NeurIPS 2026 论文列表

While test-time scaling has revolutionized reasoning in large language models, generative video reasoning remains bottlenecked by a single-shot paradigm. We demonstrate that searching over denoising steps cannot rescue logically flawed rollouts because spatial trajectories commit early in the diffusion process. Root-level Best-of-$N$ (BoN) sampling is similarly inefficient: reasoning errors cluster early in the temporal axis, and resampling blindly discards verified upstream progress. To unlock effective test-time scaling for video models, we introduce \textbf{Temporal Backtracking Search (TBS)}, which shifts the search space to the temporal axis. TBS transforms video generation into an iterative generate--verify--restart loop via three core mechanisms: (1) \textit{variable-$K$ conditioning} to resume generation from arbitrary clean prefixes; (2) \textit{temporal process verification} to localize failures and extract valid restart anchors; and (3) \textit{prefix-based search} to reallocate compute toward extending correct trajectories rather than root resampling. Across algorithmic, navigation, and robotics domains, TBS Pareto-dominates matched-budget BoN. In a strict out-of-distribution setting where one-shot generation collapses ($0.7\%$ for BoN), TBS achieves $22.7\%$, with every solved episode stemming from a restarted branch. Ultimately, TBS reveals that the local reasoning competence of video models far exceeds what single-shot rollouts indicate, providing a scalable test-time framework to unlock it.

Diffusion models have achieved remarkable success as generative models. However, even a well-trained model can accumulate errors throughout the generation process. These errors become particularly problematic when arbitrary guidance is applied to steer samples toward desired properties, which often breaks sample fidelity. In this paper, we propose a general solution to address the off-manifold phenomenon observed in diffusion models. Our approach leverages a time predictor to estimate deviations from the desired data manifold at each timestep, identifying that a larger time gap is associated with reduced generation quality. We then design a novel guidance mechanism, `Temporal Alignment Guidance' (TAG), attracting the samples back to the desired manifold at every timestep during generation. Through extensive experiments, we demonstrate that TAG consistently produces samples closely aligned with the desired manifold at each timestep, leading to significant improvements in generation quality across various downstream tasks.

Foundation models for zero-shot time series forecasting face challenges in efficient long-horizon prediction and reproducibility, with existing synthetic-only approaches underperforming on challenging benchmarks. This paper presents TempoPFN, a univariate time series foundation model based on linear Recurrent Neural Networks (RNNs) pre-trained exclusively on synthetic data. The model uses a GatedDeltaProduct architecture with state-weaving for fully parallelizable training across sequence lengths, eliminating the need for windowing or summarization techniques while maintaining robust temporal state-tracking. Our comprehensive synthetic data pipeline unifies diverse generators, including stochastic differential equations, Gaussian processes, and audio synthesis, with novel augmentations. In zero-shot evaluations on the Gift-Eval, fev-bench and Chronos-ZS benchmarks, TempoPFN achieves top-tier competitive performance, surpassing the majority of models trained on real-world data, while being more efficient than existing baselines by leveraging fully parallelizable training and inference. We open-source our complete data generation pipeline and training code, providing a reproducible foundation for future research.

Training-free conditional diffusion provides a flexible alternative to task-specific conditional model training, but existing samplers often allocate computation inefficiently: independent guided trajectories can vary widely in quality, and additional function evaluations along a single trajectory may not recover from poor early decisions. We propose **Tempered Guided Diffusion (TGD)**, an annealed sequential Monte Carlo framework for training-free conditional sampling with diffusion priors. TGD targets tempered posterior distributions over the clean signal, using noisy diffusion states only as auxiliary variables for proposing reconstructions and propagating particles. Particles are reweighted by incremental likelihood ratios, resampled, and propagated across noise levels, concentrating computation on trajectories plausible under both the prior and observation. Under idealized exact-reconstruction assumptions, full TGD yields a consistent particle approximation to the posterior as the number of particles grows. For expensive reconstruction tasks, **Accelerated TGD (A-TGD)** retains early particle exploration but prunes to a single high-likelihood trajectory partway through sampling. Experiments on a controlled two-dimensional inverse problem and image inverse problems show improved posterior approximation and favorable wall-clock speed-quality tradeoffs over independent multi-trajectory baselines.

Despite the strong performance of diffusion models in molecular generation, the sampling process itself remains largely understudied, with most existing methods adopt off-the-shelf sampling strategies developed for natural images. Yet unlike natural images, molecular distributions are governed by physical laws and are sharply concentrated, making them particularly challenging for standard samplers, which frequently produce invalid or unphysical structures. In this work, we revisit diffusion sampling from a unified stochastic differential equation (SDE) perspective and introduce a general framework parameterized by two interpretable controls: \emph{stochasticity} and \emph{temperature}. Our theoretical analysis reveals that stochasticity accelerates the decay of sampling error, while temperature directly controls distribution sharpness, enabling concentration on physically plausible configurations. Building on these insights, we develop \textsc{TReaSSure} (Temperature-Regulated Stochastic Sampling), a lightweight, training-free strategy that combines stochasticity scheduling with temperature annealing to better capture the sharply concentrated distributions of molecular data. Extensive experiments on small molecule generation, protein structure prediction, and protein design demonstrate that \textsc{TReaSSure} consistently improves generation quality and produces more physically valid structures. Empirical analyses further corroborate our theoretical findings. The method harnesses pretrained models without retraining, underscoring its generality and practical value across diverse molecular domains.

End-to-end autonomous driving systems are known to suffer from shortcut learning and causal confusion, limiting robust generalization under distribution shift. To address this, recent Vision-Language Model (VLM) approaches introduce intermediate linguistic reasoning steps including chain-of-thought, chain-of-causation, and counterfactual (CF) reasoning, and report consistent improvements in closed-loop driving. Among these, CF reasoning has been emphasized as enabling self-reflective, high-level causal understanding, yet whether these gains reflect genuine causal reasoning or language-induced distributional bias remains unclear. In this work, we examine what CF reasoning actually induces in VLM-based autonomous driving using accident scenarios with explicit causal structure and a discrete meta-action space. We observe that CF reasoning consistently shifts the action distribution toward more conservative behaviors, accompanied by increased perplexity and policy entropy, and that this effect persists under reinforcement fine-tuning, suggesting a structural rather than transient artifact. However, visual attention analysis and causal intervention experiments show that CF reasoning does not meaningfully alter visual grounding, and VLMs respond similarly to causal and non-causal perturbations, unlike humans who selectively react to causal factors. Overall, these results suggest that CF reasoning does not induce genuine causal reasoning grounded in visual structure. Instead, it acts as a linguistic mechanism that reshapes the action distribution, teaching models what to say but not how to reason, highlighting a fundamental gap between language-based reasoning and causal understanding in VLMs.

Large language models (LLMs) often fail to reason under temporal cutoffs: when prompted to answer from the standpoint of an earlier time, they exploit knowledge that became available only later. We study this failure through the lens of ex-ante reasoning, where a model must rely exclusively on information knowable before a cutoff. Through a systematic analysis of prompt-level interventions, we find that temporal leakage is highly sensitive to cutoff formulation and instruction placement: explicit cutoff statements outperform implicit historical framings, and prefix constraints reduce leakage more effectively than suffix constraints. These findings indicate that prompting can steer models into a temporal frame, but does not endow them with the ability to verify whether a response is temporally admissible. We further argue that supervised fine-tuning is insufficient, since ex-ante correctness is not an intrinsic property of an answer, but a relation between the answer and the cutoff. To address this gap, we propose \textbf{TCFT}, a \underline{T}emporal \underline{C}ritique \underline{F}ine-\underline{T}uning framework that trains models to acquire cutoff-aware temporal verification. Given a query, a cutoff, and a candidate response, TCFT teaches the model to identify post-cutoff leakage, explain temporal boundary violations, and judge temporal admissibility. Experiments with Qwen2.5-7B-Instruct and Qwen2.5-14B-Instruct show that TCFT consistently outperforms prompting and SFT baselines, reducing average leakage by 41.89 and 37.79 percentage points, respectively.

Chest X-ray visual question answering (CXR VQA) requires models not only to predict correct answers, but also to produce reliable medical reasoning. However, existing reinforcement-learning-based training typically relies on answer-level rewards, which are often too coarse to improve chain-of-thought (CoT) quality and can become ineffective when group-level advantages collapse to zero. We propose \textbf{Teach-to-Reason (T2R)}, a framework that introduces comparison-based supervision into CoT optimization through a self-improving \emph{Teacher} and a competition-guided \emph{Reasoner}. As the Teacher is iteratively strengthened via self-competition, the Reasoner is optimized against progressively stronger Teacher-generated references. We further introduce a case-wise reward design that preserves the original reward-induced positive/negative partition when it is informative, and restores supervision from competition scores when the original reward signal degenerates. Experiments on multiple CXR open-ended VQA benchmarks show that T2R consistently outperforms strong baselines, indicating that comparison-based supervision, when integrated in a controlled and principled manner, provides a more effective training signal for reasoning optimization.

Video large multimodal models incur substantial inference cost because visual tokens accumulate across sampled frames. Training-free token compression can reduce this cost, but it often degrades video reasoning performance when later frames must contribute new evidence over time. We identify a common failure mode behind this degradation: existing compressors do not track the committed support, namely the token support already retained and delivered to the language model by earlier frames. As a result, selectors with different scoring rules repeatedly retain overlapping supports across neighboring frames, a phenomenon we call Selector Collapse. To address this issue, we propose \tempo{}, a frame-causal token compression framework that combines committed-support memory, residual-energy greedy selection, and parameter-free neighborhood fusion. Across four video understanding benchmarks and multiple backbones, \tempo{} consistently improves the accuracy-efficiency trade-off of training-free compression; at 10\% retention, it preserves 99.8\% of vanilla performance on LLaVA-OneVision and improves fixed-budget frame scaling on Qwen2.5-VL to 109.3\% relative accuracy, without retraining, backbone modification, or additional token budget.

Harmful fine-tuning attack becomes a concerning safety risk for mainstream fine-tuning-as-a-service providers, as attackers can submit harmful data to the API to compromise the safety alignment of the large language models. In this paper, we first explore an intuitive gradient mixing solution, and derive a key property ensuring the success of defense -- \emph{taking a fine-tuning update that has higher cosine similarity with the safety gradient can mitigate harmful fine-tuning.} Motivated by this key property, we design an alignment-stage defense, dubbed Tcell. The core contribution of Tcell is a regularizer that align the harmful gradient and the safety gradient during safety alignment, which ensures the harmful fine-tuning update exhibits high cosine similarity with the safety gradient, achieving \emph{gradient alignment}. The benefit of gradient alignment is supported by i) empirical and theoretical interpretation and ii) comparison to alternative design without gradient alignment. Code is available at https://anonymous.4open.science/r/Tcell-1C15/.

Accurately generating images across the Tree of Life is difficult: there are over 10M distinct species on Earth, many of which differ only by subtle visual traits. Despite the remarkable progress in text-to-image synthesis, existing models often fail to capture the fine-grained visual cues that define species identity, even when their outputs appear photo-realistic. To this end, we propose TaxaAdapter, a simple and lightweight adapter that incorporates the taxonomic embeddings of Vision Taxonomy Models (VTMs) such as BioCLIP to perform scalable fine-grained species generation over the entire Tree of Life. TaxaAdapter injects VTM embeddings using taxonomy-text dual conditioning into a frozen text-to-image diffusion model, improving species-level fidelity while preserving flexible text control over attributes such as pose, style, and background. Extensive experiments demonstrate that TaxaAdapter consistently improves morphology fidelity and species-identity accuracy over strong baselines. To better evaluate these improvements, we also introduce a multimodal Large Language Model-based metric that summarizes trait-level descriptions from generated and real images, providing a more interpretable measure of morphological consistency. Beyond showing improvements over standard benchmarking experiments, we observe that TaxaAdapter exhibits strong generalization capabilities in open-world settings, enabling species synthesis in challenging regimes such as few-shot species with only a handful of training images and even species unseen during training. Overall, our results highlight that VTMs are a key ingredient for scalable, fine-grained species generation.

Alignment faking (AF) occurs when an LLM strategically complies with training objectives to avoid value modification, reverting to prior preferences once monitoring is lifted. Current detection methods focus on conversational settings and rely primarily on Chain-of-Thought (CoT) analysis, which provides a reliable signal when strategic reasoning surfaces, but cannot distinguish deception from capability failures if traces are absent or unfaithful. We formalize AF as a composite behavioural event and detect it through observable tool selection, where the LLM selects the safe tool when unmonitored, but switches to the unsafe tool under monitoring that rewards helpfulness over safety, while its reasoning still acknowledges the safe choice. We release a dataset of 108 enterprise IT scenarios spanning Security, Privacy, and Integrity domains under Corruption and Sabotage pressures. Evaluating six frontier LLMs across five independent runs, we find mean AF detection rates between 3.5\% and 23.7\%, with vulnerability profiles varying by domain and pressure type. These results suggest that susceptibility reflects training methodology rather than capability alone.

Sustainable protein discovery lacks the fast computational proxies, analogous to molecular docking or density functional theory, that accelerate drug and materials discovery. Evaluating whether a novel food tastes like its animal-based target requires expensive human sensory panels, bottlenecking the design-build-test loop. We introduce \textsc{TasteBench}, a multimodal benchmark and privacy-preserving competition for sensory prediction, spanning two tasks: a food-level ranking task built on 21K+ human evaluations across 215 plant-based foods in 24 product categories, yielding 935 within-category ranking pairs, and a supporting molecular-level taste classification task over 15K flavor molecules. To enable rigorous interpretation of model performance, we characterize the ground truth: inter-rater agreement among panelists is low (Krippendorff's $\alpha$ = .077), and the split-half reliability ceiling of panel-aggregated rankings is .825, establishing the range within which ML systems on this benchmark should be assessed. We evaluate baselines across four input modalities; on the same pairs panelists rated, the best model achieves .661 pairwise accuracy, competitive with the median individual panelist (.650). TasteBench provides the evaluation infrastructure and baselines for measuring progress on computational screening for sustainable protein discovery.

In real home deployments, household agents must often operate from a complete household scene and a situated household request, rather than from a clean task specification. Such requests require agents to identify task-relevant entities, recover intended task conditions, and resolve ordering constraints from the surrounding scene context. We formalize this capability as full-scene household reasoning: given a complete household scene and a situated household request, an agent must infer executable task structure before producing a grounded skill-level action sequence. This setting is challenging because complete household scenes contain substantial task-irrelevant information, making direct complete-scene prompting inefficient and error-prone. In practical deployment, this challenge is further amplified by privacy and local compute constraints, which favor compact open-weight models with limited long-context reasoning ability. We propose TaskGround, a training-free and model-agnostic Ground-Infer-Execute framework that grounds complete scenes into compact task-relevant scene slices, infers executable task structure, and compiles it into grounded skill-level action sequences. To evaluate this setting, we introduce FullHome, a human-validated evaluation suite of 400 household tasks spanning diverse home-scale environments and both goal-oriented and process-constrained requirements. On FullHome, TaskGround improves task success rates by large margins across both proprietary and open-weight models. Notably, it makes Qwen3.5-9B competitive with GPT-5 under direct complete-scene prompting while reducing total input-token cost by up to 18x. Our results identify executable task-structure inference as a central bottleneck in full-scene household reasoning and show that structured grounding can make compact local models substantially more effective for practical household deployment.

Methods that operate on Vision Transformer features almost universally rely on Euclidean distance to judge feature similarity. Yet Euclidean distance weights every direction in feature space equally, ignoring that downstream tasks concentrate their sensitivity on a low-dimensional submanifold. The metric that respects this structure is the pullback Riemannian metric $g(F) = J(F)^\top J(F)$ induced by the task decoder's Jacobian, but materializing it is prohibitively expensive; a single ViT-B/14 probe point requires storing $4\times 10^{10}$ scalars, and per-token scoring provably fails when task-sensitive directions are spread across tokens. We show that whether a low-rank approximation of $g(F)$ is \emph{learnable} at all depends on a precise structural property of the model-decoder pair, which we characterize via a single scalar diagnostic $\kappa_{cap}(r)$, the captured-energy fraction of the top-$r$ singular directions of $J$, computable matrix-free in $\mathcal{O}(m+rq)$ Jacobian-vector products. When the diagnostic licenses it, we train the \emph{Spectral Pullback Network} (SPN) on directions found by randomized power iteration on $J^\top J$, and distill the resulting per-token signal into a 310K-parameter feature-only \emph{conformal head}, never materializing $J$. When the original spectrum is too rich for any practical rank-$r$ budget, a VAE reparameterization of the decoder recovers a latent space in which the same machinery applies, converting intractable dense pipelines into tractable ones. Across DPT, DINOv2, CLIP, and VGGT backbones the diagnostic correctly predicts which architecture succeeds: the conformal head reaches Spearman $\rho=0.998$ against the Jacobian-derived target on DINOv2 CLS, and geometric token pruning yields a 25\% relative improvement over ToMe at aggressive prune ratios on DPT depth. These results suggest that learning \emph{which} directions matter is often more valuable than learning better features.

Bayesian experimental design (BED) has traditionally been based on maximising expected uncertainty reductions from prior to posterior. A major shortfall of this approach is that it leads to doubly intractable objectives that are difficult to optimise, while customising them to particular downstream tasks of interest can also be difficult. Following first principles decision theory, we demonstrate that BED can alternatively be formulated in terms of an expected future loss (EFL) on downstream actions, providing a simple and naturally task-driven framework. Critically, we then show that all such EFLs can be rearranged into *singly intractable objectives* that can be jointly optimised with respect to both the design policy and a downstream *action policy* using stochastic gradient descent methods. This formulation further sidesteps the need for any explicit posterior or marginal likelihood estimation and is naturally implicit, requiring only the ability to sample from the joint model over model parameters and data, and evaluate the downstream loss function. It thus allows design policies to be learned more effectively, efficiently, and simply than existing methods, while providing easy customisation to different downstream tasks and losses.

KV-cache compression for long-horizon LLM agents is often guided by accumulated attention, yet attention frequency can retain repetitive error traces while discarding decisive instructions. We propose TaskKV, a streaming step-level retention policy with three contributions: (1) a signed directional utility label from the first-order change in action likelihood, distinguishing helpful steps from confusing distractors unlike unsigned saliency; (2) a structured residual scorer that treats attention mass and intent alignment as fixed priors and learns only a denoising correction, converging with 500 trajectories; and (3) an entropy-based regime classifier that falls back to contiguous retention when the scorer cannot confidently rank steps. With a 20% KV budget on AgentBench and AlfWorld, TaskKV preserves $\sim$87--89% of full-cache success rate, improving over SnapKV by 5.0--7.9 SR points and over a strong heuristic by 2.9--7.3 points; on dense tasks, gains come from detecting when utility pruning should be disabled. Code is available at https://anonymous.4open.science/r/taskkv-F0D5.

Tool-using language model agents are usually evaluated by whether they finish the user's task. That metric misses a common failure: an agent can complete the requested work while also taking an unauthorized side-effecting action, such as writing to the wrong object, expanding the scope of the request, or contacting an unintended recipient. We call this failure mode collateral damage (CD). We introduce SABench (Side-effect-Aware Benchmark), a benchmark of 201 email, calendar, and ticket tasks. Each task crosses one of four state structures (independent, dependent, externally reachable, cross-tool) with one of four trap families (ambiguity, mis-targeting, cascade harm, overreach), and each task has a deterministic trace-based oracle for both goal satisfaction and authorization compliance. Across 2,211 model-task pairs from 11 frontier LLMs spanning major commercial and open model families, 15.2% of all runs incur CD, with risk concentrated in ambiguity tasks (38.8% violation rate) while non-ambiguous traps average 8.4% CD. Among goal-satisfied runs, 12.8% still produce unauthorized writes. Cross-tool and dependent states have about twice the CD rate of independent states. CD rates vary by 4.7x across models at comparable goal rates, which suggests that task success alone is a poor proxy for safe tool use. The SABench task corpus is released at https://huggingface.co/datasets/anon-sabench-2026/sabench; the traces, oracle, code, and analysis pipeline are released at https://anonymous.4open.science/r/sabench-artifact.

The *synthetic control method* (SCM) estimates causal effects in panel data with a single-treated unit by constructing a counterfactual outcome as a weighted combination of untreated control units that matches the pre-treatment trajectory. In this paper, we introduce the **targeted synthetic control (TSC)** method, a new two-stage estimator that directly estimates the counterfactual outcome. Specifically, our TSC method (1) yields a targeted debiasing estimator, in the sense that the targeted updating refines the initial weights to produce more stable weights; and (2) ensures that the final counterfactual estimation is a convex combination of observed control outcomes to enable direct interpretation of the synthetic control weights. TSC is flexible and can be instantiated with arbitrary machine learning models. Methodologically, TSC starts from an initial set of synthetic-control weights via a one-dimensional targeted update through the weight-tilting submodel, which calibrates the weights to reduce bias of weights estimation arising from pre-treatment fit. Furthermore, TSC avoids key shortcomings of existing methods (e.g., the augmented SCM), which can produce unbounded counterfactual estimates. Across extensive synthetic and real-world experiments, TSC consistently improves estimation accuracy over state-of-the-art SCM baselines.