every dataset (cluster-robust $p \leq 10^{-19}$). We translate the regression coefficients into deployment-ready selective-prediction lifts ($\Delta_{\text{AUARC}}$ up to $+0.088$, $\Delta_{\text{AUROC}}$ up to $+0.167$ over a tuned $\text{SC}_{K=3}$ baseline). On five frontier thinking models, where the decomposition is extracted from the model's own chain of thought, the same equal-cost comparison gives positive selective-prediction lift on all $16$ (dataset, budget, metric) cells tested.

Bayesian model calibration is central to digital twins and computer experiments, as it aligns model outputs with field observations by estimating calibration parameters and correcting systematic model bias. Classical Bayesian calibration introduces latent parameters and a discrepancy function to capture bias, but suffers from parameter–discrepancy confounding and is typically formulated as an offline procedure under a stationary data-generating assumption. These limitations are restrictive in modern digital twin applications, where systems evolve over time and may exhibit both gradual drift and abrupt regime shifts. While data assimilation methods enable sequential updates, they generally do not explicitly model systematic bias and are not robust to abrupt changes. We propose Bayesian Recursive Projected Calibration (BRPC), an online Bayesian calibration framework for streaming data under simulator mismatch and nonstationarity. BRPC extends projected calibration to the online setting by separating a discrepancy-free particle update for calibration parameters from a conditional Gaussian process update for discrepancy, preserving identifiability while enabling bias-aware adaptation under gradual system evolution. To handle abrupt changes, we integrate BRPC with restart mechanisms that detect regime shifts and reset the calibration process, yielding a unified approach to mixed nonstationarity. We establish theoretical guarantees for both components, including tracking performance under gradual evolution and false-alarm and detection behavior for restart mechanisms. Empirical results on synthetic and plant-simulation benchmarks demonstrate that BRPC improves calibration accuracy under gradual changes, while the restart-augmented framework further enhances robustness and predictive performance under abrupt regime shifts compared to sliding-window Bayesian calibration and data assimilation baselines.

Omni-modal models are expected to reason over synchronized visual events and audio cues, yet answer accuracy alone can conceal failures in where an answer is grounded and which modality supports it. In audio-video question answering (AVQA), models may produce correct answers while relying on incorrect temporal spans, dominant-modality shortcuts, or unsupported cross-modal relations. We argue that this failure arises from treating reasoning as a completion-level prediction problem, where supervision and reward signals do not distinguish between support selection and reasoning over the support. To address this, we introduce OmniSE, a support-faithful optimization framework that makes question-conditioned audio-video support the central training object. Each QA instance is rewritten into an Audio-Visual Evidence Trace, a compact optimization unit that records temporal evidence, modality-specific cues, cross-modal relations, and cited reasoning steps. Given this support object, OmniSE first uses Support-Tree Rollout (STR) to explore alternative support hypotheses before sampling support-conditioned reasoning and answers. It then applies factorized reward assignment to score temporal localization, modality evidence, cross-modal binding, reasoning faithfulness, and answer correctness at their corresponding comparison levels. Finally, reward-gated self-consistency regularizes sibling continuations under reliable support, encouraging fixed trustworthy evidence to induce stable reasoning and answers. Experiments on multiple AVQA benchmarks show that OmniSE consistently outperforms strong open-source baselines at the same scale, demonstrating the importance of support-centric optimization for evidence-centric audio-video reasoning. Code and data would be publicly available after peer review.

We present NPUsper, a live transcription system that makes Whisper efficient on mobile NPUs by eliminating redundant computation. To avoid the heavy padding used by prior streaming systems, NPUsper detects hallucinated tokens online from temporal patterns in decoder cross-attention, allowing each inference round to process short audio inputs with minimal carryover. For efficient mobile-NPU execution, we propose controlled unrolling, which executes autoregressive decoding as K-step chunk graphs, removing unnecessary KV-cache computation and reducing graph-dispatch overhead. NPUsper achieves up to 4.84x lower per-word latency, up to 33.2x lower time-to-first-token (TTFT), and up to 11.36% lower average power consumption compared with baselines, while maintaining comparable transcription accuracy. The code is available at https://github.com/npusper/NPUsper.

Contemporary large language model (LLM) chat systems treat conversation history as an immutable sequence of turns that defines the model’s working context. However, user intent in real interactions is not static: it evolves through correction, refinement, and shifting constraints. This mismatch between dynamic intent and static transcripts can result in context pollution, where outdated or irrelevant information persists and continues to influence subsequent responses. We introduce mutable transcripts, a new interaction paradigm that enables users to revise prior turns through natural language edit requests, allowing the conversation history itself to be updated rather than appended. This reframes the transcript from a passive record into an editable representation of conversational state. We present a working prototype that integrates transcript-level revision into a standard chat interface and evaluate its feasibility through a controlled user study (n=17) and an illustrative transcript analysis of representative interaction scenarios. Participants significantly preferred mutable transcripts over standard chat across measures of clarity, confidence, and ease of use, with reduced intent to restart conversations. Transcript analysis of representative user study conversations shows that mutable transcripts can reduce conversation length and eliminate obsolete retained context. These findings provide initial evidence that user-driven revision of conversational history can improve interaction quality and help maintain a more current representation of user intent.

End-to-end full-duplex speech models have brought open-source machine conversation close to human fluency, yet existing systems fall short of real-world deployment in two entangled respects: long-context conversational robustness and multi-party interaction capability. Realistic settings, including meetings, group lessons, family dinners, and social-robot reception, are inherently long-horizon and multi-party at the same time, requiring a single model to perceive, attribute, contextualize, and respond across multiple speakers over extended durations. Progress along these axes is bottlenecked by both data and evaluation. On the data side, open multi-party conversational speech corpora total only a few hundred hours and are not designed for codec-frame-level full-duplex modeling. On the evaluation side, existing long-audio benchmarks focus on passive listening, while existing speech-to-speech benchmarks remain dyadic and short. In this work, we extend the Moshi paradigm along long-horizon and multi-party axes simultaneously, in both English and Chinese, with three contributions. First, we release an open data engine and a 57.6 k-hour corpus for long, multi-party, bilingual full-duplex dialogue. The engine produces parallel-stream audio with controllable length, participant count, conversational dynamics (turn-taking, overlap, backchannels, interruption, addressee shifts, long-range co-reference), and language (English, Chinese, and intra-sentential code-switching), exceeding all prior open multi-party conversational speech corpora by more than an order of magnitude. Second, we introduce MultiTalkBench, the first benchmark to jointly evaluate long-form full-duplex dialogue with an average duration of 32.6 minutes, multi-party, and bilingual full-duplex dialogue, with explicit probes for long-range entity tracking, topic coherence, and addressee selection. Third, we train a bilingual Moshi-style full-duplex model on the released corpus that sustains coherent multi-party English-Chinese conversation over extended durations, substantially outperforming open-source baselines including Moshi, MiniCPM-o-4.5, and Qwen3-Omni-30B-A3B-Instruct.

Large Language Models (LLMs) are trained to support an increasing number of languages, yet their predefined tokenizers remain a bottleneck for adapting models to lower-resource or distinct-script languages. Existing tokenizer transfer methods typically rely on semantic heuristics to initialize new embeddings, ignoring higher-layer model dynamics and limiting transfer quality. We propose Model-Aware Tokenizer Transfer (MATT), a method that incorporates model internals into the tokenizer transfer process. MATT introduces an Attention Influence Modeling (AIM) objective that distills inter-token communication patterns from a source model into a target model with a new tokenizer, providing an efficient warm-up before standard language modeling. Unlike approaches that focus solely on embedding similarity, MATT leverages attention behavior to guide embedding initialization and adaptation. Experiments across diverse linguistic settings show that MATT recovers a large fraction of the original model’s performance within a few GPU hours, outperforming heuristic baselines. These results demonstrate that incorporating model-level signals offers a practical and effective path toward robust tokenizer transfer in multilingual LLMs.

Speech large language models (SLMs) are typically built from text large language model (TLM) checkpoints, yet they still suffer from a substantial modality gap. Prior work has mainly attempted to reduce this gap from the output side by making speech generation more text-like, but the gap remains. We argue that the key remaining bottleneck lies on the input side. We propose TextPro-SLM, an SLM that makes spoken input more closely resemble that of a prosody-aware text LLM. TextPro-SLM combines WhisperPro, a unified speech encoder that produces synchronized text tokens and prosody embeddings, with an LLM backbone trained to preserve the semantic capabilities of the original TLM while learning paralinguistic understanding. Experiments show that TextPro-SLM achieves the lowest modality gap among leading SLMs at both 3B and 7B scales, while also delivering strong overall performance on paralinguistic understanding tasks. These gains are achieved with only roughly 1,000 hours of audio training data, suggesting that reducing the modality gap from the input side is both effective and data-efficient.

Test-time alignment aims to adapt large language models (LLMs) to runtime objectives without updating their parameters. This is especially useful when the reward signal is user-specific, time-varying, or available only through a black-box evaluator. We propose a minimally invasive pre-logit steering method that optimizes additive interventions to the hidden states of a frozen LLM in order to improve reward while preserving the base model's behavior. Rather than penalizing the Euclidean norm of the steering vector, we derive an effort penalty from the local KL geometry of the induced token distribution. Specifically, we show that the per-token KL divergence between the steered and reference policies admits a second-order expansion given by a Fisher-quadratic form in the steering vector. This regularizer has an analytic gradient computable through matrix-vector products with the frozen language-model head, making it as inexpensive as a standard quadratic penalty while retaining a principled KL interpretation. We further decompose the sequence-level KL gradient into an analytic Fisher component and a trajectory-dependent score-function component, and introduce a hierarchy of Fisher surrogates with provable first-order equivalence in the small-steering regime. The resulting algorithm provides a training-free, reward-driven test-time alignment procedure that improves reward while controlling deviation from the base policy.

Audio descriptions (AD) are essential for making visual media accessible to blind and low-vision (BLV) users. While Multimodal Large Language Models (MLLMs) offer a scalable solution for generating audio descriptions on-demand, their performance on ``inline'' audio descriptions -- which must fit within existing silences in a video -- remains under-explored, particularly for diverse user-generated content (UGC). We present a comprehensive investigation into MLLM-generated inline audio descriptions. First, we introduce a professionally annotated dataset of 36 hours of videos, each averaging $\sim$5 mins. and total 7k+ descriptions, providing a high-quality benchmark for this task. Second, we propose an evaluation framework based on audio description expert guidelines that provides fine-grained actionable metrics for model improvement. Crucially, our framework evaluates the end-to-end task: it renders the generated scripts into audio tracks to assess their fit within the original video's natural silences, minimizing disruptive overlaps. Our analysis reveals that while frontier MLLMs accurately describe visual events, significant gaps persist in audio-visual timing, quality, and narrative flow. Finally, we validate the framework and conduct studies with BLV participants and sighted raters, finding that while users narrowly prefer human-authored audio descriptions, they still find MLLM-generated audio descriptions highly beneficial for comprehension. We provide guidance on the current capabilities of MLLMs for audio descriptions and release our dataset and evaluation suite to the community.

Spiking neural networks (SNNs) with learnable membrane time constants can improve temporal processing by adapting neuronal integration timescales, but they also turn the decay factor $\beta$ into an optimized dynamical parameter that is fragile under hardware-induced parameter mismatch. We develop a perturbation-based account of how time-constant variation affects trained SNNs. In controlled software simulations, this fragility is strongly task-dependent: Spiking Heidelberg Digits (SHD) and Spiking Speech Commands (SSC) settings lose $7\text{--}20$ percentage points of accuracy under $20$% coefficient-of-variation (CV) perturbations in $\beta$, whereas DVS-Gesture and CIFAR-10 lose less than $3$ pp. We show that, under a surrogate-linearized first-order analysis, the failure mode is governed by the membrane sensitivity $\Psi$. This quantity admits a closed-form online recursion requiring only one additional state per layer and no extra temporal storage. The analysis identifies time-averaged sensitivity energy as the controllable term in deployment fragility, leading to $\textbf{MARS}$ ($\textbf{M}$embrane-$\textbf{A}$ware $\textbf{R}$obustness through $\textbf{S}ensitivity$ Regularization), a training objective derived from perturbation analysis rather than heuristic noise augmentation. We prove a conservative worst-case output-perturbation bound and introduce a typical-case diagnostic explaining why sensitivity remains predictive when worst-case constants are vacuous. At $20$% CV perturbation, MARS reduces the accuracy drop in the sensitive settings to at most $0.5$ pp while preserving nominal accuracy, with minimal effect on DVS-Gesture and CIFAR-10.

Most Large Language Model (LLM) agent memory systems rely on a small set of static, hand-designed operations for extracting memory. These fixed procedures hard-code human priors about what to store and how to revise memory, making them rigid under diverse interaction patterns and inefficient on long histories. To this end, we present \textbf{MemSkill}, which reframes these operations as learnable and evolvable memory skills, structured and reusable routines for extracting, consolidating, and pruning information from interaction traces. Inspired by the design philosophy of agent skills, MemSkill employs a \emph{controller} that learns to select a small set of relevant skills, paired with an LLM-based \emph{executor} that produces skill-guided memories. Beyond learning skill selection, MemSkill introduces a \emph{designer} that periodically reviews hard cases where selected skills yield incorrect or incomplete memories, and evolves the skill set by proposing refinements and new skills. Together, MemSkill forms a closed-loop procedure that improves both the skill-selection policy and the skill set itself. Experiments on LoCoMo, LongMemEval, HotpotQA, and ALFWorld demonstrate that MemSkill improves task performance over strong baselines and generalizes well across settings. Further analyses shed light on how skills evolve, offering insights toward more adaptive, self-evolving memory management for LLM agents.

Memory systems enable LLM agents to consolidate and retrieve relevant evidence from the factual knowledge accumulated through growing interaction histories for downstream reasoning. Existing approaches have explored diverse strategies for organizing and compressing these histories. However, balancing compression with retrieval effectiveness remains challenging: retaining too much content can cause relevant evidence to be obscured by redundant entries, while discarding too aggressively may remove content that later proves relevant. This amounts to a tradeoff between compressing redundancy and preserving enough structure to retrieve target evidence, as formalized by the information bottleneck. To this end, we propose **MemCoRe**, which organizes memory as a compression hierarchy where each level compresses redundancy further while retaining the structure needed for retrieval at that level. In this hierarchy, evidence is progressively compressed from detailed records through extracted keywords to topic groups. This enables retrieval to locate target evidence by searching across levels of the hierarchy. Comprehensive experiments demonstrate that MemCoRe outperforms existing state-of-the-art baselines.

Edge-deployed personal memory assistants must handle private interpersonal conversations on-device with open-weight models. Yet, existing memory benchmarks often under-test the combination of activity-dense interaction, ego-centric perspective, and coherent multi-session worlds. MemArena fills these gaps with a single-world conversational benchmark built with its MASim agent simulator, for $50$ agents over $15$ days ($10.3$M dialog-text tokens, $24.1$K text-only ego-observed tokens/agent/day). With the interaction history, it co-generates ground truth over six recall, reasoning, and trustworthiness evaluation dimensions. We evaluate five open-weight readers with Vanilla context, BM25-RAG, Oracle retrieval, Memobase, and MemSearch as memory backends. Three results stand out: (1) Memory-backend choice matters more for content accuracy: At Qwen3-0.6B, Memobase $\to$ MemSearch gains $+32.5/+19.2$ pp, exceeding MemSearch reader scaling ($+10.6/+6.8$ pp). (2) Permission-aware access fails universally: with Oracle leaking heavily and other backends too timid to disclose. (3) Search latency bites only at very small reader: on a Spark GB10 edge node, memory-search adds a moderate and fixed $87$/$7$/$48$ ms (BM25-RAG/Memobase/MemSearch) that composes a small part of TTFT for most reader-backend combinations. We release code, the MASim simulator, and the MemArena-L benchmark at https://anonymous.4open.science/r/MemArena-NeurIPS2026-anon-3E9D

Large language models (LLMs) increasingly improve their reasoning at test time via additional computation, yet most procedures treat each problem in isolation. When problems arrive sequentially, accumulating reusable experience across them can further improve performance. Existing memory-based methods either store whole-solution templates that generalize poorly to novel problems or use heuristic step-level selection that is not optimized for final-answer correctness. Learning selection policies requires large-scale training data and fixed action spaces. We propose MILES (Modular Instruction Memory with LEarnable Selection for self-improving LLM reasoning), a framework that dynamically expands step-wise memory and applies correctness-optimized memory composition under realistic test-time constraints. MILES maintains modular memory units consisting of asymmetric pairs of sub-goal embeddings and sub-instructions, each associated with a learnable selection head. This memory structure enables a coarse-to-fine retrieval mechanism: The coarse level enables memory expansion and collects supervision for training selection heads from confident samples, while the fine stage applies learned selection heads to rerank coarse-level candidates and guide reasoning for uncertain samples. MILES consistently matches or outperforms prior methods while achieving superior accuracy–efficiency tradeoffs. Extensive experiments demonstrate its effectiveness, robustness, and transferability.

Aligning large language models (LLMs) with diverse and multifaceted user preferences is a fundamental challenge in personalized AI systems. Existing multi-objective alignment methods either rely on costly training or require pre-trained reward models for each preference, making it difficult for them to adapt to evolving preferences. Prompt-based personalization offers a training-free alternative, but prompting alone often provides limited steerability, as LLMs may overemphasize or overlook certain preferences and fail to give users reliable control over the relative importance of different objectives when conflicts arise, leading to suboptimal alignment. In this paper, we introduce MATO, a training-free framework for Multi-objective personalized Alignment with Test-time Optimization. MATO formulates personalization as a test-time optimization problem that steers the relative importance of multiple objectives through controllable weights during decoding, without modifying model parameters or requiring external reward models. Specifically, a reward discovery module recovers preference rewards directly from the backbone LLM for diverse objectives specified in natural language, while a weight optimization module dynamically adjusts objective weights based on the user's initial preferences and the partially generated response to balance competing objectives during generation. The resulting rewards and weights jointly guide an online optimization procedure over the token distribution, enabling better alignment with the target objectives. Extensive experiments across multiple datasets and backbone LLMs show that MATO consistently outperforms strong baselines, achieving Pareto-improving multi-objective alignment and stronger steerability. These results highlight test-time optimization as a promising direction for scalable, controllable, and model-agnostic personalized alignment.

We introduce the Massive Audio Embedding Benchmark (MAEB), a large-scale benchmark covering 30 tasks across speech, music, environmental sounds, and cross-modal audio-text reasoning in 100+ languages. We evaluate 56 models and find that no single model dominates across all tasks: contrastive audio-text models excel at environmental sound classification (e.g., ESC50) but score near random on multilingual speech tasks (e.g., SIB-FLEURS), while speech-pretrained models show the opposite pattern. Clustering remains challenging for all models, with even the best-performing model achieving only modest results. We observe that models excelling on acoustic understanding often perform poorly on linguistic tasks, and vice versa. We also release MAEB+, the full 98-task collection from which MAEB is curated; to our knowledge, it is the largest publicly released collection of audio embedding evaluation tasks. MAEB preserves task diversity while reducing evaluation cost and integrates into the MTEB ecosystem for unified evaluation across text, image, and audio modalities. We release MAEB and all 98 tasks along with code and a leaderboard at https://anonymous.4open.science/r/mteb-2FEB/.

While AI weather models excel at short-to-medium range forecasts (up to 15 days), they frequently suffer from ill-defined "instabilities" when rolled out over longer horizons. This work addresses the lack of a formal taxonomy by categorizing these failures into three distinct regimes — blow-up, drift, and loss of seasonality — through year-long rollouts of nine state-of-the-art AI weather models. Our analysis reveals that stability hinges on the treatment of small spatio-temporal scales: unstable models amplify high-frequency energy, while stable models act as denoisers when noise is added to their inputs. Far from reducing these models to mere stochastic parrots, our findings highlight that stable models generate unique weather trajectories, conditioned on the initial state. We verify our findings through ablation studies on architectural design choices, conducted using state-of-the-art Vision Transformer (ViT) AI weather model architectures.

Long-horizon agency requires an AI system to preserve task state across hours or days of work, many tool calls, and repeated context resets. A common response is to scale a single resource: the model’s context window, together with inference-time reasoning. We argue that this response conflates long horizon with long context. Durable long-horizon state should not live only in the prompt; it should be externalized into the harness, the non-model infrastructure that manages tools, files, plans, checkpoints, sub-agents, permissions, and recovery. Publicly documented long-running agent systems already rely on such structures, including progress files, plan files, git history, rollback checkpoints, scratchpads, sub-agent decomposition, and browser-state checkpoints. We support this position with three lines of argument. First, memory-hierarchy history shows that pressure to scale one working resource repeatedly gave way to externally managed hierarchical state. Second, we identify five structural mismatches between long-context scaling and long-horizon agency: degraded effective context, lossy compaction, a single trust region, no transactional semantics, and no working-set decomposition. Third, we map long-horizon requirements to what context scaling provides and what harness-managed state must supply. We conclude by proposing eight harness state-management patterns and arguing that harness state should become a first-class research object for long-horizon agency.