NeurIPS 2026 论文列表

Direct pairwise AUC-surrogate training is attractive because it trains toward the ranking metric used at test time, especially under class imbalance. This creates a direct alignment between the training objective and the evaluation metric, but such metric alignment alone does not guarantee that stochastic training remains corrective. We study this gap through Corrective Power: whether pairwise mini-batch updates produce stronger and more reliable improvements for more severe positive-negative ranking errors. Our main observation is that direct pairwise training can enter regions where severe errors remain common but no longer produce useful corrective motion. In such regions, even the mini-batch noise that would normally help SGD move along the steepest severe-error correction direction can become much weaker. In this paper, we seek a theoretical explanation for this by exploring the degenerate property of the U-statistics formed from the pairwise gradients therein. Our results provide a mechanistic explanation for several empirical patterns: (a) squared-hinge loss is often an effective AUC surrogate; (b) cross-entropy warm-up can help models that enter the AUC phase from a poor starting region; and (c) parameter-efficient fine-tuning from a pretrained model can reduce that need by starting closer to a favorable ranking geometry. Experiments on 8 CV tasks and 8 NLP tasks, using ResNet50, DenseNet121, DistilBERT-base, and Qwen2.5-1.5B-Instruct, support these predictions.

Modern agentic systems extend language models with external tools, enabling them to complete complex tasks through multi-turn interactions. As these agents are deployed in consequential settings, their behavior is often constrained by policies governing pricing, privacy, confirmation, safety, and other operational requirements. However, existing evaluation methods leave a critical gap: capability benchmarks measure what agents can accomplish, while content-safety red-teaming measures harmful generations, but neither tests whether agents comply with predefined policies during otherwise valid tool use. Such policy violations are difficult to audit because their effects may emerge only after several turns, span multiple tool calls, or surface in downstream systems. We propose PROBE (Policy Rule Observation via Behavioral Exploration), an algorithm for training LLM-based auditors that interact with deployed target agents as realistic users and probe them for policy violations through multi-turn, tool-mediated dialogue. We formalize agent auditing as a two-player Markov game and propose a composite reward that balances violation discovery, behavioral plausibility, conversation progress, and policy diversity. To study the role of training objectives and reward design, we compare GRPO, DPO, and GFlowNet-based optimization in a controlled setting. Across extensive experiments, auditors trained with PROBE uncover substantially broader and higher-yield policy violations than same-architecture zero-shot adversaries, even when the auditor is significantly smaller than the target agent. Our results show that multi-turn behavioral auditing is a distinct and necessary axis of agent evaluation, complementing capability and content-safety benchmarks. More broadly, PROBE offers a scalable path toward automated policy auditing for real-world agentic systems and suggests a future co-evolutionary loop in which auditors expose failures and agents improve from auditor feedback.

The transition from Large Language Models (LLMs) to agents shifts safety stakes from toxic text to irreversible environmental harm. While current defenses remain largely retrospective, proactive runtime intervention is bottlenecked by the lack of large-scale, causally-consistent data. We propose PROACT-Agent, a framework for synthesizing high-fidelity trajectories to enable real-time guardrails. We identify a critical "safety drift" in prior benchmarks, where lenient annotation paradigms fail to enforce temporal consistency. PROACT-Agent addresses this through: (1) Progressive Trajectory Unrolling to reveal risks hidden in long-context interactions; (2) Reasoning-Augmented Causal Rectification to enforce monotonic causal consistency; and (3) Culturally-Aware Data Localization for cross-border robustness. We introduce PROACT-Bench, the largest bilingual safety benchmark to date, featuring 140,000+ trajectories adjudicated with industrial-grade rigor. Experiments show that models trained via our framework achieve superior zero-latency intervention, establishing a new standard for real-time autonomous agent safety.

Large language models (LLMs) have recently improved sequential recommendation, yet the task remains a retrieval problem without explicit user queries: the system must infer the next item from user histories where diverse intents and preferences are intertwined. Existing methods typically compress this heterogeneous evidence into a single embedding or query, while LLM-based recommenders can produce overly general queries that are semantically plausible but weakly aligned with retrieval. We propose PRISMIC (Preference Reconstruction via Intent Synthesis and Multi-signal Inference Consolidation), which reformulates sequential recommendation as multi-signal intent consolidation. PRISMIC first generates multiple candidate queries from different historical signals, then trains an LLM-based consolidator with GRPO using an NDCG-based retrieval reward to merge relevant signals into a single query. The resulting consolidated query is not directly used at inference time; instead, it serves as a semantic supervision target for training a user encoder, enabling LLM-free inference. Across four real-world datasets, PRISMIC consistently outperforms strong baselines. Ablations further show that the gains come not from GRPO alone, but from consolidating multiple intent-signals and distilling the consolidated intent into an encoder.

Parameter-efficient fine-tuning (PEFT) methods such as low-rank adaptation have become essential for adapting large language models to downstream tasks. Existing approaches often apply low-rank updates uniformly across all layers, ignoring the structural coupling inherent in transformer architectures. We observe that transformers contain \emph{weight pairs} whose interactions govern model behavior including query-key products, value-output products, and feedforward networks. We propose PRincipal Interaction Subspace Matching (PRISM), a principled framework that exploits this coupling through spectral parameterization. For each weight pair, we parameterize adaptations using the spectral basis of the paired matrix, ensuring that updates occur in the subspace where the paired matrix has maximal effect. We validate our approach through controlled experiments on synthetic models, demonstrating faster convergence. Empirical evaluations on arithmetic reasoning, code generation, and vision tasks show that PRISM outperforms existing PEFT methods.

RGB cameras struggle on the road surface itself: pavement is nearly texture-less at the centimetre scale, and dry, wet, and icy asphalt can look photometrically identical. Polarization is the natural complement, since Fresnel optics ties the angle of linear polarization (AoLP) to surface-normal azimuth and the degree of linear polarization (DoLP) to refractive index, but no public dataset has made it possible to test whether this physical promise translates into measurable gains on real roads. We introduce PRISM, a polarimetric road-surface dataset of 47,098 time-synchronized frames combining trichromatic linear polarization, co-boresighted RGB, 128-channel LiDAR, and RTK-GNSS/INS, collected across proving-ground and open-road environments under clear, overcast, rainy, foggy, and snowy conditions. Dense road-surface elevation ground truth, validated against as-designed proving-ground geometry, accompanies frame-level labels over five materials and five surface states. Two benchmark tracks are defined on these data: road-surface condition classification and bird's-eye-view elevation estimation. Both share a five-variant input ablation crossing RGB, monochromatic and trichromatic polarization, and their combinations. Reference baselines turn the physical promise into a measurable one. Adding trichromatic polarization to RGB reduces elevation MAE from 3.61 to 3.26cm and improves every threshold metric, while yielding consistent gains on condition classification across backbones. The improvement is task-conditional: trichromatic resolution drives the elevation gain, where AoLP encodes geometric information that intensity cannot, and stratified evaluation localizes where polarimetric channels matter most. PRISM, together with reference baselines, evaluation and stratification scripts, and a datasheet, is publicly available at \url{https://huggingface.co/datasets/NeurIPS-2026-PRISM/PRISM-Dataset}.

Standard single-cell differential expression (DE) analysis identifies genes that change across conditions, but it usually overlooks two key sources of heterogeneity that single-cell data are uniquely positioned to reveal: which cells within a donor are truly disease-affected, and how disease effects depend on cell state or subject-level context. Assigning every cell the diagnosis of its donor can contaminate DE signal when disease penetrance is partial, while modeling each gene with a single disease effect cannot capture heterogeneous responses across cell populations. We present PRISM (Phenotype-Resolved Inference in Single-cell Mixed models), a negative-binomial mixed-effects framework that augments standard DE analysis with three outputs: a context-DE vector $\theta_g$ that tests whether disease effects vary along a biological axis $z_{ij}$ (such as cell state, sex, or age); a cell-level disease posterior $q_{ij}$ that provides unsupervised disease annotation; and a subject-level disease burden $\rho_i$. Naive likelihood-only inference of $q_{ij}$ is not identified onto the disease axis when nuisance variation (batch, cell-cycle, dominant subtypes) competes with disease; we resolve this with a closed-form 1-D Wasserstein-2 projection of the disease-arm posterior onto a bimodal reference marginal that enforces the \{healthy, affected\} structure implied by the disease label.

We introduce Post-Optimization Model Edit (POME), a new algorithm that enhances the performance of fine-tuned large language models using only their pretrained and fine-tuned checkpoints, without requiring extra data or further optimization. The core idea is to apply a muon-style projection to $\Delta W$, the difference between the fine-tuned and pretrained weights. This projection uses truncated singular value decomposition (SVD) to equalize the influence of dominant update directions and prune small singular values, which often represent noise. As a simple post-processing step, POME is completely decoupled from the training pipeline. It requires zero modifications and imposes no overhead, making it universally compatible with any optimizer or distributed framework. POME delivers consistent gains, boosting average performance by +2.5\% on GSM8K and +1.0\% on code generation. Its broad applicability—from 7B foundation models to 72B RLHF-instructed models—establishes it as a practical, zero-cost enhancement for any fine-tuning pipeline.

Diffusion-based solvers have become a prominent approach to neural combinatorial optimization. At inference time, they assign each decision variable a prediction score. A greedy decoder ranks candidates by these scores and checks hard constraints to ensure feasibility. Top-ranked candidates should be both promising and jointly feasible, making prediction--constraint alignment critical. When this alignment is weak, greedy decoding skips constraint-violating candidates and falls back to lower-ranked alternatives, degrading solution quality. Existing test-time methods improve search, sampling, or decoding, but do not directly strengthen this alignment before decoding. We observe that trained denoisers already carry constraint-aligned structural signals. Building on this observation, we propose Perturbed Instance-Structure Guidance (PISG), a training-free method that amplifies these signals at test time. At each denoising step, PISG constructs a structure-degraded prediction and extrapolates the unperturbed prediction away from it. The resulting guided prediction improves candidate rankings before constrained decoding. Signal analysis shows that this guidance concentrates on constraint-critical variables rather than uniformly rescaling scores. Under greedy decoding and without downstream refinement, PISG improves all 16 TSP solver-scale pairs across DIFUSCO, T2T, FastT2T, and StruDiCO, and all four MIS-ER solvers, with gains up to +8.4\%. These results identify prediction-constraint alignment as an effective test-time enhancement direction for diffusion-based combinatorial optimization.

Backpropagation (BP) is widely viewed as biologically implausible, in part because it requires feedback weights to be the transpose of forward weights for error propagation. Interestingly, when training a network with fixed random feedback weights to circumvent this issue, learning aligns the forward weights with the feedback weights, leading the backpropagated error signal to become an approximation of the standard gradient used by BP. This process, called Feedback Alignment (FA), occurs in MLPs and very shallow CNNs but does not scale well to deeper architectures. In this work, we first investigated differences between BP and FA models, trained on CIFAR10, specifically focusing on the effective rank of the signal. We found that the FA error has a considerably lower rank and hence is constrained to a lower-dimensional subspace compared to BP, limiting exploration of the parameter space. Motivated by this observation, we evaluated two mechanisms for increasing the effective dimensionality of FA: Muon, an optimiser that orthogonalises weight updates; and hidden activity normalisation, which promotes activation orthogonality. Across larger architectures and benchmarks, we find that these methods consistently improve over FA baselines, for example, on CIFAR100 with a Resnet-18, accuracy increases by 9 percentage points. Our results identify low-dimensional gradient dynamics as a key obstacle to scaling FA and suggest that inducing higher-dimensional update geometry is a promising route toward scaling alternatives to backpropagation.

Characterizing the relationship between brain anatomical wiring and functional coordination remains a fundamental challenge in computational neuroscience. This difficulty primarily arises because structure-function coupling in the human brain is spatially heterogeneous, often localized to subnetworks, and inherently difficult to align across disparate representational modalities. To address this methodological gap, we introduce $\textbf{O}$rthogonal $\textbf{S}$parse $\textbf{S}$ubgraph $\textbf{A}$lignment ($\textbf{Ossa}$), a principled and interpretable framework designed to identify compact structural cliques that exhibit maximal concordance with functional connectivity (FC) across subjects. Within each FC-derived community, $\textbf{Ossa}$ optimizes an alignment objective over an orthogonal transformation, which absorbs coordinate mismatch between the two connectivity modalities, and a sparse node-selection vector, which identifies the induced structural connectivity (SC) subgraph. Furthermore, we establish theoretical properties for the $\textbf{Ossa}$ estimator, demonstrating its statistical consistency in recovering the true underlying SC clique within each functional community. Extensive empirical evaluations on simulated data, alongside analyses of the Adolescent Brain Cognitive Development (ABCD) baseline and longitudinal cohorts, demonstrate that $\textbf{Ossa}$ successfully recovers aligned structural cliques under latent cross-modal rotations. Ultimately, the proposed methodology significantly improves out-of-sample SC-FC alignment over size-matched baselines and robustly identifies stable, single-core topological structures within human brain networks.

We study online alignment of large language models under misspecified preference feedback, where the observed preference oracle deviates from an ideal but unknown ground-truth oracle. The online LLM alignment problem is a bi-level reinforcement problem due to the coupling between data collection and policy updates. Recently, the problem has been reduced to a tractable single-level objective in the SAIL (Self-Improving Efficient Online Alignment) framework. In this paper, we introduce a pointwise oracle uncertainty set in this problem and formulate an oracle-robust online alignment objective as a worst-case optimization problem. For log-linear policies, we show that this robust objective admits an exact closed-form decomposition into the original loss function plus an explicit sensitivity penalty. We develop projected stochastic composite updates for the resulting weakly convex objective and prove $\widetilde{O}(\varepsilon^{-2})$ oracle complexity for reaching approximate stationarity.

We study the contextual combinatorial semi-bandit (CCSB) problem with general reward function approximation. At each round, the learner observes a context, selects a combinatorial action consisting of a subset of basic arms, and receives the reward of each selected arm; the goal is to maximize the cumulative reward over time. We propose SquareCB.Comb, a computationally efficient algorithm that, at each round, solves a convex optimization problem to sample a combinatorial action that balances exploration and exploitation. SquareCB.Comb scales to large arm sets and imposes no structural assumptions on the action set beyond a cardinality bound of $m$ on each combinatorial action. We prove that SquareCB.Comb achieves a minimax optimal regret bound of $\tilde{O}{\sqrt{m A T \log |\mathcal{F}|}}$, where $A$ is the number of arms, $m$ is the maximum number of arms in a combinatorial action, $T$ is the time horizon, and $\mathcal{F}$ is the reward function class. In the realizable setting, this bound matches the state-of-the-art regret guarantees achieved by policy search-based algorithms in the more restricted slate recommendation settings, while simultaneously generalizing to arbitrary combinatorial action structures and general reward function approximation.

Machine learning models often suffer performance degradation under subpopulation shift, particularly when spurious correlations cause models to rely on shortcut features that fail to generalize across subgroups. A recent line of work mitigates this issue by using loss-based signals to identify informative samples, but these signals can become severely distorted under label noise: mislabeled samples may also incur large losses and contaminate subsequent reweighting or retraining. Despite its practical importance, this intersection remains largely underexplored. We propose POTER, a reweighting framework based on optimal transport that derives sample importance from the transport geometry between the training distribution and a reference distribution constructed from limited validation group annotations. By measuring alignment at the individual-sample level rather than relying on loss, POTER downweights mislabeled or strongly bias-aligned samples while assigning higher importance to samples better aligned with the reference distribution. In addition, POTER requires only a single ERM training stage, moving beyond the retraining paradigm common in recent work. Across standard benchmarks and noisy-label settings, POTER achieves state-of-the-art worst-group accuracy, including cases where label corruption is concentrated within minority subgroups.

Learning to generalise from limited data is a fundamental challenge for both artificial and biological systems. A common strategy is to extract reusable structure from abundant unlabelled data, enabling efficient adaptation to new tasks from limited labelled data. This two-stage paradigm is now standard in modern training pipelines, where pretraining is followed by fine-tuning or linear probing. We provide an analytical model of this process: structure extraction is formalized as principal component analysis on unlabelled data, and downstream learning as linear regression on a separate labelled dataset. In the high-dimensional regime, we derive exact expressions for training and generalisation error showcasing their dependence on representation dimensionality, unlabelled and labelled sample sizes, and task alignment. Our results show that pretrained representations strongly influence downstream generalisation, and we characterize the optimal representation size as a function of task parameters: with abundant pretraining data but scarce downstream data, maximally compressed representations are optimal, whereas with limited pretraining data, higher-dimensional representations generalise better. Furthermore, we establish an exact trade-off between pretraining and supervision, quantifying how much unlabelled data is required to replace a single labelled sample. Beyond our idealised model, we observe similar phenomenology in autoencoders and pretrained LLMs. Altogether, we highlight that optimising representation size is critical, giving conditions for when compression during pretraining improves generalisation.

We study in-context learning problems where a Transformer is pretrained on tasks drawn from a mixture distribution $\pi=\sum_{\alpha\in\mathcal{A}} \lambda_{\alpha} \pi_{\alpha}$, called the pretraining prior, in which each mixture component $\pi_{\alpha}$ is a distribution on tasks of a specific difficulty level indexed by $\alpha$. Our goal is to understand the performance of the pretrained Transformer when evaluated on a different test distribution $\mu$, consisting of tasks of difficulty $\beta\in\mathcal{A}$, and with potential distribution shift relative to $\pi_\beta$. In particular, we consider nonparametric regression problems with random smoothness, and multi-index models with both random smoothness and random effective dimension. We prove that a Transformer pretrained on such mixture distributions can adapt to the difficulty of a new task in-context and achieve the optimal prediction rate, uniformly over test distributions in a chi-squared divergence ball. Thus, the pretrained Transformer is able to achieve faster rates of convergence on easier tasks and is robust to distribution shift at test time. Finally, we prove that even if an estimator had access to the test distribution $\mu$, the convergence rate of its expected risk over $\mu$ could not be faster than that of our pretrained Transformers, thereby providing a more appropriate optimality guarantee than minimax lower bounds.

Understanding model predictions is essential for physical applications, where outputs often inform safety-critical decisions, such as structural load assessment, weather warnings, and clinical diagnosis. Shapley values satisfy many desirable properties as an attribution method, but their computational cost during inference hinders their practical use. Current amortized explainers, such as FastSHAP, are limited to homogeneous inputs, which is problematic for physical applications where data often comes from irregular grids and geometries. We introduce OperatorSHAP, a grid-agnostic attribution method and training procedure that allows us to train FastSHAP-like explainers for neural operators. We establish a theoretical framework for attributions in function space, connecting to Aumann–Shapley values. We further show that OperatorSHAP's explanations are consistent with state-of-the-art discrete Shapley values across resolutions and transfer across grid sizes without retraining.

Backpropagation is the cornerstone of deep learning, but its reliance on symmetric weight transport and global synchronization makes it computationally expensive and biologically implausible. Feedback alignment offers a promising alternative by approximating error gradients through fixed random feedback, thereby avoiding symmetric weight transport. However, this approach often struggles with poor learning performance and instability, especially in deep networks. Here, we show that a one-time soft alignment between forward and feedback weights at initialization enables deep networks to achieve performance comparable to backpropagation, without requiring weight transport during learning. This simple initialization condition guides stable error minimization in the loss landscape, improving network trainability. Spectral analyses further reveal that initial alignment promotes smoother gradient flow and convergence to flatter minima, resulting in better generalization and robustness. Notably, we also find that allowing moderate deviations from exact weight symmetry can improve adversarial robustness compared to standard backpropagation. These findings demonstrate that a simple initialization strategy can enable robust and effective learning in deep networks without dynamic weight transport, providing a resource-efficient alternative to standard backpropagation.

Aligned models can misbehave in several ways: they are often sycophantic, fall victim to jailbreaks, or fail to include appropriate safety warnings. Consistency training is a promising new alignment paradigm to mitigate such failures by training invariants into the model using contrastive input pairs. Existing consistency training procedures generate the supervision signal once, offline, and use supervised fine-tuning (SFT) to update the model. Unfortunately, the resulting models tend to merely memorize the surface forms of the training distribution and thus generalize poorly and regress in their capabilities. We introduce On-Policy Consistency Training (OPCT), a new consistency training approach where the objective is computed over the model's own responses to prompts, supervised by itself conditioned on corresponding contrastive prompts. We evaluate OPCT on three safety axes: sycophancy, jailbreaking, and safety awareness. Across three model families, OPCT outperforms its SFT counterpart on all safety desiderata. It nearly halves the sycophancy rate relative to baseline (8.1\% vs. 15.4\%, compared to 11.2\% for SFT). Under an adaptive per-target attacker, OPCT holds jailbreak defense success near 99\% on held-out jailbreak behaviors, whereas SFT achieves 87\% on average. On safety awareness, OPCT outperforms SFT in two out of three models, and matches it on the other. OPCT also largely avoids the capability regressions that SFT induces, such as a 28-point drop on MATH-500. Our results suggest that consistency training is best implemented as OPCT rather than as SFT, especially when generalization beyond the training distribution is desired.