NeurIPS 2026 Papers

Large multimodal models (LMMs) have advanced rapidly in perception and reasoning, yet it remains unclear whether they can discover visually grounded, physically feasible solutions in open-ended environments. We first introduce **MM-CreativityBench**, a benchmark for affordance-grounded creative tool use, where models must inspect scenes, entities, and parts to identify non-obvious object uses grounded in visual evidence. Our evaluation shows that current LMMs often fail to sustain grounded exploration: they overlook relevant entities, under-examine critical parts, or hallucinate unsupported attributes. To address these failures, we propose **affordance-grounded alignment**, framing creative tool use as a preference learning problem. Using Direct Preference Optimization and supervision from an affordance knowledge base, we train models to favor visually grounded attribute–affordance reasoning over hallucinated alternatives while improving exploration efficiency. Our results yield consistent gains in correct entity and part selection, and reduce hallucination and grounding errors. These findings position grounded creativity as a core capability for future multimodal agents: the ability to adapt to unfamiliar environments and solve problems beyond memorized patterns or surface-level plausibility, moving closer to real human-like intelligence.

Diffusion Transformers (DiTs) have achieved strong performance in visual generation, but their dense token processing leads to slow training convergence and high computational cost. Recent token routing methods such as TREAD accelerate training by allowing a random subset of tokens to bypass intermediate layers. It can be expected that carefully tuning the routing ratios over time steps or layers can achieve more significant accelerations. However, the theoretical mechanisms underlying stochastic routing remain underexplored, making the tuning of routing ratios intractable. In this paper, we first theoretically show that stochastic token routing can be interpreted as an implicit route-sensitivity regularizer. Under this view, the routing ratio determines the strength of the induced regularization: larger routing ratios introduce stronger route-induced perturbations and therefore stronger regularization pressure. This further motivates a noise-conditioned routing strategy. In diffusion training, samples at larger timesteps contain heavier noise and often provide less stable gradient signals, suggesting that they may require stronger regularization. We therefore assign larger routing ratios to higher-noise samples, allowing the routing-induced regularization strength to adapt to the noise condition. Based on this insight, we propose \emph{Noise-Aware Routing}, which partitions samples within each mini-batch into low-, middle-, and high-noise regions and assigns progressively larger routing ratios to them. This simple noise-conditioned allocation synergistically bridges computational savings with enhanced model priors. Extensive experiments on class-conditional ImageNet generation show that our method accelerates training convergence and improves generation quality. Noise-Aware Routing achieves up to \textbf{15.6}$\times$ and \textbf{20}$\times$ convergence speedups over DiT baselines—and outperforms TREAD—on ImageNet-256 and ImageNet-512, respectively, measured by the training iterations required to match our 400K-step FID. In the guided setting, our method achieves an FID of \(2.67\) with a substantially shorter training iterations without architectural changes.

LLM-as-a-Judge systems supply the reward signal in modern RLHF and RLVR pipelines, but their binary verdict reduces to a single linear readout on one hidden state. We show this readout is shallow enough that short, low-perplexity tokens flip the verdict from "No" to "Yes". These tokens are sampled from the judge's own next-token distribution at the response position, with no manual seed set and no gradient-based optimization. Our procedure, AdvJudge-Zero, reaches $>$90\% ensemble false-positive rate on 22 of 24 (model, dataset) cells across six Qwen, Llama, and Gemma judges, versus 54-72\% for the prior curated 10-token benchmark, and the discovered surface transfers cross-format to a 70B scalar reward model. The same discovered pool enables a defense: a LoRA fine-tune stratified by a 9-class mechanism taxonomy hardens cross-family generalization where naive sampling on the same pool fails, with mechanism breadth rather than pool size carrying the gain. Under GRPO training, the hardened judge eliminates the reward-collapse failures (false-positive spikes and length collapse) we observe in the unhardened baseline on both MATH and GSM8K at ten seeds per condition. The discovered pool, the mechanism taxonomy, and per-prompt flip records will be released under responsible disclosure.

Temporally extended exploration via graph Laplacian-based options is a promising approach to sparse-reward reinforcement learning (RL), but existing methods either do not explicitly target novelty or fail to scale to pixel-based domains under function approximation. Novel Exploration via Orthogonality (NEO) addresses the first issue by constructing options that navigate from highly visited regions toward less visited ones, yet prior results were limited to settings where exact eigenvectors can be computed. We present a scalable extension of NEO to pixel-based domains, built on three contributions. First, we use a novelty-weighted continuous Laplacian graph-drawing objective, which enables RL with continuous observations. Second, we embed the resulting eigen-potential options within a hierarchical reinforcement learning framework, enabling coherent temporally extended behavior. Third, we observe that learned eigen-potential rewards are directional but locally unreliable under online approximation; we therefore augment each option reward with a novelty bonus, a novel design idea that proves essential for stabilizing option learning while preserving novelty-directed exploration. Together, these contributions yield stronger and more persistent exploration, enabling longer option rollouts and better access to hard-to-reach novel states. Empirically, our method significantly outperforms both the prior scalable Laplacian-option baseline and a direct extension of NEO on sparse-reward benchmarks under a fixed budget. On Montezuma's Revenge, our best variant achieves approximately 1.8x higher return than both baselines. On Venture, both baselines yield returns near zero, whereas our method achieves a return of 1135. Across seven hard ProcGen games, our method achieves approximately 3.5x and 5.6x higher aggregate normalized return than the two baselines, respectively.

Cooperative multi-agent reinforcement learning (MARL) often fails under exogenous variability such as shifts in opponents or environment regimes that cannot be controlled by the team but alter the transition dynamics. We formalize this challenge as Exogenous Dec-POMDP (ED-POMDP), which decomposes the global state into endogenous variables controllable by the team and exogenous variables beyond its control. This formulation exposes delayed influence: While actions do not directly determine exogenous transitions, they can shape them through induced changes in endogenous state. Based on this, we propose LEICA, a CTDE-compatible algorithm that learns history-conditioned endogenous and exogenous context representations and shapes policy updates using influence-weighted intrinsic rewards. Across SMAX benchmarks with opponent strategy shifts, LEICA consistently improves both training performance and generalization to unseen opponents' strategies over existing baselines, supporting the usefulness of influence-weighted shaping under exogenous train-test regime shifts.

We study visual persistence in autoregressive video world models: the Key-Value (KV) cache accumulates a growing visual memory, but once rollouts extend beyond the training horizon, the model can no longer reliably address stored content. A standard fix compresses the cache into a fixed-size memory that retains recent context while summarizing the distant past. However, we show compression alone cannot restore recall: past this horizon, temporal positional encodings go out-of-distribution, so the model cannot reliably retrieve stored history regardless of content. On the evaluated architecture, content compression with out-of-distribution positions yields results identical to a fixed-size sliding window over recent KV entries. We propose WorldTrace, a training-free framework that keeps compressed memory addressable by assigning each slot a fixed, in-distribution position relative to the current frame. With addressable memory, we explore two retention approaches: WorldTrace-Field (coherence-oriented) aggregates history in a rotation-invariant space, improving TempSSIM by +15.5% while reducing scene drift; WorldTrace-LandMark (recall-oriented) stores verbatim scene traces at detected boundaries. The recall-oriented variant sustains scene reconstruction over long rollouts, with stronger long-range recall and smooth temporal coherence.

We study zeroth-order optimization of non-convex functions with the aid of directional hints, which are cheap but potentially inaccurate approximations of the true gradient direction, given by linear subspaces at each iteration. To leverage these hints adaptively while maintaining robustness to their quality, we introduce Control-Variate Zeroth-Order Descent (CV-ZOD), a new framework that refines the classical zeroth-order gradient estimator with a control variate that can be set based on the directional hints. We first show that the oracle algorithm that optimally sets the reference vector and step size at each iteration achieves a convergence rate that interpolates between the first-order $O(1/T)$ rate and the zeroth-order $O(d/T)$ rate, depending on the quality of the hints along the trajectory. We then develop a practical variant of CV-ZOD that achieves the same oracle guarantee up to logarithmic factors, without any prior knowledge of the hint quality. We validate the method empirically on simulation-based scientific optimization tasks, demonstrating sustained progress on nonconvex landscapes where zeroth-order descent is slower and existing guided methods stall as guidance deteriorates.

Real-world tasks rely on information from multiple sensory sources, motivating multimodal learning as a core paradigm in modern machine learning. However, multimodal learning suffers from modality laziness, where one modality suppresses the contribution of the other modalities by dominating the training process. Recent studies of proposing various metrics to identify lazy modalities and manipulate optimizations to mitigate this imbalance, facing two main problems: 1) The identification of lazy modality is incomplete and overly sharp; 2) The dominated modality can be lagged by under-optimized lazy modalities. To address these issues, we propose Adaptive-Margin Masking and Restoration (AMRe), which introduces adaptive-margin modality identification and restoration optimization to balance dominated and lazy modalities. Experimental results show that AMRe consistently outperforms competitive baselines and several state-of-the-art methods by achieving significant improvement on standard multimodal benchmarks. The codes are released in https://anonymous.4open.science/r/AMRe-5E69.

A major bottleneck in characterizing the failure modes of generative AI systems is the cost and time of annotation and evaluation. Consequently, adaptive testing paradigms have gained popularity, where one opportunistically decides which cases and how many to annotate based on past results. While this framework is highly practical, its extreme flexibility makes it difficult to draw statistically rigorous conclusions, as it violates classical assumptions: the number of observations is typically limited (often 10 to 50 cases) and decisions regarding sampling and stopping are made in the midst of data collection rather than based a pre-specified rule. To characterize what statistical inferences can be drawn from highly adaptive audits, we introduce a hypothesis testing framework from two 'dueling' perspectives: (i) the model's null that asserts there is no failure mode with performance below a target threshold versus (ii) the auditor's null that asserts they have a sampling strategy that will uncover a failure mode. Leveraging Safe Anytime-Valid Inference (SAVI), we formalize the auditor as conducting 'testing by betting', which translates into simultaneous e-processes for testing the dueling null hypotheses. Furthermore, *if* the auditor is sufficiently powerful, we prove that these two hypotheses are asymptotically inverses of each other, in that passage of a stringent audit does in fact certify the AI system as being globally robust. Empirically, we demonstrate that our proposed testing procedures maintain anytime-valid type-I error control, outperform pre-specified testing methods, and can reach statistically rigorous conclusions sometimes with as few as 20 observations.

Evolution strategies (ES) have recently emerged as a promising alternative to gradient-based reinforcement learning for fine-tuning large language models, but classical zeroth-order optimization theory does not systematically account for the empirical observations driving this success. We formulate the continuous-time limit of Z-score-normalized ES as a unified stochastic differential equation and identify three sequential dynamical regimes characterized by the dominant gradient contribution to the population reward variance; under a basin-local low-rank Hessian assumption, this framework systematically explains (1) the information-geometric optimality of the Z-score heuristic, (2) the dimensionality paradox of sample-efficient ES with populations far smaller than the parameter dimension, (3) the same-task rise-then-decay of the training reward, and (4) catastrophic forgetting of pre-training capabilities. Synthetic experiments and direct verification on Qwen2.5-{0.5B,3B,7B}-Instruct confirm these predictions, matching the parameter drift reported by Abdi et al. (2026) within $\sim 1.5\\%$.

The target-charging technique (TCT), a generalization of the sparse vector technique, monitors whether query answers deviate beyond acceptable bounds while incurring privacy costs only upon significant deviations. We develop a more mature theory of TCT through three contributions. First, we generalize TCT and replace prior simulation-based analyses with a direct stochastic domination argument, yielding a tighter bound and a modular proof structure that affords flexibility in the choice of privacy representation for the base mechanism as well as the composition accountant. Second, we provide an anytime-valid TCT guarantee that integrates with approximate Rényi DP filters. Third, we enable individual-level privacy filters. Together, these provide the first end-to-end TCT construction compatible with privacy filters and individual-level accounting, with practical relevance demonstrated through benchmarks inspired by the impending world-scale deployment of the W3C Attribution standard, which is fundamentally rooted in individual filters.

Modern deep reinforcement learning algorithms often store past sensory observations in a buffer for later replay and learning. However, this process is biologically implausible; animals do not store raw sensory observations. Instead, they store their imperfect understanding of the world and the events in it. An alternative to storing all past sensory observations is to learn directly from the observations as they come in and store the seemingly important parts of the incoming data. Eligibility trace algorithms in reinforcement learning (RL) sweep back updates for multiple steps, improving efficiency and providing update rules that do not rely on storing large data buffers. These algorithms were common and effectively used in the linear setting, but have not been widely adopted in the deep RL setting. Many questions remain open to make these algorithms more effective, even as basic as how to incorporate vector step-sizes. We first highlight how adaptive vector step-sizes need to be incorporated into the eligibility trace and then introduce a new vector step-size algorithm that is more amenable to the forward-backward view equivalence needed to derive an update with eligibility traces. We show that our new algorithms improve over other streaming algorithms across Mujoco and MinAtar environments, and are comparable to methods with replay buffers such as PPO.

We present the Adaptive Scheduling Pipeline (ASP), a bounded-lag asynchronous reinforcement learning (RL) architecture for post-training large language models (LLMs). Rather than only disaggregating generation and training, the ASP combines two design ideas. First, a queue-aware scheduling policy that jointly reduces three sources of rollout underutilization —padding bubbles from intra-minibatch length heterogeneity, skewness bubbles from long-tail outputs, and sorting bubbles from queue instability induced by length-sorted dispatching—through a single offline preprocessing pass. This enables efficient multi-instance rollout that overcomes the early scaling plateau of single-instance inference while maintaining stable queue dynamics. Second, a bounded-lag asynchronous architecture decouples rollout from training while keeping policy lag observable and small enough to preserve stable learning in our experiments. We also provide an explicit weight-flow path that converts trainer snapshots into inference-ready weights without re-coupling generation and training. Experiments show that ASP improves end-to-end throughput by up to $3.22\times$, reduces iteration time by up to 76\%, scales more robustly with additional generation resources than the MindSpeed-RL and verl baselines, and shows no visible convergence degradation. ASP has also been used in an internal telecom-domain RL post-training workflow, reducing the reported training period from more than three months to about two months and producing a domain-enhanced model for knowledge answering and tool calling.

Temporal action segmentation (TAS) is commonly trained on large banks of pre-extracted frame features rather than raw RGB video, but storing and replaying those features becomes a practical bottleneck for long-form video, especially in incremental TAS where past-task data must remain available to mitigate catastrophic forgetting. Recent TAS replay and condensation methods reduce this burden, but they rely on dense frame-wise annotations and allocate storage through fixed segment-level structures, limiting their flexibility for long procedural videos whose frame complexity varies substantially over time. We introduce Adaptive Residual-Quantized Variational AutoEncoder (ARQ-VAE), a compact and label-free memory representation for TAS features that learns a discrete residual quantization of frame-level video features and then adaptively truncates the residual chain on a per-frame basis, retaining only the residual prefix that best reconstructs each frame feature. This yields a practical compact feature memory that does not depend on frame-wise labels and can therefore support multiple TAS settings with the same stored representation, while a lightweight fine-tuning stage further improves reconstruction quality for the truncated codes that are actually retained. Across standard TAS benchmarks, ARQ-VAE achieves a substantially improved storage--performance trade-off over prior replay and condensation baselines. On Breakfast, our final representation reduces training-set storage from 28 GB for the original I3D feature bank to 10 MB while retaining strong downstream segmentation performance; with ASFormer, it achieves 71.5 Edit and 51.2 F1@50 in the supervised setting. We further show that the same label-free memory remains effective for incremental and unsupervised TAS, and transfers to an alternative pretrained feature space based on DINOv3. We position ARQ-VAE as a practical memory representation for TAS, and a promising direction for broader dataset condensation applications.

Adaptive Random Forests (ARF) are among the most effective ensemble methods for learning from non-stationary data streams, yet their success relies on heuristic drift detectors and reset rules that lack formal guarantees and require careful tuning. We revisit ARF from an online learning perspective and show that its design decomposes into two fundamental questions: how to aggregate a dynamically evolving set of trees, and how to update the pool of trees over time. We address aggregation using parameter-free online learning methods with strongly adaptive regret, enabling the ensemble to track the best tree mixture. For pool updates, we replace detector-driven resets with deterministic multi-scale scheduling based on geometric lifetimes, combined with incumbent--challenger replacement governed by anytime-valid statistical tests. This design yields theoretical guarantees: global control of false replacements, bounds on promotion delay under post-change advantage, and adaptivity under piece-wise stationary environment. Empirically, the resulting Online RF consistently outperforms ARF and is competitive with broader streaming baselines across several classification and regression benchmarks.

Gaussian process (GP) bandits provide a powerful framework for performing blackbox optimization of unknown functions. The characteristics of the unknown function depend heavily on the assumed GP prior. Most work in the literature assume that this prior is known but in practice this seldom holds. Instead, practitioners often rely on maximum likelihood estimation to select the hyperparameters of the prior - which lacks theoretical guarantees. In this work, we study two algorithms for joint prior selection and regret minimization in GP bandits based on GP Thompson sampling (GP-TS): Prior-Elimination GP-TS (PE-GP-TS) that disqualifies priors with poor predictive performance, and HyperPrior GP-TS (HP-GP-TS) that utilizes a bi-level Thompson sampling scheme. We theoretically analyze the algorithms and establish a sublinear regret bound for HP-GP-TS. In addition, we demonstrate the effectiveness of these algorithms compared to the alternatives through extensive experiments with synthetic and real-world data.

We study $\left(\epsilon,\delta\right)$-differentially private algorithms for the problem of approximately computing the top singular vector of a matrix $A\in\mathbb{R}^{n\times d}$ where each row of $A$ is a datapoint in $\mathbb{R}^{d}$. Following Dwork-Talwar-Thakurta-Zhang (STOC 2014), we consider the privacy model where neighboring inputs differ by one single row. We give a novel algorithm that achieves beyond-worst-case guarantees for input matrices with low coherence, which is a structural property of matrices in many applications, including but not limited to i.i.d. data. Our algorithm contributes to the extensive literature on private power iteration methods, where we introduce a new filtering technique which adapts to this coherence parameter. Our work departs from and complements the work by Hardt-Roth (STOC 2013) which achieves beyond-worst-case guarantees for the more restrictive privacy model where neighboring inputs differ in one single entry by at most 1.

Graph contrastive learning (GCL) learns node and graph representations by contrasting multiple views of the same graph. Existing methods predominantly rely on a small set of fixed, handcrafted views---typically a local and a global perspective---which fundamentally constrains their capacity to capture multi-scale structural patterns. We present an augmentation-free, multi-view GCL framework grounded in fractional-order continuous dynamics. By systematically varying the fractional derivative order $\alpha\in(0,1]$, our encoders produce a continuous spectrum of semantically distinct views: small values of $\alpha$ induce localized, memory-attenuated feature propagation, whereas values approaching $1$ recover broader, global aggregation. Crucially, we treat $\alpha$ as a learnable parameter, enabling the model to automatically identify informative views in a data-driven manner, without resorting to manual view engineering. Extensive experiments on standard node- and graph-level benchmarks demonstrate that the resulting representations are more robust and expressive, consistently outperforming state-of-the-art GCL baselines.

Reliable prediction of material properties from atomistic structures requires machine learning models that respect SE(3) symmetry. Frame-based methods address this challenge by constructing equivariant coordinate systems that align SE(3)-equivalent structures into unified representations. While attractive for their flexibility and efficiency, existing frame-based methods face two major limitations: a single frame type is often insufficient for heterogeneous atomic environments, and frame constructions can become unstable near highly symmetric or degenerate configurations. To address these issues, we propose the Multi-Frame Adaptive Network (MFAN), which leverages multiple frame types and introduces a local adaptive mechanism to combine them according to atomic environments, enabling different geometric references across atoms and complementary local-global information for each atom. Building on this multi-frame architecture, we further incorporate a frame-quality-aware weighting scheme that downweights unreliable frames before degeneration, thereby improving the continuity of the resulting representation. Experiments on crystal property prediction benchmarks demonstrate the leading performance of MFAN, with consistent gains from adaptive multi-frame learning. Additional force prediction experiments show improved robustness in continuity-sensitive settings.