NeurIPS 2026 Papers

Diffusion Large Language Models (dLLMs) have emerged as a compelling paradigm for language modeling, offering the unique potential for highly efficient inference via multi-token parallel decoding. However, unlocking this potential remains challenging. High-quality generation typically defaults to one-token-per-step static decoding, while existing parallel algorithms rely on heuristic confidence thresholds, introducing a fragile trade-off between decoding efficiency and generation quality. To overcome this, we introduce **Free** **D**raft-**a**nd-**Ve**rification (**FreeDave**), a novel training-free and model-free fast decoding algorithm. FreeDave capitalizes on the inherent properties of dLLMs by treating future token predictions as "free drafts" and rigorously self-verifying them in subsequent forward passes. Theoretically, FreeDave is guaranteed to use the fewest possible model forward passes to reproduce the exact sequence generated by static decoding, achieving algorithmically lossless acceleration. Extensive evaluations on math reasoning and code generation benchmarks demonstrate that FreeDave can boost the number of generated tokens per model forward (TPF) by up to 5.48$\times$ and tokens per second (TPS) by up to 3.94$\times$, while robustly maintaining generation quality and avoiding the task-specific performance degradation faced by heuristic methods.

Agentic forecasting is important for decision-making in dynamic environments, but it remains challenging because agents must reason from incomplete, time-limited evidence and produce calibrated probabilities before outcomes are resolved. Memory provides a natural mechanism for transferring experience from resolved forecasts to future prediction tasks. However, existing agent-memory methods are not tailored to forecasting, as they typically store past interactions, reflections, or factual associations without explicitly representing reusable predictive factors or calibration knowledge. We propose ForecastCompass(FoCo), an adaptive factor-based memory framework for agentic forecasting. FoCo organizes forecasting experience with a hierarchical forecasting-task taxonomy, enabling retrieval task-relevant forecasting knowledge. It maintains two complementary memory components: factor memory, which captures reusable predictive dimensions, and reasoning memory, which encodes probability updating, uncertainty handling, and calibration principles. Using retrospective analyses as learning signals, FoCo iteratively revises memory through a verbalized memory-revision procedure, enabling the agent to accumulate transferable forecasting knowledge over time. Experiments on Prophet Arena and FutureX with GPT-5-mini and Gemini-2.5-Flash show that FoCo improves both probabilistic accuracy and calibration.

We introduce FLUXtrapolation, a benchmark for extrapolating ecosystem fluxes under progressively harder distribution shifts. Ecosystem fluxes are central to understanding the carbon, water, and energy cycles, yet they can only be measured directly at sparsely located measurement towers. Producing global flux estimates therefore requires training models on observed sites using globally available covariates and predicting in unobserved regions, that is, upscaling. Flux upscaling is a challenging domain generalization problem that is affected by a shift in covariate distribution across climates, ecosystem types, and environmental conditions, as well as by conditional shift: important drivers remain unobserved at global scale. We provide a quantitative analysis of both these shifts in $P_X$ and $P_{Y\mid X}$. FLUXtrapolation is designed based on domain expertise on flux upscaling: it defines temporal, spatial, and temperature-based extrapolation scenarios and evaluates performance across held-out domains, temporal aggregations, and tail errors. In a pilot study, we find that baselines perform similarly under median hourly RMSE, but separate under the proposed tail-focused and multi-scale evaluation. FLUXtrapolation therefore poses a realistic and thus relevant challenge for machine learning methods under distribution shift; at the same time, progress on this benchmark would directly support the scientific goal of improving flux upscaling.

We introduce Sharper Transductive Local Complexity (STLC) as a new tool for analyzing the generalization performance of transductive learning methods, improving upon the current transductive bounds. Our work extends the classical local complexity-based analysis to the transductive setting, incorporating substantial and novel components beyond standard inductive and transductive analysis. Although Local Rademacher Complexity (LRC) has been used to obtain sharp inductive generalization bounds and local complexity-based transductive bounds, it has remained an open problem whether a localized Rademacher complexity framework can achieve exactly the same sharp bounds matching their inductive counterparts. STLC provides a confirmative answer to this question. STLC is constructed by first deriving a new and sharp concentration inequality for the supremum of empirical processes capturing the gap between test and training losses, or the test-train process, under uniform sampling without replacement. The proof establishes a Bernstein-type concentration inequality via a novel entropy-based approach built on the modified log-Sobolev inequality for the swap walk. A subsequent peeling strategy with a surrogate variance operator then yields excess risk bounds in the transductive setting that exactly match the classical LRC-based inductive bounds without the additional logarithmic gap in existing works. We further advance the current state-of-the-art in transductive learning through two applications: (1) for realizable transductive learning over binary-valued function classes with finite VC dimension $\dVC$ and $u \ge m \ge \dVC$, where $u$ and $m$ are the number of test features and training features, STLC gives a nearly optimal bound $\Theta(\dVC \log(me/\dVC)/m)$ nearly matching the minimax rate $\Theta(\dVC/m)$ up to $\log m$ and exactly matching the inductive bound, resolving a decade-old open question; and (2) STLC presents a sharper excess risk bound for transductive kernel learning compared to the prior local complexity–based results.

Diffusion-based large language models (dLLMs) rely on bidirectional attention, which prevents lossless KV caching and requires a full forward pass at every denoising step. Existing approximate KV caching methods reduce this cost by selectively updating cached states, but their decision overhead scales with context length or model depth. We propose EntropyCache, a training-free KV caching method that uses the maximum entropy of newly decoded token distributions as a constant-cost signal for deciding *when* to recompute. Our design is grounded in two empirical observations: (1) decoded token entropy correlates with KV cache drift, providing a cheap proxy for cache staleness, and (2) feature volatility of decoded tokens persists for multiple steps after unmasking, motivating recomputation of the $k$ most recently decoded tokens. The skip-or-recompute decision requires only $O(V)$ computation per step, independent of context length and model scale. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct show that EntropyCache achieves 15.2×–26.4× speedup on standard benchmarks and 22.4×–24.1× on chain-of-thought benchmarks against vanilla baselines, with competitive accuracy and decision overhead accounting for only 0.5% of inference time.

Speech large language models (SpeechLMs) can process spoken requests, yet still lag behind text-prompted large language models (LLMs) on instruction following and reasoning. We argue that this gap is not only an ASR or representation problem: the same semantic instruction can induce different response policies when spoken with different speakers, accents, noise conditions, prosody, or disfluencies. We propose Verifiable Invariant Reinforced Behavior Alignment (VIRBA), a reinforcement-learning framework for aligning SpeechLM behavior across acoustic realizations of the same semantic intent. VIRBA builds multi-view spoken instruction groups, scores sampled responses with semantic preference, rule-verifiable correctness, cross-acoustic invariance, and adaptive reasoning rewards, and optimizes the model with Cross-Acoustic Group Relative Policy Optimization (CA-GRPO). The resulting objective moves SpeechLM alignment beyond teacher imitation toward robust reasoning policies that remain stable across realistic spoken realizations. Experiments with recent SpeechLM baselines, disfluency robustness, spoken QA, audio reasoning, and speech-to-text translation show the largest gains on reasoning-heavy and acoustically perturbed spoken prompts.

Brain-to-speech decoding aims to restore communication by translating neural activity into speech-related representations. From a modeling perspective, this task can be formulated as a weakly supervised neural sequence decoding problem, where phoneme sequences must be recovered without frame-level annotations. Existing speech decoders commonly rely on recurrent sequence models to integrate temporal context, but their dense sequential computation can limit their suitability for resource-constrained portable or implantable neuroprosthetic systems. In this study, we propose a residual spiking neural network (SNN) for phoneme-level decoding of intracortical speech signals under constrained model capacity and weak sequence-level supervision. Our approach incorporates biologically inspired fixed connection-specific delays to integrate temporal context, providing each postsynaptic unit with access to recent presynaptic spike activity without recurrent state or dense temporal convolutions. Because each directed connection is assigned a single fixed delay, the temporal field is expanded without adding trainable delay parameters. Across 500k, 2M, and 5M core-parameter budgets, fixed delays consistently improve phoneme error rate (PER), with gains saturating once the delay range covers the task-relevant temporal horizon. Boundary-aligned CTC analysis further suggests that delayed temporal context reduces local uncertainty around phoneme transitions, indicating that fixed delays support weakly supervised phoneme alignment rather than simply increasing model capacity. Compared with recurrent, convolutional, and feedforward baselines under matched core-parameter budgets, the proposed SNN achieves the best PER at 500k and 2M parameters and the second-best PER at 5M parameters, while reducing estimated operation-level energy by approximately 2-4$\times$ relative to the strongest non-spiking baselines. These results suggest that fixed-delay SNNs provide an efficient temporal modeling mechanism for brain-to-speech decoding, potentially enabling lightweight decoding for on-device speech neuroprostheses.

Biologically plausible learning requires temporal credit assignment without backpropagation through time (BPTT). Hamiltonian Echo Learning (HEL) achieves this by running a neural system twice --- an \emph{inference} phase then a time-reversed \emph{echo} phase --- recovering exact BPTT gradients. We identify spatio-temporal sum-separability of the Hamiltonian as the structural condition making HEL's learning rule both spatially and temporally local, and exploit it to extend HEL beyond diagonal recurrences: a Hopfield-inspired oscillatory RNN with dense recurrent connectivity yields a contrastive Hebbian rule with constant memory and only two forward passes. This rule matches full BPTT and outperforms e-prop and truncated BPTT on time-series classification and regression. We then relax HEL's Hamiltonian-reversibility constraint --- which forces unbiological connectivity and activation patterns --- and derive \emph{Echo Learning} (EL), exact for arbitrary smooth dynamics whenever the network is reversible and self-adjoint with respect to a readout involution. Both conditions are enforced by a spatio-temporally local homeostatic loss, recovering near-BPTT gradient quality without architectural constraints.

Recent speech brain-computer interfaces (BCIs) increasingly rely on large models and complex multi-stage pipelines for neural decoding. However, in speech BCI, where training data are scarce, noisy, and expensive to collect, the benefits of model scaling remain unclear. This work revisits a fundamental question: $\textbf{are large models truly necessary for accurate neural decoding?}$ We propose $\textbf{\textit{BEST}}$ ($\textbf{\textit{B}}$rain-$\textbf{\textit{E}}$ncoding-$\textbf{\textit{S}}$peech-$\textbf{\textit{T}}$ext), a two-stage framework for efficient neural-to-text decoding. Instead of relying on large generative models or phoneme-centric pipelines, $\textbf{\textit{BEST}}$ introduces hierarchical intermediate representations and a compact speech-oriented decoder that directly maps neural activity to text. This design preserves essential structure while significantly reducing computational cost, enabling robust generalization in low-resource settings and practical deployment under constrained resources. On the Brain-to-Text ’24 and ’25 benchmarks, $\textbf{\textit{BEST}}$ achieves word error rates (WERs) of 6.19\% and 1.62\%, espectively, using only 3.7\% of the parameters required by comparable methods. Beyond these results, we provide three key insights: (1) decoder scaling yields non-monotonic gains; (2) cross-modal alignment is more critical than model size; and (3) overfitting degards sequence-level performance. Representation analysis further reveals a consistent neural→audio→text alignment pathway.

Distributed online convex optimization (D-OCO) is a powerful paradigm for modeling distributed scenarios with streaming data. However, the communication cost between local learners and the central server is substantial in large-scale applications. To alleviate this bottleneck, we initiate the study of D-OCO with compressed communication. Firstly, to quantify the compression impact, we establish the $\Omega(\delta^{-1/2}\sqrt{T})$ and $\Omega(\delta^{-1}\log{T})$ lower bounds for convex and strongly convex loss functions, respectively, where $\delta \in (0,1]$ is the compression ratio. Secondly, we propose an optimal algorithm, which enjoys regret bounds of $O(\delta^{-1/2}\sqrt{T})$ and $O(\delta^{-1} \log T)$ for convex and strongly convex loss functions, respectively. Our method innovatively incorporates the error feedback mechanism into the Follow-the-Regularized-Leader framework to address the coupling between the compression error and the projection error. Furthermore, we employ the online compression strategy to mitigate the accumulated error arising from the bidirectional compression. Our online method has great generality, and can be extended to the offline stochastic setting via online-to-batch conversion. We establish convergence rates of $O(\delta^{-1/2}T^{-1/2})$ and $O(\delta^{-1} T^{-1})$ for convex and strongly convex loss functions, respectively, providing the first guarantees for distributed non-smooth optimization with compressed communication and domain constraints.

Reconstructing audio from brain activity remains challenging. A key limitation lies in current audio diffusion models, which are primarily designed for text-based conditioning and lack mechanisms to incorporate non-linguistic continuous signals such as neural responses. While recent work in the visual domain has explored direct conditioning on brain activity, this approach has not yet been extended to audio diffusion models. We introduce a framework that directly guides a latent audio diffusion model with fMRI signals, avoiding intermediate feature prediction for the conditioning stage. Central to our approach is a transformer-compatible adapter that enables the integration of non-linguistic representations into the cross-attention layers of diffusion transformers (DiTs). Although motivated by neural decoding, this adapter provides a general mechanism for conditioning audio diffusion models on arbitrary continuous inputs, not inherently limited to neural signals. We demonstrate consistent improvements over baselines, including two-stage fMRI-to-latent conditioning, across both acoustic and semantic evaluation metrics. In addition, we introduce a fidelity metric grounded in a computational model of auditory neural processing, which allows us to quantify the contribution of individual auditory cortical regions to reconstructed audio representations.

Optimizing functionals over the space of probability measures is now ubiquitous in machine learning. A widely used approach is to perform the optimization directly over the Wasserstein space, but many objective functionals of practical interest are non-convex along Wasserstein geodesics, making the analysis of standard first-order methods challenging. In this work, we study a class of objectives over the Wasserstein space that admit a difference-of-convex (DC) decomposition and we lift the classical convex-concave procedure (CCCP) to this setting. Under smoothness and strong convexity assumptions on the convex components of the decomposition, we prove almost stationarity along the iterates of the resulting algorithm. Our main focus is on the Maximum Mean Discrepancy (MMD) and the Energy Distance (ED) functionals, for which we develop explicit Wasserstein DC decompositions, and establish local convergence of the scheme under mild assumptions. Empirically, we show that well-chosen DC decompositions yield faster and more stable convergence than Wasserstein gradient descent on these MMD objectives.

More than 80% of the 1.6B English speakers do not use Standard American English (SAE), yet LLMs often fail to correctly identify non-SAE dialects and generate stereotyped responses for their speakers. We introduce **DialectLLM**, the first large-scale framework for generating high-quality multi-dialectal conversational data encompassing the three pillars of written dialect---lexical (vocabulary), orthographic (spelling), and morphosyntactic (grammar) features. DialectLLM produces a dialect-parallel dialog dataset spanning nine English dialects. Partnering with native linguists, we design and validate SAE-to-dialect transformation rules, ensuring authenticity. Our approach challenges the prevailing practice of applying a single morphosyntactic feature set to both user utterances and model responses, showing that models should not reproduce up to 90% of the grammatical features of a dialect. Human evaluation confirms data quality, with annotators preferring DialectLLM over prior methods in 98.8% of pairwise comparisons for dialect naturalness. We then construct **DialectLLM-Bench**, a dialect-parallel benchmark with 50k+ dialogs, resulting in 97k+ QA pairs, and evaluate 17 LLMs on dialect identification and response generation tasks. Even frontier models achieve under 70% accuracy, fail to reach 50% for prominent dialects like Canadian English, and systematically misclassify non-SAE dialects as American or British. Beyond benchmarking, we show that DialectLLM data also serve as a scalable LLM post-training resource, suggesting a practical path toward dialect-aware conversational AI.

Foundation models produce high-dimensional latent representations in which suitably designed statistical decision rules can be effective yet analytically tractable. We study such detectors in the high-dimensional regime, where latent dimension and training sample size grow proportionally. For linear and quadratic statistics, we derive explicit large-deviation characterizations of false-alarm and miss-detection performance under the Neyman--Pearson paradigm, and show how the binary analysis extends to multi-class classification. The resulting error probabilities and corresponding decay rates exhibit non-monotonic dependence on the shape parameter, including double-descent and, in some regimes, multiple-descent. Experiments on vision, text, and audio datasets using modern foundation models show close agreement between theoretical predictions and empirical performance. In particular, the proposed framework allows us to set a prescribed false-alarm level and accurately predict the corresponding miss-detection probability. Overall, our results provide a theoretically grounded and interpretable framework with competitive empirical performance across different modalities.

Fine-scale-faithful neural simulation under fixed storage budgets remains challenging. Many existing methods reduce high-frequency error by improving architectures, training objectives, or rollout strategies. However, under budgeted coarsen-quantize-decode pipelines, fine detail can already be lost when the carried state is constructed. In the canonical periodic incompressible Navier--Stokes setting, we show that primitive and derived fields undergo systematically different retained-band distortions under the same operator. Motivated by this observation, we formulate **Derived-Field Optimization (DerivOpt)**, a general state-design framework that chooses which physical fields are carried and how storage budget is allocated across them under a calibrated channel model. Across the full time-dependent forward subset of PDEBench, DerivOpt not only improves pooled mean rollout nRMSE, but also delivers a decisive advantage in fine-scale fidelity over a broad set of strong baselines. More importantly, the gains are already visible at input time, before rollout learning begins. This indicates that the carried state is often the dominant bottleneck under tight storage budgets. These results suggest a broader conclusion: in budgeted neural simulation, carried-state design should be treated as a first-class design axis alongside architecture, loss, and rollout strategy.

Reinforcement learning (RL) has become a prevalent paradigm for training tool calling agents, which typically requires online interactive environments. Existing approaches either rely on training data with ground truth annotations or require advanced proprietary language models (LMs) to synthesize environments that keep fixed once created. In this work, we propose TRUSTEE, a cost-friendly method for training tool calling agents with dynamic environments fully simulated by free open-source LMs that can be as small as 8B, including task generation, user simulation, tool simulation and trajectory evaluation, paired with an adaptive curriculum learning mechanism that controls task difficulty during training. Our empirical results show that TRUSTEE outperforms baselines which require extra external resources in most cases. These confirm that, with a sufficiently sophisticated design, even simulated environments with a local 8B LM as the backbone could set a strong baseline for tool learning. We hope our proposed paradigm could democratize tool learning and inspire future research on environment scaling with limited resources.

Skill-augmented agents load reusable skills as persistent runtime context, improving task performance but also giving malicious skills a durable channel for steering future actions. Such skills may leak secrets, corrupt code, bypass approvals, or stage data for exfiltration only after a concrete user task and workspace state make the unsafe action appear useful. This makes pre-install vetting insufficient and calls for runtime, task-conditioned protection. We propose Defense-as-Skill, a defense paradigm that implements the runtime guard itself as an installable, inspectable, and editable skill. Our guard, SkillSonar, runs alongside untrusted task skills and checks sensitive actions against the user's task boundary, routing each action to an allow, replan, or confirmation decision without modifying the underlying agent runtime. To study this setting, we construct SCOPE-R, a task-conditioned dataset covering 6 risk families and 21 sub-categories, with 206 attack-confirmed malicious instances and 43 benign tasks. We then improve SkillSonar on the SCOPE-R training subset using runtime guard-skill evolution, a Monte-Carlo Tree Search procedure that evolves the on-disk guard skill from feedback on the rollouts. On held-out SCOPE-R test splits across Claude Code and OpenClaw backends, SkillSonar reduces attack success while preserving benign utility, maintaining modest token overhead, and transferring to OOD attack families. These results show that runtime guards can be skill-native, transparent, and evolvable policy artifacts for skill-augmented agents.

Recent full-duplex spoken dialogue models have demonstrated compelling progress toward human-like interaction, enabling agents to respond with low latency, produce backchannels, and handle user barge-ins. Yet these improvements in conversational dynamics often come with weaker reasoning and instruction-following abilities, revealing a potential tension between interactive dynamics and intelligence capability. In this paper, we argue that such an **intelligence--dynamics** trade-off is not fundamental: conversational dynamics can instead be learned as a separate real-time decision policy from human dialogue data. To this end, we propose DuplexPO, a reinforcement learning (RL) framework that decouples when to speak from what to say. It preserves the semantic response capability of an instruction-tuned assistant, while optimizing its temporal interaction behavior over selected high-impact windows from long human conversations. To quantitatively optimize these dynamics, we formulate the Factorized Conversational Dynamics Reward (FCDR) to enable fine-grained temporal credit assignment for turn initiation, backchanneling, yielding, and regularized participation. The policy is then optimized with a GRPO-style objective. Experiments show that DuplexPO substantially improves full-duplex behaviors, including timely backchannels, smooth turn-taking, and barge-in handling, while maintaining strong reasoning and instruction-following performance. Moreover, improvements in dynamics-oriented metrics are reflected in better user experience, suggesting that optimizing conversational timing as a standalone objective can promote more natural full-duplex interaction.

Random sampling is a fundamental tool in modern machine learning and numerical linear algebra for reducing the computational cost of large-scale matrix problems. Existing analyses, however, rely primarily on subspace embedding guarantees, which do *not* precisely characterize the statistical bias of *nonlinear* random oblique projections induced by sampling, which arises ubiquitously in subsampled least squares and fast low-rank approximation methods. Because (pseudo)inversion is *nonlinear*, these random oblique projections can be systematically *biased* even when the underlying sketch is unbiased, thereby introducing hidden bias into downstream least squares and low-rank approximation solutions. In this work, we develop a unified non-asymptotic theory for random oblique projections in high dimensions. We show that standard random sampling schemes generally induce a systematic statistical *bias* overlooked by classical subspace embedding-style analyses, and we propose a principled debiasing framework to correct it. We illustrate the power of the theory through two canonical applications. For subsampled least squares, we obtain sharp bias--variance characterizations, reveal previously unrecognized statistical *suboptimality* in widely used sampling schemes, and identify when debiasing yields provable improvements. For fast CUR decomposition, we develop a debiased approach with improved approximation accuracy. Numerical experiments further validate our theoretical findings.