NeurIPS 2026 Papers

As Embodied Agents System (EAS) move into physical domains such as autonomous navigation and household assistance, reliable execution becomes critical for physical task completion and collaboration trust. Recent EAS reliability evaluation works typically focus on outcome-centric metrics as success rate or safety-only scores to assess an agent's performance, which collapse diverse execution trajectories into coarse outcomes. Therefore they ignore a critical property of EAS -- which we define as the **Resilience** -- that reflects how EASs recover, stabilize, and extend under perturbations and across iterative updates. The lack of resilience is particularly critical in open-world environments due to continuous unexpected disruptions, thus directly affecting the quality of EAS deployment. To address this problem, we gain insight from the resilience-engineering concepts to EAS groundings and propose a novel resilience evaluation framework that can be flexibly applied to any EAS. Specifically, we define the first comprehensive resilience metrics suite for EASs system that exposes *Rebound*, *Stability*, and *Graceful Extensibility* across embodied tasks execution, providing a practical grounding for EAS resilience analysis. We further implement the resilience evaluation layer that transforms execution evidence into assessments for comparison, diagnosis and optimization. Across 400 household tasks with 10 representative EAS baselines, our evaluation reveals process-level distinction hidden by outcome metrics, including recovery cost differences among successful episodes ($\Delta C_{\mathrm{rec}}=25.2$), increased instability under semantic perturbations, and task-family-specific degradation under stress. Metrics-guided optimizations reduce recovery cost by 42.94%, increase stability by 19.87%, and improve graceful extensibility completion by 10.08%, showing the diagnostic effect of resilience evaluation. Beyond these targeted gains, our results reveal trade-offs among resilience characteristics, suggesting that a resilient EAS construction should be evaluated and configured according to deployment-specific requirements. Our code and data are available at .

Asynchronous stochastic gradient descent (ASGD) is a standard way to exploit heterogeneous compute resources in distributed learning: instead of forcing fast workers to wait for slow ones, the server updates the model whenever a gradient arrives. Vanilla ASGD applies each arriving gradient with the same weight. When local data distributions are heterogeneous, this becomes problematic: faster workers contribute more updates, and we show theoretically that the method is biased toward a frequency-weighted average of the local objectives rather than the desired global objective. Existing remedies typically move away from the simple ASGD template by introducing gathering phases, buffering, or extra memory. We show that this is unnecessary. Keeping the standard ASGD mechanism, we recover the correct objective by rescaling worker-specific stepsizes in proportion to their computation times, so that each worker contributes the same aggregate learning rate over a cycle. In the non-convex setting, under smoothness and bounded heterogeneity assumptions, we prove that the resulting method, Rescaled ASGD, converges to stationary points of the correct global objective in the fixed-computation model. Its time complexity matches the known lower bound in the leading term, while the effects of staleness and data heterogeneity appear only in lower-order terms. Experiments confirm that the method converges to the correct objective and is competitive with state-of-the-art baselines.

Upgrading a deployed face-recognition model usually makes new query embeddings incompatible with the legacy gallery, forcing costly backfilling of stored embeddings. We show that this incompatibility is often much simpler than it appears: independently trained angular-margin face models exhibit representation rigidity, where their embedding spaces are well approximated by a single global orthogonal transformation. We formalize this phenomenon as Procrustes rigidity, derive a finite-calibration recoverability bound, and estimate the orthogonal adapter from a small paired calibration set via closed-form Procrustes alignment. The resulting deployment protocol is backfill-free: transport upgraded queries into the legacy coordinate system, then reuse the existing gallery without re-embedding stored identities. Across 45 models spanning three architectures, two losses, three data scales, and five seeds, we find strong rigidity within architecture families, non-monotonic scaling with data size, and robust cross-model retrieval. On the 1.6M-image IFRT benchmark, Procrustes-aligned queries recover at least $99.8\%$ of mono-model Rank-1 accuracy for same-family upgrades, while lightweight residual correction substantially narrows the cross-architecture gap.

Repository-level documentation generation is vital for human comprehension and autonomous agents, yet existing approaches fail to align high-level architectural intent with granular implementation details, resulting in documentation that lacks both global coherence and local fidelity. To address this, we propose RepoScope, a repository documentation generation framework that achieves alignment via two synergistic innovations. Specifically, RepoScope employs Physico-Logical Partitioning by jointly modeling directory structures and call graphs, effectively bridging the gap between physical layouts and logical dependencies while creating coherent documentation units. Within this partitioned space, we introduce SAIL (Seeded Architectural Intent Lift), a mechanism that projects architectural intent top-down to steer granular analysis, which is then revised and lifted bottom-up via code evidence to ensure global-local consistency. Our experimental results on CodeWikiBench demonstrate state-of-the-art performance in aligning with human-authored documentation. Furthermore, the validation on real-world industrial systems shows that RepoScope yields superior correctness and completeness in empowering autonomous agents for CLI extraction compared to existing baselines.

The recent empirical success of non-Euclidean optimizers, such as Muon (Jordan et al., 2024) and Shampoo (Gupta et al., 2018), has transformed the training of Large Language Models (LLMs) by exploiting the geometry of matrix spaces. Despite this progress, the theoretical understanding of stochastic optimization in general normed spaces remains limited, particularly in distributed settings. In this work, we present a novel analysis of Minibatch Stochastic Gradient Descent (SGD) in non-Euclidean geometries. Leveraging the framework of $\\kappa$-regular normed spaces (Juditsky & Nemirovski, 2008) and a light-tail noise model defined directly in the dual norm, we derive high-probability oracle complexity bounds for smooth non-convex optimization. Crucially, our analysis reveals that measuring stochastic variance within the intrinsic non-Euclidean geometry avoids the suboptimal dimension-dependent factors inherent in Euclidean norm equivalence. Building on this framework, we introduce Rennala-NSGD, a semi-asynchronous distributed algorithm, and establish, to the best of our knowledge, the first wall-clock time complexity guarantee for asynchronous non-Euclidean stochastic optimization.

Clustering and vector quantization are core primitives for representation learning and large-scale retrieval, yet widely used methods often produce clusters or codewords of highly uneven quality, especially in noisy or heterogeneous data. In this paper, we propose a distortion-constrained optimal transport (DCOT) framework that explicitly enforces per-cluster bounds on the average assignment distortion, defined as the expected distance between data points and their assigned representatives, ensuring more consistent and reliable representations. Our formulation couples an optimal transport assignment objective with cluster-level distortion constraints to control cluster quality. To solve this constrained problem, we develop an efficient alternating optimization algorithm that iteratively updates transport plans, representatives (codewords), and dual variables. We further provide theoretical guarantees including entropic consistency, coordinate descent, and sublinear convergence. Importantly, DCOT serves as a plug-in assignment module that can be seamlessly integrated into a wide range of vector quantization methods. Replacing the assignment step with DCOT consistently improves reconstruction quality across diverse architectures and training settings, demonstrating strong robustness and general applicability.

Direct Preference Optimization (DPO) has recently been applied to medical image segmentation, since comparing candidate masks is substantially cheaper than producing new dense pixel-level annotations, making pairwise preference signals an appealing source of supervision. However, existing work derives preferences from ground-truth masks, leaving open whether the approach is sound under the imperfect feedback available in practice. In our analysis of this reformulation, we identify that candidate masks tend to agree over most of the image and differ only in localized regions, yet standard DPO averages the preference signal over the full spatial domain. This couples update strength to disagreement area, diluting small correct refinements and amplifying large misranked differences. Crucially, this bias even degrades performance even under oracle preferences derived from ground truth. We propose Region-Normalized DPO (RN-DPO), which normalizes the likelihood ratio over the disagreement region between candidate masks, removing this coupling. We further provide a systematic empirical study of preference-based segmentation fine-tuning under controlled noisy judges, analyzing the effects of mining strategies, judge types, and reliability regimes. Across two medical segmentation benchmarks, multiple judge configurations, and two backbone architectures, RN-DPO consistently improves over vanilla DPO and robust DPO variants, with gains persisting under noiseless oracle preferences.

Existing LLM-based policy optimizers see only scalar rewards: that a policy scored 0.45, but not whether the agent got stuck in a loop, fell into a hole on the third step, or performed well on 19 out of 20 rollouts and failed catastrophically on one. We propose Reflective Prompted Policy Optimization (R2PO), a two-stage LLM framework for policy search over compact policy classes that augments scalar reward feedback with trajectory-level behavioral evidence. A Search-LLM acts as a global policy optimizer and proposes candidate policy parameters; the environment executes them; a Critic-LLM then inspects the resulting rollouts and proposes targeted parameter revisions grounded in observed states, actions, and rewards. Across ten environments, ablations show R2PO's gains arise from a design that explicitly separates global search from behavior-grounded revision and uses selection to filter high-variance edits. We further identify a dominant failure mode, salience bias. When presented with multiple rollouts, the Critic-LLM fixates on improving a single failure even when most trajectories succeed. In a three-trajectory variant, where the Critic-LLM is shown the best, worst, and median rollout from each evaluation, this behavior explains 76.6\% of regressions on CartPole. R2PO mitigates this by reasoning over aggregate rollout statistics, median-trajectory selection, and a revision rule. Using a relatively small open-weight 20B-parameter model, R2PO achieves the highest mean best reward across all ten environments, while reaching near-optimal performance substantially earlier in training (e.g., near-maximum CartPole reward within $\sim$500 episodes), and training far more stably than both deep RL and prior LLM-based methods. Together, these results show that treating trajectories as first-class in-context evidence, rather than background artifacts reduced to scalar returns, fundamentally changes how even comparatively small LLMs can search over policy spaces, enabling them to learn faster, diagnose more precisely, and reliably improve external controllers rather than tune them by trial and error.

Ensuring high returns while providing strong safety guarantees in offline imitation learning (IL) is a fundamental challenge: existing approaches either rely on explicitly specified cost functions, which are difficult to design for complex real-world safety constraints, or require expert demonstrations that are both safe and high-performing, which are often too costly to collect. We introduce *dual-expert imitation learning*, which leverages two complementary data sources: (i) safety-compliant demonstrations that faithfully follow safety guidelines but may be suboptimal, and (ii) performance-oriented demonstrations that achieve high returns but may violate safety guidelines. To address this problem, we propose DexDICE (**D**ual-**ex**pert offline imitation learning via stationary **DI**stribution **C**orrection **E**stimation). Instead of conventional distribution constraints, DexDICE combines DICE-style offline learning with safe support constraints, enabling agents to exploit high-return behaviors while remaining within safety-compliant regions. We evaluate DexDICE on a re-curated DSRL benchmark and validate it on a real-world mobile-robot navigation task trained directly from human teleoperation demonstrations, where it significantly outperforms offline IL baselines and remains robust to data scarcity and quality degradation.

Humans rapidly learn abstract knowledge when encountering novel environments and flexibly deploy this knowledge to guide efficient and intelligent action. Can modern AI systems learn and plan in a similar way? We study this question using a dataset of complex human gameplay with concurrent fMRI recordings, in which participants learn novel video games that require rule discovery, hypothesis revision, and multi-step planning. We jointly evaluate models by their ability to play the games, match human learning behavior, and predict brain activity during the same task, comparing a suite of frontier Large Reasoning Models (LRMs) against model-free and model-based deep reinforcement learning agents and a Bayesian theory-based agent. We find that frontier LRMs most closely match human behavioral patterns during game discovery and predict brain activity an order of magnitude better than both reinforcement learning alternatives across cortical and subcortical regions, with effects robust to permutation controls. Through targeted manipulations, we further show that brain alignment reflects the model's in-context representation of the game state rather than its downstream planning or reasoning. Our results establish LRMs as compelling computational accounts of human learning and decision making in complex, naturalistic environments.

AI systems are increasingly deployed in conversational settings where users may be uncertain whether they are speaking with a human or an AI. Despite mounting regulatory attention to this known safety risk, existing evaluations of AI disclosure are typically English-only, based on machine-generated questions, and restricted to text. We present RealityTest to comprehensively test whether AI systems disclose their identity when asked. The benchmark is the first large-scale multimodal and multilingual evaluation, grounded in human data on how people actually encounter and question AI identity in the real-world. Alongside the benchmark, we release the underlying dataset of 3,152 identity-probing queries collected from $\textasciitilde$750 participants across 49 countries and five languages, in text and speech scenarios. We find that only 31\% of people ask about identity directly, and that the questions people ask are far more diverse than machine-generated queries. We test 17 text and 6 speech models, and find substantial variation in disclosure behaviour. However, a single suppression instruction reduces disclosure rates to below 30\%, even in the best-performing models. Validating our investment in diverse, human-grounded evaluation data, we find that how the question is phrased and the context of the conversation matter more for disclosure than which model is being tested. Safety evaluations built on narrow or synthetic query sets risk mischaracterising how models behave in realistic deployment settings.

Pre-training large language models on massive GPU clusters has made hardware faults routine rather than rare, driving the need for resilient training systems. Yet existing frameworks either focus on specific parallelism schemes or risk drifting away from a failure-free training trajectory. We propose ReCoVer, a resilient LLM pre-training system that upholds a single invariant: each iteration keeps the number of microbatches constant, ensuring per-iteration gradients remain stochastically equivalent to a failure-free run. The framework is organized as three decoupled protocol layers: (1) Fault-tolerant collectives that isolate faults from propagating across replicas; (2) in-step fine-grained recovery that preserves intra-iteration progress and prevents gradient corruption; (3) versatile-workload policy that dynamically redistributes microbatch quotas across the survivors. The design is parallelism-agnostic, integrating directly with both 3D parallelism and Hybrid Sharded Data Parallel (HSDP) as a drop-in substrate. We evaluate our implementation on end-to-end pre-training tasks for up to 512 GPUs, ReCoVer successfully preserves the training trajectory from a failure-free reference despite of 256 GPUs lost spread across the run. For comparison with checkpoint-and-restart baselines, ReCoVer demonstrates $2.23\times$ higher effective throughput after successive failures. This advantage results in ReCoVer processing 74.9% more tokens at 234 GPU-hours, with the gap widening as the training prolongs.

Discrete autoregressive (AR) text-to-image (T2I) models adopt a two-stage paradigm in which a VQ tokenizer maps images to discrete codes and an AR policy models their distribution. Current post-training methods optimize only the AR policy while keeping the VQ decoder frozen. We show that this practice introduces Latent Covariate Shift: as the policy evolves, its token distribution progressively diverges from the ground-truth distribution on which the decoder was trained, such that reward scores improve while decoded image quality degrades. To address this mismatch, we propose **RankE**, the first end-to-end post-training framework for discrete T2I generation. Rather than optimizing the policy against a fixed decoder, RankE co-evolves the policy and decoder through alternating optimization: the policy is refined via group-relative preference optimization, while the decoder is jointly adapted via reward-aware adversarial training. This co-evolution ends the fidelity--alignment trade-off that plagues frozen-decoder approaches: on LlamaGen-XL (775M), standard RL improves CLIP but degrades FID, whereas RankE simultaneously improves both (FID 15.21, CLIP 33.76 on MS-COCO 30K). Consistent joint gains on Janus-Pro (1B) confirm that decoder co-evolution reliably converts reward optimization into pixel-space quality improvements.

V2X collaborative perception improves single-vehicle perception by aggregating multi-view features from collaborative agents. However, existing methods typically rely on simple spatial warping to align the collaborative features, which can cause semantic misalignment during fusion. Meanwhile, these deterministic warping-based methods are highly sensitive to the positional noises. In this paper, we propose a plug-in module called Robust Semantic and Spatial Aligner (RSSA), which performs semantic and probabilistic spatial transformations for robust feature alignment. Specifically, a Semantic Transformation (SemT) block modulates feature channels based on the observed relative pose to correct semantic inconsistencies. Subsequently, a Probabilistic Spatial Transformation (P-SpaT) block performs probabilistic warping by sampling from the Gaussian-modeled relative pose space. A Relative Transformation Augmentation (RT-Aug) strategy is further introduced to augment the relative poses and facilitate the training of the RSSA. Extensive experiments on both simulated and real-world datasets demonstrate that RSSA improves the state-of-the-art methods by a large margin under the configuration of both clean and noisy positions.

Online reinforcement learning (RL) has achieved significant success in decision-making tasks, but modern online RL remains highly sensitive to the quality of training experience. Understanding the role of training data in online RL is challenging because the training distribution evolves together with the policy, causing the influence of the training data to propagate across future optimization and future data collection. Existing data-attribution methods for online RL primarily focus on local influence within a single training round, overlooking the non-local data influence across multiple training rounds. In this work, we formalize data attribution in online RL through a trajectory-level leave-one-out quantity, RL-LOO, which measures how one training sample influences a downstream target after subsequent online updates have taken place. We then derive RL-Inf, a first-order estimator that propagates data influence through both optimization effects and policy-induced sampling effects, and provide theoretical guarantees on its approximation accuracy under smoothness assumptions. Our analysis further disentangles these two effects and shows both theoretically and empirically that the full RL-Inf estimator can often be well approximated using the optimization effect alone. Building on this observation, we develop RINSE, a practical data filtering method for online RL. Experiments on standard control benchmarks and an RLHF-style toxicity-mitigation setting show that RINSE improves training efficiency and final performance over standard PPO and local attribution baselines. These results suggest that RL-Inf provides a practical and principled framework for understanding and improving online RL.

On modern production text-to-image systems, successful policy violations are rare, and previously effective human-written seeds are often patched out. Current automated red-teamers are poorly matched to this regime in two ways: unreliable success measurement and poor exploration. First, we find that judges widely used in prior T2I red-teaming work are unreliable under vague unsafe-content targets: they either miss true violations or reward benign borderline images on hardened APIs. We therefore define strict category-specific success criteria and calibrate strong VLM judges against human labels. Second, we show that broadly used prompt-modification pipelines do not solve the exploration problem: on harder guardrail settings they remain tied to seed prompts, fail to transfer, or cannot bootstrap positive examples. We introduce RISE, which evolves reusable strategies used to generate prompts rather than rewriting them one by one. The best discovered strategies are then reused to generate attacks across new scenarios. On DALL·E 3, Nano Banana 2 (Google) and GPT-Image-2, RISE reaches up to 13\% human-verified ASR; under the same calibrated evaluation, prior methods with reported ASR as high as roughly 30\% fall to near zero.

Specifying segmentation targets via text descriptions provides flexible semantic conditioning for medical image segmentation. However, existing methods still struggle to robustly translate high-level textual semantics into accurate pixel-level spatial prompts. In this paper, we propose RADAR (Residual Aggregation and Dense Alignment Representations), a parameter-efficient framework for text-guided medical image segmentation. RADAR explicitly decouples and synergizes the medical semantic anchoring capabilities of BiomedCLIP with the prompt-driven mask decoding abilities of MedSAM2, achieving low-cost task adaptation via LoRA. Specifically, RADAR introduces Guided Residual Depth Aggregation, which adaptively aggregates historical layer features under semantic guidance to mitigate textual semantic injection lag and deep feature dilution. Furthermore, we propose Structured Dense Representation Alignment, which leverages a frozen BiomedCLIP to construct crop-level semantic anchors, namely target, boundary, and context, to enhance the local discriminative power of dense features and the reliability of automated prompts. Additionally, Multi-modal Prompt Generation is designed to transform cross-modal semantic responses into composite prompts consisting of text, points, and masks. Experimental results on three public medical image segmentation datasets demonstrate that RADAR achieves competitive performance while introducing merely 0.74M trainable parameters, validating the effectiveness of the proposed approach in parameter-efficient adaptation and the enhancement of cross-layer and local representations. The source code is provided in the supplementary material.

We study adversarial multi-armed bandits with and without delayed feedback under a **safety-aware** goal: achieving minimax-optimal worst-case regret while keeping nearly **constant** regret relative to a designated "safe" baseline policy. Existing approaches can balance this trade-off with immediate feedback for smooth comparators, but arbitrary delays can mistime transitions between **conservatism** and **exploration**, endangering the safety guarantee. To bridge this gap, we propose PRUDENT-BANKER, a novel algorithm that combines a delay-adapted variant of Online Mirror Descent with a modified phased-aggression mechanism. Its key technical contribution is a delay-calibrated restart threshold that rigorously accounts for the worst-case distortion induced by unobserved feedback and reliably detects comparator suboptimality. We also establish new lower bounds for safety-constrained adversarial delayed bandits, showing that the regret guarantees of PRUDENT-BANKER are unimprovable, up to logarithmic factors, under the baseline-safety requirement. To the best of our knowledge, PRUDENT-BANKER is the first algorithm to achieve the optimal safety–robustness trade-off: pseudo-regret $\tilde{O}(\sqrt{T}+\sqrt{D})$ together with $\tilde{O}(1)$ regret against the safe comparator, both with and without delays. Experiments across diverse delay distributions show that, unlike standard delay-robust baselines, PRUDENT-BANKER effectively balances safety and learning.

We consider the problem of *regularized* best-response max-regret minimization in online RLHF under general preferences and bandit feedback. While various regularizers are utilized to robustify alignment, known polylogarithmic regret guarantees remain heavily specific to KL. To investigate whether such fast rates extend beyond KL, we adopt the *Generalized Bilinear Preference Model (GBPM)*—capturing intransitive preferences over $d$-dimensional item-wise features via a rank-$2r$ skew-symmetric matrix—to isolate the impact of generic regularization. Crucially, under GBPM, we prove that the dual gap of any greedy policy is bounded by the *squared* estimation error, derived using *only* strong convexity and skew-symmetry. Under a feature coverage assumption, we establish polylogarithmic $\tilde{\mathcal{O}}(\eta d^4 (\log T)^2 \wedge d^2 \sqrt{T})$ regret with Greedy Sampling and $\mathrm{poly}(d)$-free $\tilde{\mathcal{O}}(\sqrt{\eta r T} \wedge r^{1/3} T^{2/3})$ regret with Explore-Then-Commit, where $\eta^{-1}$ is the regularization coefficient and $T$ is the time horizon. This demonstrates that ``fast'' regrets are *not* KL-specific, but rather a fundamental consequence of generic strongly convex geometry.