NeurIPS 2026 论文列表

LLM agents excel when environments are mostly static and their required information fits in a model’s context window, but they struggle with IT enterprise diagnostic tasks—such as incident management, IT vulnerability analysis, and FinOps anomaly explanation—where operators iteratively mine massive observability data over a curated resource graph to identify the origins that explain observed symptoms, enabling correct remediation. The data in these domains inherently has hidden dependency structure: entities interact, signals co-vary, and the importance of a fact may only become clear after other evidence is discovered. To cope with bounded context windows, agents must summarize intermediate findings before their significance is known, increasing the risk of discarding key evidence. ReAct-style agents are especially brittle in this regime. Their retrieve-summarize-reason loop makes conclusions sensitive to exploration order and introduces run-to-run non-determinism, producing a reliability gap where Pass-at-k may be high but Majority-at-k remains low. Simply sampling more roll-outs or generating longer reasoning traces does not reliably stabilize results, since verifying a diagnosis requires remediation actions that carry operational risk---demanding consistent correctness, not occasional success---and ReAct provides no mechanism for belief revision as evidence accumulates. In addition, ReAct entangles semantic reasoning with controller duties such as tool orchestration and state tracking; execution errors and plan drift degrade reasoning while consuming scarce context. We address these issues by formulating the investigation as abductive reasoning over a dependency graph and proposing EoG (Explanations over Graphs), a disaggregated framework where an LLM performs bounded local evidence mining and labeling (cause vs symptom) while a deterministic controller manages traversal, state, and belief propagation to compute a minimal explanatory frontier. On representative ITBench diagnostics tasks, EoG improves accuracy and run-to-run consistency over ReAct baselines, including a 7x average gain in Majority-at-k F1 score.

As the demand to integrate Artificial Intelligence into high-stakes environments continues to grow, explaining the reasoning behind neural-network predictions has shifted from a theoretical curiosity to a strict operational requirement. Our work is motivated by the explanations of autoregressive neural predictions on dynamic physical fields, as in weather forecasting. Gradient-based feature attribution methods are widely used to explain the predictions on such data, in particular due to their scalability to high-dimensional inputs. It is also interesting to remark that gradient-based techniques such as SmoothGrad are now standard on images to robustify the explanations using pointwise averages of the attribution maps obtained from several noised inputs. Our goal is to efficiently adapt this aggregation strategy to dynamic physical fields. To do so, our first contribution is to identify a fundamental failure mode when averaging perturbed attribution maps on dynamic physical fields: stochastic input perturbations do not induce stationary amplitude noise in attribution maps, but instead cause a geometric displacement of the attributions. Consequently, pointwise averaging blurs these spatially misaligned features. To tackle this issue, we introduce WassersteinGrad, which extracts a geometric consensus of perturbed attribution maps by computing their entropic Wasserstein barycenter. The results, obtained on regional weather data and a meteorologist-validated neural model, demonstrate promising explainability properties of WassersteinGrad over gradient-based baselines across both single-step and autoregressive forecasting settings.

SHAP explanations are widely used in high-stakes settings to justify decisions, yet they can differ substantially across repeated runs, even when the model, the input instance, and the prediction are held fixed. Prior work has documented disagreement *between* explanation methods; we show that substantial disagreement arises even *within* SHAP across reruns of the same estimator on the same trained model and instance. We call this phenomenon *explanation multiplicity* and develop an evaluation methodology for characterizing it under deployment-realistic computational budgets, combining a dual-seed protocol that disentangles model-induced from explainer-induced variability, a hierarchy of magnitude-, rank-, and set-based metrics, and randomized Dirichlet and Mallows null models that calibrate when observed disagreement exceeds chance. Across multiple datasets, model classes, and sampling strategies, we find that explanation multiplicity is pervasive and persists even for high-confidence predictions. The dominant source depends on the data regime: model-induced variability dominates in small-data settings, while explainer-induced variability dominates at scale. Commonly used $\ell_2$ distance is largely insensitive to this instability, while rank-based metrics reveal severe reordering of top-ranked features. Improved sampling methods such as CTE do not eliminate rank-level multiplicity, and deterministic alternatives such as K-Means achieve stability only by targeting a different estimand that diverges sharply from Shapley values defined against the empirical training distribution. Practitioners should treat single-run SHAP outputs as realizations of a distribution rather than as authoritative artifacts.

Reinforcement learning from human feedback (RLHF) typically assumes a static or non-strategic reward model (RM). In iterative deployment, however, the policy generates the data on which the RM is retrained, creating a feedback loop. Building on the Stackelberg game formulation of this interaction, we derive an analytical decomposition of the policy's true optimization gradient into a standard policy gradient and a parameter-steering term that captures the policy's influence on the RM's future parameters. We show that standard iterative RLHF, which drops this steering term entirely, suffers from alignment collapse: the policy systematically exploits the RM's blind spots, producing low-quality, high-reward outputs whose feedback reinforces the very errors it exploits. To mitigate this, we propose foresighted policy optimization (FPO), a mechanism-design intervention that restores the missing steering term by regularizing the policy's parameter-steering effect on RM updates. We instantiate FPO via a scalable first-order approximation and demonstrate that it prevents alignment collapse on both controlled environments and an LLM alignment pipeline using Llama-3.2-1B.

In the realm of multi-objective alignment for large language models, balancing disparate human preferences often manifests as a zero-sum conflict. Specifically, the intrinsic tension between competing goals dictates that aggressively optimizing for one metric (e.g., helpfulness) frequently incurs a substantial penalty on another (e.g., harmlessness). While prior work mainly focuses on data selection, parameter merging, or algorithmic balancing during training, these approaches merely force compromises between divergent preferences along a fixed Pareto frontier, failing to fundamentally resolve the inherent trade-off. In this work, we approach this problem from a novel perspective of multi-dimensional rewards. By scaling up the model's rollouts and analyzing the outputs across different reward dimensions, we arrive at a critical conclusion: the conflict among multiple objectives stems from the fact that the prompt itself inherently restricts the achievable multi-dimensional rewards. Based on this core observation, we propose MORA: Multi-Objective Reward Assimilation (MORA). Specifically, MORA isolates single-reward prompts through pre-sampling and expands their reward diversity by rewriting the original questions to incorporate multi-dimensional intents. Extensive experiments demonstrate that: (I) in sequential alignment, MORA achieves single-preference improvements ranging from 5% to 12.4%, with exceptional gains in harmlessness, after multiple-preference alignment across helpful, harmless, and truthful dimensions. (II) In simultaneous alignment, MORA achieves an average overall reward improvement of 4.6%. Our codes are available at [https://anonymous.4open.science/r/MORA-MPA](https://anonymous.4open.science/r/MORA-MPA).

Explainability, the capability of an artificial intelligence system (AIS) to explain its outcomes in a manner that is comprehensible to human beings at an acceptable level, has been deemed essential for critical sectors, such as healthcare. Is it really the case? In this perspective, we consider two extreme cases, Oracle (without explainability) versus AI Colleague (with explainability) for a thorough analysis. We discuss how the level of automation and explainability of AIS can affect the determination of liability among the medical practitioner/facility and manufacturer of AIS. We argue that explainability plays a crucial role in setting a responsibility framework in healthcare, from a legal standpoint, to shape the behavior of all involved parties and mitigate the risk of potential defensive medicine practices.

Solving optimal transport (OT) on random minibatches is a common surrogate for exact OT in large-scale learning. In flow matching (FM), this surrogate is used to obtain OT-like couplings that can straighten probability paths and reduce numerical integration cost. Yet, the population-level coupling induced by repeated minibatch OT remains only partially understood. We formalize this coupling as the expected batch OT plan $\overline{\pi}\_{k}$, obtained by averaging empirical OT plans over independent minibatches of size $k$. We then establish its large-batch consistency and, in the semidiscrete case relevant to generative modeling, derive rates for both the transport-cost bias and the convergence of $\overline{\pi}\_{k}$ to the OT plan. For FM, this yields a population coupling whose induced velocity field is regular enough to define a unique flow from the source to the discrete target. We finally quantify how OT batch size interacts with numerical integration in a tractable two-atom model and in synthetic and image experiments.

Reinforcement learning (RL) has demonstrated success in post-training large language models (LLMs) as agents for tasks such as computer use, tool calling, and coding. However, exploration remains a central challenge in RL for LLM agents, especially as they operate in language-action spaces with complex observations and sparse rewards. In this work, we address exploration for LLM agents by leveraging the ability of LLMs to plan and reason in language about the environment to shift exploration from low-level actions to higher-level language strategies. We thus propose Strategy-Guided Exploration (SGE), which first generates a concise natural-language strategy that describes what to do to make progress toward the goal, and then generates environment actions conditioned on that strategy. By exploring in the space of strategies rather than the space of actions, SGE induces structured and diverse exploration that targets different environment outcomes. To increase strategy diversity during RL, SGE introduces mixed-temperature sampling, which explores diverse strategies in parallel, along with a strategy reflection process that grounds strategy generation on the outcomes of previous strategies in the environment. Across UI interaction, tool-calling, coding, and embodied agent environments, SGE consistently outperforms exploration-focused RL baselines, improving both learning efficiency and final performance. We show that SGE enables the agent to learn to solve tasks too difficult for the base model.

Flow maps have enabled remarkable progress in few-step generative modeling across both continuous and discrete state spaces. Despite their promise, existing parameterizations are restricted to flows over fixed dimensions or fixed sequence lengths. Here, we introduce **Expanding Generative Flows** (EFlows), which define flows between distributions of *increasing dimensionality* through an expanding interpolant built from *augmented distributions of conditional noise*. Building on this framework, we propose **Expanding Flow Maps** (EFMs), a new class of flow maps that distill EFlows into efficient few-step generative models. Each EFM factors the map between any two timesteps into two learned components: an *expand operator*, which augments the state with new coordinates or tokens, and a *transport operator*, which pushes the expanded state forward along the interpolant. Composing these operators yields a single map that jointly *expands* and *denoises* the state, recovering existing fixed-canvas flow maps as the special case in which the expand operator is the identity. We further extend the framework to the discrete simplex, enabling variable-length sequence generation via token insertions. Across domains, EFlows and EFMs provide a principled approach to generative problems in which output size is itself learned, controllable degree of freedom.

Current large reasoning models (LRMs) have shown strong ability on challenging tasks after reinforcement learning (RL) based post-training. However, previous work mainly focuses on English reasoning in expectation of the strongest performance, despite the demonstrated potential advantage of multilingual thinking, as well as the requirement for native thinking traces by global users. In this paper, we propose ExpLang, a novel LLM post-training pipeline that enables on-policy thinking language selection to improve exploration and exploitation during RL with the use of multiple languages. The results show that our method steadily outperforms English-only training with the same training budget, while showing high thinking language compliance for both seen and unseen languages. Analysis shows that, by enabling on-policy thinking language selection as an action during RL, ExpLang effectively extends the RL exploration space with diversified language preference and improves the RL exploitation outcome. The method is orthogonal to most RL algorithms and opens up a new perspective on using multilinguality to improve LRMs.

Advancing spatial intelligence in Multimodal Large Language Models (MLLMs) is bottlenecked by the scarcity of complex, scalable 3D question-answer (QA) data. While manual annotation is labor-intensive, directly utilizing LLMs to synthesize these QA pairs often fails due to their inherent deficiencies in spatial and geometric computation. We introduce $\textbf{Exemplar2VQA}$, a scalable exemplar-driven visual question answering generation framework that rapidly synthesizes large-scale spatial QA pairs in simulated environments via multi-agent coding. By equipping collaborative agents with a meticulously designed library of geometric utilities, Exemplar2VQA bypasses LLMs' spatial reasoning flaws through deterministic code execution. Crucially, the framework exhibits remarkable versatility: taking diverse static object-centric spatial query templates as exemplars, it seamlessly and autonomously scales them into massive, high-fidelity synthetic datasets. Fine-tuning Qwen2.5-VL (3B/7B) exclusively on Exemplar2VQA-generated synthetic indoor data yields significant performance improvements across various diverse benchmarks. Furthermore, its effectiveness is not limited to in-domain indoor datasets but also robustly extends to outdoor and mixed-scene benchmarks. These results establish Exemplar2VQA as a scalable and powerful paradigm for bridging the sim-to-real gap in Embodied AI. The code will be available upon acceptance.

Text-to-image (T2I) diffusion models often synthesize policy-violating content, necessitating robust safeguards and rigorous evaluation frameworks. However, current approaches remain largely monolithic in both execution and evaluation, limited to coarse-grained categorization within broad unsafe concepts. Such approaches fail to account for the specific constituent factors of a violation, resulting in weak learning signals and allowing model collapse to bypass safety checks. To overcome these limitations, we propose Example-Based Spatial Guidance (EBSG), a training-free method that utilizes user-editable exemplar packs to provide granular spatial guidance for precise concept steering. Unlike existing methods that coarsely steer the entire image away from a broad concept, EBSG decomposes these categories into specific sub-concepts via exemplar text-image pairs, providing explicit, localized steering signals for more robust safety control. Furthermore, we introduce a vision language model (VLM)-based evaluation protocol that provides a fine-grained assessment, avoiding the pitfall of conventional binary evaluators that permit model collapse to bypass safety checks. Empirically, EBSG achieves state-of-the-art erasure performance across diverse safety datasets and concept categories, spanning four nudity benchmarks, seven I2P harmful-concept slices, and four MJA metaphor categories, while supporting multi-concept removal and transferring to a modern backbone, SD3.

Autoregressive scientific forecasters often enforce physical or structural constraints by repairing each predicted state before feeding it back into the model. However, it remains unclear when stronger physical rule enforcement becomes reliable and when it becomes a source of distribution shift. We study this question through operator exactness, meaning whether the repair map is the identity on the target manifold and is aligned with the target geometry. We compare raw forecasting, post hoc repair, and in-loop repair across periodic incompressible Navier--Stokes, non-periodic CFDBench flows, and a hierarchical-forecasting support task. In the exact periodic regime, Fourier projection substantially improves rollout accuracy. On the NS-128 benchmark, a strong Raw-FNO has a final-step rollout MSE at horizon 100 of $(9.390 \pm 6.290)\times 10^{-5}$, and post hoc and in-loop projection reduce it to $(1.130 \pm 0.165)\times 10^{-6}$ and $(5.370 \pm 0.113)\times 10^{-7}$. However, once an exact projection is unavailable and only approximate boundary-preserving cleanup is available, the ordering changes. Across cavity, tube, dam, and cylinder flow, stronger Poisson-based cleanup can reduce divergence while worsening rollout error; target-distortion MSE predicts this harm far better than a linear-system residual. Controlled mismatch, screened cleanup, adaptive gating, and external-backbone checks show that the best approximate-regime operating point can be raw or near-identity. Hierarchical forecasting gives the same broader pattern. Exact forecast reconciliation is a stable baseline, whereas blended top-down repair, a validation-tuned interpolation toward historical-proportion top-down reconciliation, is dataset-dependent. Thus, constraint enforcement should be benchmarked by operator--data alignment before enforcement strength. Use in-loop projection when the operator is exact, and validate approximate cleanup strength using rollout metrics otherwise.

Variable-order Markov models generate by backing off to the longest usable suffix of the generated history, while regular constraints describe finite horizon controls such as fixed endings, metrical patterns, and anti-copy rules. Existing belief-propagation methods give exact regular-constrained sampling for first-order Markov chains, but a first-order state merges histories that a variable-order generator deliberately keeps distinct. We give the corresponding variable-order construction: replace the Markov state by the sparse observed context state and compose this context graph with the regular constraint automaton. For a fixed context graph and automaton, inference is linear in the horizon and in the number of reachable product edges, without expanding to all $|\mathcal{V}|^K$ histories; the same source-row interface handles reversible augmentation by inverse count lookup, without materializing transformed corpora. Tiny enumerated examples verify exact partition functions and conditional probabilities; Bach Prelude experiments show sparse regular-constrained sampling at $K=6$ and 12-key virtual transposition semantics from $592$ stored events.

Nonlinear independent component analysis (nICA) is generally unidentifiable due in part to measure-preserving automorphisms (MPAs), which induce indistinguishable latent representations. We show that for non-Gaussian sources with mild regularity assumptions, such MPAs are not Lipschitz, motivating Lipschitz continuity as a structural assumption. Under this condition, we prove global identifiability of nICA within the class of Orthogonal Coordinate Transformations (OCTs) with bounded sources, recovering the true components up to permutation and scaling. Our analysis reduces the learning problem to a linear program over the Birkhoff polytope, and extends to functions with symmetric Jacobians, revealing connections to optimal transport. We introduce OCTNet and SYMNet, normalizing flow architectures that enforce these constraints in practice. Experiments demonstrate strong recovery performance and robustness to model misspecification, outperforming existing approaches.

Diffusion and flow-based models learn powerful data priors by training a denoiser to reverse Gaussian corruption. To use this prior to solve a linear inverse problem, one needs to sample from the posterior, but the score that the prior provides is the unconditional score, not the posterior score. Existing methods either steer a fixed pretrained denoiser with approximate measurement-matching corrections, or train a conditional restoration model that abandons the denoising structure of the prior. We derive the exact posterior score in closed form for linear Gaussian inverse problems under general Gaussian interpolants, and show that posterior sampling reduces to a denoising problem at an operator-dependent shifted pivot under an anisotropic noise covariance. We turn this identity into **Exact Posterior Score (EPS)**, a denoising training objective that preserves the input/output structure of standard pretraining and can therefore be trained from scratch or fine-tuned from a pretrained denoiser. At inference, EPS uses the same sampler as the underlying backbone, with no likelihood gradients or projections. We evaluate EPS on five linear inverse problems across FFHQ and ImageNet, where it outperforms training-free and training-based baselines on fidelity, perceptual, and distributional metrics, while using roughly an order of magnitude fewer denoiser evaluations than gradient-based posterior samplers.

Diffusion Transformers (DiTs) have emerged as the state-of-the-art architecture in high-fidelity generative modeling, but their massive computational demands and multi-step inference process limit their deployment in practical scenarios. Post-training quantization (PTQ) offers a promising solution for accelerating inference, but DiTs suffer from severe activation outlier issues, which can lead to catastrophic accuracy loss in integer formats. Although some recent outlier mitigation methods employ orthogonal transformations or diagonal scaling to smooth the activation distribution, they either solve the transformation matrix solely for activations or fail to fully balance the inter-channel variance between activations and weights. Additionally, the construction of the calibration set by taking the weighted average of the activation values for each time step introduces spurious temporal correlations. To address these issues, we propose a joint diagonalization of the activation-weight covariance matrix based on the solution of a generalized eigenvalue problem (GEVP). We also transform the feature fusion of the calibration set's temporal dimension into probabilistic sampling and assembly in the spatial dimension using Uniform Token Splicing (UTS). Specifically, compared to suboptimal baseline methods, our approach reduces the FID by 16.06 and 6.49 on both the W3A4-quantized DiT-XL/2 and PixArt-$\Sigma$ models, respectively. This demonstrates that our method successfully suppresses outliers in activations and weights, maintaining DiT's performance under low-bit quantization conditions. Code is available at the anonymous repository \url{https://anonymous.4open.science/r/JDUS-DiT-74E1}.

Building agentic systems that can autonomously self-improve from experience is a longstanding goal of AI. Large language models (LLMs) today primarily self-improve via two mechanisms: self-reflection for context updates, and reinforcement learning (RL) for weight updates. In this work, we propose Evolutionary System Prompt Learning (E-SPL), a method for jointly improving model contexts and model weights. In each RL iteration, E-SPL samples trajectories under multiple system prompts in parallel, then jointly applies RL updates to weights and evolutionary updates to system prompts via LLM self-reflection. E-SPL encourages a natural division between declarative knowledge encoded in prompts and procedural knowledge encoded in weights. Most notably, in an easy-to-hard generalization setting (AIME $\rightarrow$ BeyondAIME), E-SPL improves RL success rate from 38.8% $\rightarrow$ 45.1%. E-SPL also improves RL on AIME 2025 (56.3% $\rightarrow$ 60.6%), HMMT 2025 (50.0% $\rightarrow$ 52.7%), and agentic search (44.2% $\rightarrow$ 48.6% on gpt-oss-120b). Across all settings, E-SPL outperforms both RL-only and evolution-only baselines, demonstrating that weight updates and context updates are deeply synergistic and can together yield gains in generalization that neither achieves alone.

Would experience designing faster GPU kernels also help close in on a long-standing open mathematical conjecture? Large Language Models (LLMs) integrated into evolutionary search have recently produced state-of-the-art solutions on optimization tasks, including open mathematical conjectures, GPU kernel design, scientific law discovery, and combinatorial puzzles. To achieve this, prior work applied search scaffolds to one target task at a time, so every new problem is approached from scratch and the experience accumulated during search is discarded once the model finishes its attempt. This leaves the capability of iteratively evolving a solution (\textit{e.g.}, knowing which part to mutate and how, deciding when to backtrack) entirely in the scaffold rather than in the model itself. Whether the model itself could acquire this capability and reuse it across different tasks has been largely unexamined. To address this, we introduce \textbf{\textsc{Evolution Fine-Tuning} (EFT)}, a mid-training paradigm that teaches LLMs to evolve solutions across tasks by converting evolutionary search trajectories into supervision. We construct \datasetName, a 156K-trajectory dataset spanning 10 domains and 371 optimization tasks, and fine-tune open-source LLMs from 2B to 9B parameters. Empirically, EFT confers cross-task generalization: across 22 held-out tasks, our models surpass their base counterparts by 10.22\% on average. Furthermore, when paired with test-time RL, our model match state-of-the-art performance on two circle-packing tasks and outperforms its base-model counterparts on the Erdős minimum-overlap problem. EFT thus serves as a ``practice phase'' for general-purpose discovery agents that doesn't solve new problems from scratch.