NeurIPS 2026 Papers

Reinforcement learning (RL) post-training has become an important paradigm for improving multimodal large language models (MLLMs), especially with automatically verifiable rewards. For vision-centric MLLMs, self-supervised visual pretext tasks provide such rewards without human annotations. However, existing self-supervised RL methods mostly exploit within-instance structure, such as spatial or temporal ordering within a single image or video, offering limited supervision for cross-instance discrimination and fine-grained comparison. We introduce Group-and-Order Self-Supervised Reinforcement Learning (GO-SSL), a cross-instance jigsaw framework for vision-centric MLLM post-training. Given two visual instances, GO-SSL mixes their local elements and trains the model to group them by instance of origin while recovering the order within each group, coupling instance-level comparison with spatial, temporal, or geometric structure recovery. This paired formulation turns the same unlabeled data into richer verifiable supervision through diverse pairings, shuffles, and hard-pair curricula. We instantiate GO-SSL mainly on cross-image jigsaw tasks and further extend it to 3D depth and video temporal ordering. Across diverse benchmarks, GO-SSL consistently improves over the base MLLM and prior baseline methods, with clear gains in fine-grained perception and spatial understanding. These results suggest a data-efficient direction for self-supervised RL in MLLMs: constructing comparative cross-instance contexts can provide richer transferable supervision than merely scaling data volume or adding isolated pretext tasks. Code will be released later.

On-policy distillation (OPD) has emerged as an efficient post-training paradigm for large language models. However, existing studies largely attribute this advantage to denser and more stable supervision, while the parameter-level mechanisms underlying OPD's efficiency remain insufficiently understood. In this work, we argue that OPD's efficiency stems from a form of ``foresight'': it establishes a direct and stable path from the initial model to the final model early in training. This foresight manifests in two aspects. First, at the \textbf{Module-Allocation Level}, OPD identifies regions with low marginal utility and concentrates updates on modules that are more critical to reasoning. Second, at the \textbf{Update-Direction Level}, OPD exhibits stronger low-rank concentration, with its dominant subspace aligning closely with the final update subspace early in training. Motivated by this theory, we propose \textbf{EffOPD}, a plug-and-play acceleration method that speeds up OPD by searching for an effective step size and extrapolating along the current update direction. EffOPD requires no additional trainable modules or complex hyperparameter tuning, and achieves up to $3.5\times$ training acceleration while maintaining comparable final performance. Overall, our findings provide a parameter-dynamics perspective for understanding the efficiency of OPD and offer practical insights for designing more efficient post-training methods for large language models. Our code is available at: https://anonymous.4open.science/r/EffOPD-7C58.

In-context watermarking (ICW) prepends an instruction to a query asking the model to embed a statistically detectable signal in its response. It thus equips LLMs with a watermarking interface that third parties can invoke without access to model internals. Its reliability hinges on the LLM following the instruction without degrading answer quality, yet how well current LLMs do so has not been measured. We introduce ICWBench, a benchmark of three verifiable ICW instruction families, each scored on both detectability and answer quality. Evaluating 14 frontier proprietary and open-source LLMs, we find that none of the evaluated LLMs achieves both objectives across all three families. To address this, we propose a self-contained two-stage training method, requiring no distillation from a stronger model, no manual annotation, and no pre-existing ICW IF ability. The first stage, self-distillation with logits perturbation (SDLP), uses the same base LLM as both teacher and student: an instruction-equivalent decoding-time logits perturbation makes the teacher follow the ICW instruction, and the student is trained to match the teacher's output distribution. The second stage applies reinforcement learning with the automatic verifier as the reward. Applied to Qwen3-14B, the weakest of the 14 evaluated LLMs in ICW IF, our method raises average TPR@$1\%$FPR across three ICW instructions from $0.100$ to $0.974$, achieving a more favorable trade-off than the frontier proprietary LLMs we evaluate.

Reinforcement Learning with Verifiable Rewards (RLVR) enhances large language model (LLM) reasoning, yet growing evidence indicates an \emph{exploration ceiling}: they tend to reweight existing solution traces and produce low-diversity rollouts. This restricts the discovery of novel solution perspective and weakens out-of-distribution generalization. We address this limitation by directly perturbing the rollout policy's parameters while updating the clean policy model as usual, an approach that achieves better temporal consistency and yields more complex behavior patterns. To stabilize learning from these perturbed rollouts, which are inherently off-policy, we incorporate truncated importance sampling. Furthermore, we introduce a lightweight, dynamic noise scheduler based on semantic diversity and model self-certainty to optimally adjust the noise scale. Instantiated on GRPO, our method (PSN-GRPO) demonstrates its efficacy through four key contributions: (1) it improves high-budget pass@$k$, increases semantic and operational diversity, increases CoT consistency, and remains complementary to exploration-oriented RLVR techniques; (2) it validates efficacy across multiple models, including Qwen2.5, Qwen3, and Llama-3.1; (3) it shows robust performance on out-of-distribution science and coding benchmarks; and (4) it uncovers novel solution perspectives for problems the original model could not solve, as shown by qualitative analysis.

Understanding human behavior while interacting with the surrounding world is crucial for many applications of embodied AI. First-person videos are particularly informative for this problem, as they well capture how activities reshape the scene over time. However, existing approaches often rely on implicit visual or language-aligned representations, disregarding structured reasoning over the scene dynamic. In this paper we argue that explicit, compositional and editable representations of human-environment interactions can play a crucial role for rich grounded activity understanding. To this end, we introduce SG-Ego, a large scale annotation set extending Ego4D with spatio-temporal scene graphs, where relations triplets are consolidated over time into explicit time-evolving descriptions of the scene state. To effectively reason over this representation, we propose GLEN, a graph-based model that operates over scene graph sequences to both align them with textual actions and model their temporal evolution. In addition, we formulate the activity-driven graph-edit forecasting (A-GEF) problem, a novel task that casts scene dynamics as a sequence of structured transformations conditioned on ongoing actions, enabling explicit reasoning about how scenes change over time. We demonstrate the effectiveness of our approach across multiple downstream tasks, spanning retrieval benchmarks as EgoMCQ and the more complex compositional EgoCVR, as well as long-horizon reasoning benchmarks as EXPLORE-Bench and the newly introduced A-GEF. GLEN achieves competitive or state-of-the-art performance compared to raw video baselines and it excels in reasoning settings, typically addressed only with MLLMs, while enabling controllable and structured predictions of scene dynamics driven by human activities. We believe our results establish spatio-temporal scene graphs, together with models that reason over them, as strong compositional and interpretable representations for egocentric video understanding and potentially beyond. Code will be released upon acceptance.

Self-supervised monocular depth estimation is appealing for its potential to scale with unlabeled data, yet existing methods often degrade when trained on diverse scenes. This study finds that, under self-supervision, models tend to encode dataset-specific structural biases rather than transferable scene geometry, leading to cross-domain optimization conflicts. To address this issue, this study proposes the MaSS framework, aiming to make self-supervised depth learning scalable to diverse data. MaSS introduces a dedicated structural representation pathway explicitly decoupled from contextual features, which is guided by a novel scene-structure discrimination objective to learn generalized, layout-aware structural representations for depth decoding. Trained and evaluated on four diverse datasets, both individually and jointly, MaSS substantially outperforms prior self-supervised methods and is the first to consistently benefit from mixed-domain training. Zero-shot evaluation on eight additional datasets further demonstrates its remarkable generalizability, particularly on those entirely out-of-distribution scenes. The source code will be publicly available upon publication.

We introduce **AutoSpec**, a neural network framework for discovering iterative spectral algorithms for large-scale numerical linear algebra and numerical optimization. Our self-supervised models adapt to input operators using coarse spectral information (e.g., eigenvalue estimates and residual norms), and predict recurrence coefficients for computing or applying a matrix polynomial tailored to a downstream task. The effectiveness of **AutoSpec** relies on three ingredients: an architecture whose inference pass implements short, executable numerical linear algebra recurrences; efficient training on small synthetic problems with transfer to large-scale real-world operators; and task-defined objectives that enforce the desired approximation or preconditioning behavior across the range of spectral profiles represented in the training set. We apply **AutoSpec** to discovering algorithms for representative tasks on spd matrices: accelerating matrix function approximation; accelerating sparse linear solvers; and spectral filtering/preconditioning for eigenvalue computations. On real-world matrices, the learned procedures deliver orders-of-magnitude improvements in accuracy and/or reductions in iteration count, relative to basic baselines. We find clear connections to classical theory: the induced polynomials may exhibit equioscillation behavior characteristic of Chebyshev polynomial approximation.

It is widely agreed that when AI models assist decision-makers in high-stakes domains, they should communicate the confidence of their predictions. However, empirical evidence suggests that decision-makers often struggle to determine when to trust a prediction based solely on this communicated confidence. In this context, recent theoretical and empirical work suggests a positive correlation between the utility of AI-assisted decision-making and the degree of alignment between the AI confidence and the decision-makers' confidence in their own predictions. Crucially, these findings do not yet elucidate the extent to which this alignment influences the complexity of learning to make optimal decisions through repeated interactions. In this paper, we address this question in the canonical case of binary predictions and binary decisions. We first show that this problem is equivalent to a two-armed online contextual learning problem with full feedback, and establish a lower bound of $\Omega ( \sqrt{|\mathcal{H}| \cdot |\mathcal{B}| \cdot T} )$ on the expected regret any learner can attain, where $\mathcal{H}$ and $\mathcal{B}$ denote the sets of possible human and AI confidence values. We then demonstrate that, by leveraging the alignment between AI and human confidence, a learner can attain an expected regret of $\mathcal{O}(\sqrt{|\mathcal{H}| \cdot T\log T})$ and, in the regime where $\mathcal{B}$ is countable and $\sqrt{|\mathcal{H}|} = \mathcal{O}(\log T)$, a non-trivial generalization of the Dvoretzky-Kiefer-Wolfowitz inequality improves the regret bound to $\mathcal{O}(\sqrt{T\log T})$. Taken together, these results reveal that alignment can reduce the complexity of learning to make decisions with AI assistance. Experiments on real data from two different human subject studies where participants solve simple decision-making tasks assisted by AI models with different degrees of alignment complement our theoretical results.

Benders decomposition (BD) is a widely used solution approach for solving two-stage stochastic programs arising in real-world decision-making under uncertainty. However, it often suffers from slow convergence as the master problem grows with an increasing number of cuts. In this paper, we propose Reinforcement Learning for BD (RLBD), a framework that adaptively selects cuts using a neural network-based stochastic policy. The policy is trained using a policy gradient method via the REINFORCE algorithm. We evaluate the proposed approach on a two-stage stochastic electric vehicle charging station location problem and compare it with vanilla BD and LearnBD, a supervised learning approach that classifies cuts using a support vector machine. Numerical results demonstrate that RLBD achieves substantial improvements in computational efficiency and exhibits strong generalization to problems with similar structures but varying data inputs and decision variable dimensions.

Extremal combinatorics seeks rare mathematical structures that satisfy hard constraints while maximizing a global objective. Most existing neural systems for mathematical discovery train a model to generate complete configurations from scratch. We argue that this framing mismatches the dynamics of effective search: high-quality solutions are more effectively found by preserving a feasible partial core and conditionally regenerating the remaining uncertain objects. We formulate extremal structure discovery as object-level active-core completion: the model is trained to conditionally complete masked objects given a feasible partial core, capturing inter-object relations rather than fitting a distribution over complete configurations. We instantiate this with a full-attention Transformer over object-aligned slots, and propose Inpainting PatternBoost, a variant of PatternBoost. In each iteration, elite constructions from a top pool are partially destroyed by random or structured remasking; the model learns to reconstruct the missing objects, and their completions are reinserted into the pool. We evaluate on the no-three-in-line problem and maximum sum-of-radii circle packing in the unit square, under a single shared abstraction. On no-three-in-line, our framework recovers the optimal $2n$-point configuration for $n$ up to $18$, and reaches $2n-1$ points for $n=19, 20, 25$, outperforming prior neural baselines. On circle packing, where FlowBoost combines neural generation with gradient-based local search, our approach matches or exceeds FlowBoost when paired with the same local search; even without local search, it reaches an average of 88.1% of FlowBoost's reported score across instances.

Large language model (LLM)-based coding agents achieve strong results on controlled benchmarks yet routinely produce pull requests that real maintainers reject. The cause is not functional incorrectness but a lack of organicity: generated code ignores project conventions, duplicates internal APIs, and violates implicit architectural constraints. The latest repository snapshot is insufficient-it reveals the final codebase state, not the change patterns that shaped it. We make three contributions. (1) Paradigm: we propose next-commit prediction as a paradigm for learning agentic coding skill from a repository's own history-each historical commit is a self-supervised target whose oracle diff supplies dense supervision. (2) Evaluation: we establish Organicity Evaluation as a measurable objective for coding agents and contribute the Learning to Commit benchmark-5 curated open-source GitHub repositories under strict per-repository temporal splits-scoring patches along file localisation, internal API reuse, patch bloat, and code-style consistency. (3) Method: we instantiate the paradigm with the Learning to Commit framework, in which the agent performs supervised contrastive reflection: it blindly attempts each historical commit, contrasts its prediction against the oracle, and incrementally distils a reusable skill document that conditions subsequent generations. On the Learning to Commit benchmark and on SWE-bench Pro reframed under our temporal-split protocol, our framework consistently improves organicity on held-out future tasks and further lifts test-pass rates on SWE-bench Pro, narrowing the gap between benchmark success and real-world mergeability.

We study the problem of learning to bid when the bidder’s *value is dynamic*, i.e., when the current value depends on past outcomes. Specifically, we consider a bidder participating in repeated second-price auctions whose value depends on the time elapsed since their last successful bid, with auctions arriving in continuous time and only aggregated feedback revealed at the end of the horizon. Such a bidder must **(1)** balance the immediate benefit of winning the current auction against its impact on future values and **(2)** learn unknown environmental parameters. We derive regret bounds for a class of learning methods that combine plug-in estimators with a differential-equation characterization of the optimal policy, and show that a specific confidence bound algorithm learns the optimal policy with a near optimal regret of $\tilde{\mathcal{O}}(\log N)$ for piecewise linear primitives, and $\tilde{\mathcal{O}}(N^{1/3})$ for general, smooth primitives, achieving these regrets without explicit randomization. These theoretical results are supported by numerical experiments.

When auditing machine learning models, human experts sequentially inspect model predictions and explanations, gradually building a mental model of the model’s behavior to uncover errors. This manual process is critical for safe deployment in high-stakes domains but fundamentally limited in scalability. In this work, we propose a novel method to simulate this human auditing behavior for automating the auditing process. We formalize auditing as a sequential decision-making problem, modeled as a Markov decision process. Our state representation summarizes the auditor’s accumulated knowledge of model predictions and feature usage. Building on this formalization, we solve the MDP using reinforcement learning, yielding RLAuditor, which learns to select informative test samples to efficiently uncover model errors. We further connect our formalization to Theory of Mind, showing that our state mirrors how human auditors build mental models: its update rule is equivalent to Bayesian belief revision and it converges to the model’s behavioral pattern at an interpretable rate. We validate our approach on multiple ML models and datasets across different data modalities. In a human subject study, we demonstrate that RLAuditor helps human auditors produce more accurate audit reports in less time.

Edge-cloud collaboration between large language models (LLMs) and small language models (SLMs) offers a promising way to leverage the problem-solving capability of LLMs while keeping private data on device. Existing methods typically have the LLM generate guidance from an initial user query, which the on-device SLM combines with local private data to produce the final response. However, the LLM's general guidance often does not align with the user's personalized needs, limiting the utilization of the LLM's capabilities. It has been observed that the on-device SLM exhibits metacognitive capabilities, which allow it to recognize its limitations and adjust its reasoning process. Motivated by this, we transform the on-device SLM into a metacognitive inquirer agent that can assess its internal state and actively decide *when*, *what*, and *how* to query the LLM for personalized guidance. Specifically, we propose Active Inquiry via Metacognitive Actions (AIMA). AIMA learns an on-device policy that jointly optimizes the selection of high-level metacognitive actions and the action-conditioned query refinement. We further introduce a privacy-constrained rewriting mechanism that detects and eliminates sensitive information leakage in the refined query. Extensive experiments and theoretical analysis demonstrate that our method outperforms existing methods.

Fine-grained image retrieval (FGIR) typically relies on supervision from seen categories to learn discriminative embeddings for retrieving unseen categories. However, such supervision often biases retrieval models toward the semantics of seen categories rather than the underlying appearance characteristics that generalize across categories, thereby limiting retrieval performance on unseen categories. To tackle this, we propose GAPan, a Generative Appearance Prior alignment network that reformulates the learning objective from category prediction toward appearance modeling. Technically, GAPan treats retrieval features with an invertible density model based on normalizing flows. In the forward direction, the flow maps all instance features into a latent density space, where each seen category is modeled by a class-conditional Gaussian prior and optimized via exact likelihood estimation. This formulation preserves richer appearance details by leveraging the invertible property of the flows. In the reverse direction, samples from the high-density regions of these learned priors are mapped back to the feature space to produce appearance-aware anchors that reflect intra-category variation. These anchors supervise a prior-driven alignment objective that aligns retrieval embeddings with category-specific appearance distributions, thereby improving generalization to unseen categories. Evaluations demonstrate that our GAPan achieves state-of-the-art performance on both widely-used fine- and coarse-grained benchmarks.

All prior membership inference attacks for fine-tuned language models use hand-crafted heuristics (e.g., loss thresholding, Min-K\%, reference calibration), each bounded by the designer's intuition. We introduce the first transferable learned attack, enabled by the observation that fine-tuning any model on any corpus yields unlimited labeled data, since membership is known by construction. This removes the shadow model bottleneck and brings membership inference into the deep learning era: learning what matters rather than designing it, with generalization through training diversity and scale. We discover that fine-tuning language models produces an invariant signature of memorization detectable across architectural families and data domains. We train a membership inference classifier exclusively on transformer-based models. It transfers zero-shot to Mamba (state-space), RWKV-4 (linear attention), and RecurrentGemma (gated recurrence), achieving 0.963, 0.972, and 0.936 AUC respectively. Each evaluation combines an architecture and dataset never seen during training, yet all three exceed performance on held-out transformers (0.908 AUC). These four families share no computational mechanisms, their only commonality is gradient descent on cross-entropy loss. Even simple likelihood-based methods exhibit strong transfer, confirming the signature exists independently of the detection method. Our method, Learned Transfer MIA (LT-MIA), captures this signal most effectively by reframing membership inference as sequence classification over per-token distributional statistics. On transformers, LT-MIA achieves 2.8$\times$ higher true positive rate at 0.1\% false positive rate than the strongest baseline. The method also transfers to code (0.865 AUC) despite training only on natural language texts. Our results imply that leakage from memorization is intrinsic to cross-entropy training; architectural innovation within this paradigm did not escape our attack. Code and the trained classifier are provided.

Understanding how complex molecular systems transition between metastable states is a central challenge in computational chemistry and biophysics. A key tool for characterizing these transitions is the committor function, defined as the probability that a trajectory initiated from a given configuration reaches one state before another, which encodes detailed mechanistic information about transition pathways. First, we curate a public benchmark of empirically determined committor values for alanine dipeptide. Second, we develop a "committor-in-the-loop" framework that combines a self consistent loss function with weighted ensemble (WE) simulation: WE adaptive bins are biased by the learned committor, which is in turn trained on WE samples. Across a two-channel 2D system and alanine dipeptide settings, our WE-based framework consistently outperforms greedy baselines, demonstrating the importance of principled exploration. Together, the benchmark and framework provide both a standard for evaluating committor learning methods and a new method for studying rare transitions in high-dimensional molecular systems.

Recurrent neural networks are widely used in neuroscience and machine learning, but their weight matrices often appear dense and unstructured, defying clean interpretation. The Schur decomposition expresses any recurrent weight matrix in quasi-triangular form, allowing network dynamics to be interpreted as a latent circuit of "functionally feedforward" interactions between orthogonal modes. However, its application for interpreting recurrent weight matrices has been limited due to its well-known combinatorial non-uniqueness. We introduce SchurMO (Schur Manifold Optimization), a method for discovering structured Schur forms via Riemannian gradient descent on the manifold of orthogonal similarity transformations, enabling recovery of target motifs. Compared to a greedy baseline, SchurMO more reliably and more efficiently recovers ground-truth latent motifs (chain-like, banded, and modular structure) from noisy weight matrices. Applying SchurMO to nonlinear recurrent networks enables recovery of latent chain-like structure, providing a link between connectivity and the underlying dynamics of sequence production. Our results highlight SchurMO as a scalable, computationally efficient approach to discovering latent structure in the Schur forms of recurrent weights.

In-context reinforcement learning (ICRL) leverages the in-context learning capabilities of transformer models (TMs) to efficiently generalize to unseen sequential decision-making tasks without parameter updates. However, existing ICRL methods rely on explicit reward signals during pretraining, which limits their applicability when rewards are ambiguous, hard to specify, or costly to obtain. To overcome this limitation, we propose a new learning paradigm, \emph{In-Context Preference-based Reinforcement Learning} (ICPRL), in which both pretraining and deployment rely solely on preference feedback, eliminating the need for reward supervision. We study two variants that differ in the granularity of feedback: \emph{Immediate Preference-based RL} (I-PRL) with per-step preferences that generalize dueling bandits, and \emph{Trajectory Preference-based RL} (T-PRL) with trajectory-level comparisons. We first show that supervised pretraining, a standard approach in ICRL, remains effective under preference-only context datasets, demonstrating the feasibility of in-context solving unseen tasks using only preference signals. To further improve data efficiency, we introduce alternative preference-native frameworks for I-PRL and T-PRL that directly optimize TM policies from preference data without requiring reward signals nor optimal action labels. Experiments on dueling bandits, navigation, and continuous control tasks demonstrate that ICPRL enables strong in-context generalization to unseen tasks, achieving performance comparable to ICRL methods trained with full reward supervision.