NeurIPS 2026 Papers

This paper proposes a guaranteed defense method for large language models (LLMs) to safeguard against jailbreaking attacks. Drawing inspiration from the denoised-smoothing approach in the adversarial defense domain, we propose a novel smoothing-based defense method, termed Disrupt-and-Rectify Smoothing (DR-Smoothing). Specifically, we integrate a two-stage prompt processing scheme—first disrupting the input prompt, then rectifying it—into the conventional smoothing defense framework. This disrupt-and-rectify approach improves upon previous disrupt-only approaches by restoring out-of-distribution disrupted prompts to an in-distribution form, thereby reducing the risk of unpredictable LLM behavior. In addition, this two-stage scheme offers a distinct advantage in striking a balance between harmlessness and helpfulness in jailbreaking defense. Notably, we present a theoretical analysis for generic smoothing framework, offering a tight bound for the defense success probability and the requirements on the disruption strength. Our approach can defend against both token-level and prompt-level jailbreaking attacks, under both established and adaptive attacking scenarios. Extensive experiments demonstrate that our approach surpasses current state-of-the-art defense methods in terms of both harmlessness and helpfulness.

Spatial and temporal resource constraints are critical for both biological and artificial intelligent systems. Here we define differentiable cost terms for breadth, depth, and time within a recurrent convolutional neural network conceived as a finite subset of an infinite lattice architecture. We optimize these costs jointly with task errors via backpropagation, and efficient computational graphs emerge organically through training. We find that breadth, depth, and time can be traded off against each other to achieve a given level of performance. Networks grow in all three dimensions with task complexity and spontaneously take more recurrent steps when inputs are occluded. Surprisingly, time used by the model correlates with human reaction times in an object recognition task. Our framework provides a normative account of how resource constraints shape neural architectures, connecting to questions about brain design in neuroscience, and may help illuminate the diversity of neural solutions found in nature.

Multi-task reinforcement learning (MT-RL) trains a single agent to solve multiple tasks by leveraging shared knowledge across task domains. However, most MT-RL methods assume centralized access to task data, making them impractical for privacy-sensitive or large-scale distributed settings. Multi-task federated reinforcement learning (MT-FRL) mitigates this issue by enabling agents to collaborate through shared models rather than raw trajectories, but often suffers from unstable performance under task heterogeneity, negative transfer, and imbalanced learning across tasks. A fundamental question in MT-FRL is how each agent should decide with whom to share skills and how strongly to rely on shared knowledge. To answer this question, we propose GAdapHeadFRL, a personalized MT-FRL framework based on geometry-aware grouping and adaptive decision-layer task-head mixing. First, we decompose the local Bellman residual into representation and task-head estimation errors, and derive a closed-form optimal sharing strength governed by task mismatch and estimation uncertainty. Second, we instantiate this insight by identifying compatible client groups from local updated geometry, constructing group-specific task-head prototypes, and dynamically mixing them with local heads using lightweight practical proxies. Experiments on heterogeneous benchmarks, including MiniGrid (up to 12 tasks) and MetaWorld (up to 50 tasks), show that GAdapHeadFRL consistently outperforms state-of-the-art personalized MT-FRL baselines. It achieves up to 24.5\% relative improvement on MiniGrid, and increases the converged-task ratio from 50.0\% to 72.9\% over the personalized baseline. In addition, the proposed method reduces performance standard deviation by over 82.9\% on the largest task settings compared with centralized MT-RL references. These results demonstrate that geometry-aware grouping and adaptive task-head mixing provide an effective and scalable principle for heterogeneous MT-FRL.

Large Language Model (LLM) agents increasingly serve as personal assistants and workplace collaborators, where their utility depends on memory systems that extract, retrieve, and apply information across long-running conversations. However, both existing memory systems and benchmarks are built around the dyadic, single-user setup, even though real deployments routinely span groups and channels with multiple users interacting with the agent and with each other. This mismatch leaves three properties of group memory unmeasured: (i) group dynamics that go beyond concatenated one-on-one chats, (ii) speaker-grounded belief tracking, where the per-user memory modeling is needed, and (iii) audience-adapted language, where Theory-of-Mind shifts produce role-specific vocabulary. We introduce GroupMemBench, a benchmark that exposes all three. A graph-grounded synthesis pipeline produces multi-party conversations with controllable reply structure and conditions each message on per-user personas and target audiences. An adversarial query pipeline then binds every question to a specific asker across six categories, spanning multi-hop reasoning, knowledge update, term ambiguity, user-implicit reasoning, temporal reasoning, and abstention, and iteratively searches challenging, realistic queries that reflect comprehensive memory capability. Benchmarking leading memory systems exposes a sharp collapse: the strongest one reaches only 46.0% average accuracy, with knowledge update at 27.1% and term ambiguity at 37.7%, while a simple BM25 baseline matches or exceeds most agent memory systems. This indicates current memory ingestion erases the structural and lexical features group memory depends on, leaving multi-user memory far from solved.

Group-based Reinforcement Learning (RL) has significantly enhanced Large Language Models (LLMs) in agentic scenarios. To achieve finer-grained policy updates, recent agentic RL frameworks have shifted from trajectory-level to step-level training. However, long-horizon agentic RL suffers from severe reward sparsity and delay, as feedback is often deferred for dozens of interaction steps. While existing step-level frameworks refine training granularity, their credit assignment remains coarse-grained and still treats agent exploration as isolated, linear trajectories. This oversimplified perspective ignores the inherent graph structure of state transitions, leading to high-variance state-value estimation and myopic, localized credit assignment. To overcome these critical bottlenecks, we propose Group-Graph Policy Optimization (G2PO), a novel group-based RL algorithm tailored for multi-turn agentic tasks. G2PO explicitly transforms linear interaction trajectories into a global state-transition graph. By aggregating identical observations across different trajectories, we introduce group-aggregation state-value estimation that reduces sampling variance and trajectory-dependent bias. Furthermore, we redefine agent actions as transitions between state nodes and propose an edge-centric advantage estimation strategy. By globally standardizing Temporal Difference (TD) errors across the entire graph, G2PO explicitly identifies and prioritizes critical transitions that drive absolute task progress. Extensive experiments on representative long-horizon benchmarks—WebShop, ALFWorld, and AppWorld—demonstrate that G2PO substantially outperforms state-of-the-art prompt-based and RL baselines, achieving remarkable success rate improvements of up to 22.2\% over GRPO.

Skill-augmented agents increasingly rely on large reusable skill libraries, but retrieving relevant skills is not the same as presenting usable context. Existing methods typically return atomic skills or dependency-aware bundles whose internal roles remain implicit, leaving the agent to infer the execution entry point, support skills, visible requirements, and failure-avoidance guidance. We introduce **G**roup **o**f **Skills** (**GoSkills**), an inference-time group-structured retrieval method that changes the agent-facing retrieval object from a flat skill list to a compact, role-labeled execution context. **GoSkills** builds anchor-centered skill groups from a typed skill graph, expands support groups through a group graph, bottlenecks the selected group plan into a bounded set of atomic skill payloads, and renders a fixed execution contract with **START**, **SUPPORT**, **CHECK**, and **AVOID** fields, without changing the downstream agent, skill payloads, or execution environment. Experiments on SkillsBench and ALFWorld show that \goskills preserves visible-requirement coverage under a small skill budget, improves over flat skill-access baselines, and often improves reward and agent-only runtime relative to structural retrieval references. Code is available at https://anonymous.4open.science/r/Group-of-Skills-E861.

Multi-Agent Debate (MAD) improves the reasoning performance of Large Language Models (LLMs) through multi-round interaction. However, LLMs in MAD are highly susceptible to blind conformity. Existing individual evaluation methods, typically based on confidence or perplexity, fail to reflect the correctness of reasoning and may even exacerbate blind conformity. To address this, we shift the perspective from individual evaluation to group interaction. We define mutual referencing among LLMs as \textbf{Debate Relationships} and recognize that regulating these relationships is the key to mitigating blind conformity. In this paper, we propose a novel framework for \textbf{D}ynamically r\textbf{E}gulating deb\textbf{A}te \textbf{R}elationships (DEAR) from the group perspective. At first, DEAR quantifies consensus and divergence as \textit{group evidence} to capture the debate state. Then, DEAR operates through three stages: 1) What: perceiving group consultation tendency and uncertainty; 2) Who: introducing a Selection RL-Agent to dynamically select reference peers; and 3) How: adopting a Behavior RL-Agent to adaptively adjust generation behaviors.

Group distributionally robust optimization (GDRO) aims to learn models that perform well across $m$ distributions simultaneously. Existing methods either (i) query $m$ samples per round, which imposes strict implementation requirements, or (ii) query only a single sample per round, which requires many rounds to converge and therefore leads to long runtime. To avoid these limitations, we introduce a novel setting called \emph{flexible sample queries} (FSQ), which allows the number of samples that can be queried to vary across rounds. Under the FSQ setting, we cast GDRO as a two-player game: one performs non-oblivious online convex optimization, while the other tackles a non-oblivious prediction with limited advice (PLA) problem. In this game, the first player updates its decision using follow-the-regularized-leader with stochastic gradients. For the second player, we develop a new PLA method equipped with an adaptive exploration strategy that can leverage an arbitrary per-round sample size. We then establish the *first* high-probability regret bound for non-oblivious PLA. By integrating these components, we develop an anytime GDRO algorithm that supports FSQ, and derive a high-probability optimization error bound of $O\left(\frac{1}{t}\sqrt{\sum_{j=1}^t \frac{m}{r_j}\log m}\right)$, where $r_j$ is the sample size at round $j$. This result demonstrates that the optimization error decreases as the per-round sample sizes increase, and implies the same near-optimal sample complexity of $O(m\log (m)/\epsilon^2)$ for any fixed sample size $r\in[m]$.

Reinforcement learning with verifiable rewards (RLVR) has become a powerful tool for improving multimodal reasoning. However, standard outcome supervision suffers from credit assignment ambiguity, as a model may receive positive rewards by exploiting language priors or dataset shortcuts rather than grounding its reasoning in the requisite visual evidence. This makes high-quality negative responses critical, since effective negatives should remain plausible under the current policy while exposing the shortcut that should be suppressed. In this paper, we propose EAN, a multimodal reinforcement learning framework utilizing Evidence-Ablated Negatives to train LMMs against their own language-prior failures. EAN first identifies highly visually dependent examples using a scaled image information gain metric and preserves data diversity with a mixed-ranking strategy. It then constructs evidence-ablated images by masking the most discriminative image patches, eliciting plausible but visually ungrounded responses from the current policy. During policy optimization, these negative responses are incorporated as bounded auxiliary signals under the clean image, penalizing shortcut learning while maintaining RL stability. Experiments across challenging multimodal reasoning benchmarks demonstrate that EAN mitigates modality imbalance and improves over strong RLVR baselines.

Long-horizon LLM agents trained with outcome-based RL exhibit a systematic failure mode: reasoning diverges from environmental observations at some step, and the error compounds across subsequent turns. We find that 74\% of failed trajectories contain reasoning–observation inconsistency, and once it appears, 62\% of subsequent turns remain inconsistent. Process rewards reduce but do not eliminate this: under free-form reasoning, agents learn to avoid penalties without genuinely tracking task progress or anticipating action outcomes. We propose ARC (Anchored Reasoning with Commitment), which requires the agent to make two explicit commitments at each step: a global progress statement that reconciles the agent's understanding of task state with accumulated observations, and a local outcome prediction that must be borne out by the next observation. Together, these commitments anchor the agent's reasoning to the actual environment state across turns, and provide well-defined targets for process supervision. To prevent interference between sparse outcome and dense process signals, we decouple their advantages via independent group normalization. On ALFWorld, WebShop, and WebArena, ARC improves over the strongest baseline by 6.4, 3.5, and 9.6 points respectively.

3D affordance grounding aims to highlight the actionable regions on 3D objects, which is crucial for embodied AI. While previous research primarily leverages static cues from language or single images, which often lack the rich interaction context necessary to accurately locate functional zones. To alleviate this predicament, we collect a comprehensive video-based 3D affordance dataset, VIDA, which contains 38K human-object-interaction (HOI) videos covering 16 affordance types, 38 object categories, and 22K point clouds. Based on VIDA, we propose a strong baseline: VideoAfford, a unified framework that extends Multimodal Large Language Models (MLLMs) with fine-grained affordance segmentation capabilities. VideoAfford incorporates a latent action encoder to distill dynamic interaction knowledge from demonstration videos and a spatial-aware loss to encourage geometrically consistent affordance predictions. Extensive experiments on VIDA show that VideoAfford significantly outperforms strong baselines in both seen and unseen settings, demonstrating its effectiveness in video-driven 3D affordance reasoning and open-world generalization. The dataset and code will be released soon to facilitate future research in this area.

Generating egocentric video from a single exocentric video is an emerging and important topic for AR/VR and physical AI. Compared with conventional novel view synthesis, exo-to-ego generation is a significantly harder task due to the extreme viewpoint changes, extensive occlusions and unobserved regions, where standard geometric conditioning becomes highly unreliable. We present Grounded-Exo2Ego, a principled framework that addresses these challenges at both the architectural and data levels. Architecturally, Grounded-Exo2Ego is a dual-branch video diffusion model that couples a geometric anchoring branch, which renders a 3D reconstruction into the ego view for overall scene layout, with a semantic grounding branch, which anchors per-object context onto the scene geometry so that heavily distorted regions and unobserved regions can be hallucinated in context. Additionally, to support robust training at scale, we substantially improve the labeling pipeline for real-world data and further develop a fully automated synthetic data engine that renders high-quality dynamic humans in procedurally generated environments. Evaluations on the challenging EgoExo4D dataset show that our method substantially outperforms recent state-of-the-art approaches. Extensive ablations further demonstrate that both structured semantic grounding and rigorous data processing are essential for robust exo-to-ego video generation in extreme scenarios.

Hallucinations in large language models (LLMs) that are plausible-sounding but factually incorrect or unsupported pose a major challenge for deploying these models in high-stakes applications such as medical diagnosis, legal reasoning, and knowledge-based question answering. Existing detection methods primarily rely on either generating multiple responses to check consistency or supervised training on labeled hallucinations. Multiple-response approaches tend to be computationally expensive and may be less practical for real-time scenarios, while supervised methods are limited to known hallucination types and cannot generalize to unseen cases. To address these limitations, we propose an efficient unsupervised hallucination detection method. Our approach measures the embedding discrepancy between the LLM’s response considered independently and the response contextualized with the question. A large discrepancy indicates that the question significantly influences the response, suggesting it is grounded, whereas a small discrepancy signals a higher likelihood of hallucination. This method is efficient, interpretable, and does not require labeled hallucinations. Extensive experiments across multiple tasks demonstrate that our approach consistently achieves superior or comparable detection performance while reducing the detection cost compared to the multiple-response methods.

Deep neural networks exhibit periodic loss spikes during unregularized long-term training, a phenomenon known as the "Slingshot Mechanism." Existing work usually attributes this phenomenon to intrinsic optimization dynamics, but its triggering mechanism remains unclear. This paper shows that, under commonly used floating-point precision, the Cross-Entropy loss itself can produce a deterministic numerical instability. As training enters a high-confidence stage, the difference between the correct-class logit and the other logits may exceed the absorption-error threshold of float-point number. Then during backpropagation, the gradient of the correct class is rounded exactly to zero, while the gradients of the incorrect classes remain nonzero. This breaks the zero-sum constraint of gradients across classes and introduces a systematic drift in the parameter update of the classifier layer. We prove that this drift forms a positive feedback loop with the feature, causing the global classifier mean and the global feature mean to grow exponentially. We call this mechanism _Numerical Feature Inflation_ ($\mathcal{NFI}$). This mechanism explains the rapid norm growth before a Slingshot spike, the subsequent reappearance of gradients, and the resulting loss spike. We further show that $\mathcal{NFI}$ is not equivalent to an observed loss spike: in more practical tasks, absorption errors may affect only a subset of samples, so spikes do not necessarily occur, while the same zero-sum-breaking mechanism can still drive rapid growth of parameter norms. Our results reinterpret Slingshot as a numerical dynamic of finite-precision training, and provide a testable explanation for abnormal parameter growth and logit divergence in late-stage training.

Long-video understanding in VLMs is bottlenecked by a single monolithic forward pass over thousands of frames at quadratic attention cost. A common mitigation is to first $select$ a small subset of informative frames before the forward pass; common for training-free selectors via auxiliary encoder-space similarities. Such signals are capped by contrastive pretraining, which usually fails on reasoning-heavy queries (negation, cross-frame counting, holistic summarization). We propose $\textbf{GridProbe}$, an efficient training-free posterior-probing inference paradigm that scores evidence in $\textbf{answer space}$ using a frozen VLM's own reasoning and then selects question-relevant frames $\textbf{adaptively}$, resulting in sub-quadratic attention cost with little to no accuracy loss. We arrange frames on a $K{\times}K$ grid and run lightweight row $R$ and column $C$ probes, where each probe reads its peak posterior as a query-conditioned confidence. The outer product of $R$ and $C$ yields an interpretable $\textbf{importance map}$ whose skewness and kurtosis drive $\textbf{Shape-Adaptive Selection}$, a closed-form rule that reliably replaces the fixed frame budget $M$ with a per-question $M_{\mathrm{eff}}$. We show empirically that $M_{\mathrm{eff}}$, surprisingly, tracks intrinsic question difficulty without ever seeing the answer, a sign of test-time adaptive compute. On Video-MME-v2, $\textbf{GridProbe}$ matches the monolithic baseline within $1.6$ pp Avg Acc at $3.36\times$ TFLOPs reduction, while on LongVideoBench it Pareto-dominates the baseline ($+0.9$ pp at $0.35\times$ compute). Because the selector and QA models can be decoupled, pairing a small 2B selector with a stronger 4B or 8B QA is strictly Pareto-dominant over the 2B monolithic baseline (up to $+4.0$ pp at $0.52\times$ compute, on average), with no retraining. Finally, the interpretability of the importance maps opens future avenues for behavioral diagnostics, grounding, and frame-selection distillation.

A major recent advance in quantization is given by microscaled 4-bit formats such as NVFP4 and MXFP4, quantizing values into small groups sharing a scale, assuming a fixed floating-point grid. In this paper, we study the following natural extension: assume that, for each group of values, we are free to select the "better" among two or more 4-bit grids marked by one or more bits in the scale value. We formalize the power-of-two-grids (PO2) problem, and provide theoretical results showing that practical small-group formats such as MXFP or NVFP can benefit significantly from PO2 grids, while the advantage vanishes for very large groups. On the practical side, we instantiate several grid families, including 1) PO2(NF4), which pairs the standard NF4 normal grid with a learned grid, 2) MPO2, a grid pair that is fully learned over real weights and activations, and 3) SFP4, a TensorCore-implementable triple which pairs NVFP4 with two shifted variants. Results for post-training quantization of standard open models and pre-training of Llama-like models show that adaptive grids consistently improve accuracy vs single-grid FP4 under both weight-only and weight+activation quantization.

Recent works have shown that gradient-update alignment is a powerful signal for modulating optimizer updates and improving training dynamics. We promote this update-wise heuristic into a mathematically grounded principle for selecting and tuning optimizer hyperparameters. By treating gradients and updates as signals, and an optimizer as a causal filter that maps between them, we formulate optimizer selection as maximizing the expected drop rate in loss over a prescribed family of optimizers. We show that this objective is exactly the inner product between the optimizer filter and the gradient autocorrelation, and prove that a greedy optimum exists and has a stability bound under perturbations of the estimated gradient statistics. Specializing in momentum-based optimizers, the theory yields simple dynamic momentum selection rules for both SGD+Momentum and Adam/AdamW. Experiments across image classification, language model fine-tuning, and vision transformer fine-tuning show that the resulting dynamic momentum rules match or improve upon the best fixed hyperparameters found via manual sweeps, reducing the need for exhaustive momentum sweeps.

In robotic manipulation, the tight coupling between grasping and motion planning often obscures the true source of failure, leading to inefficient trial-and-error. To enable efficient long-horizon manipulation, we propose GTP-FA (Grasp-Then-Plan with Failure Attribution), a task-oriented two-stage ‘grasp-then-plan’ framework that generates grasp candidates and performs downstream motion planning conditioned on the selected grasp. Given a failed manipulation trajectory, we learn a failure attribution model that generalizes to unseen grasps and produces a stable distribution over failure modes for diagnosis-guided optimization. Based on these attribution results, we then optimize both modules in a diagnosis-driven manner: on the grasping side, we inject task-level priors and risk penalties into grasp candidate scoring and optimization to suppress unstable or task-incompatible grasps; on the planning side, we target high-risk initial states through data collection and fine-tuning to address genuine planning bottlenecks. We evaluate the proposed framework in both simulation and real-robot experiments, and show that GTP-FA improves the corresponding base learners across RL, IL, diffusion-policy, and VLA-based settings, achieving substantially higher overall task success rates. Project page: here.

Models that recognize when their predictions are unreliable inspire greater trust than those that make confident but incorrect decisions. While recent advances enable uncertainty quantification in GNNs, understanding *why* a model is uncertain, whether due to noisy features, conflicting neighborhoods, or sparse training coverage, remains unsolved. We address this gap with GraphUAT, the first *uncertainty attribution* framework for GNNs, which identifies the sources of uncertainty rather than aiming to improve its quantification. GraphUAT employs a teacher-student paradigm, where the trained GNN's behavior is distilled into an uncertainty-aware student that disentangles aleatoric and epistemic uncertainties. We then use targeted probes to trace the student's uncertainty back to the responsible features and structural components. Experiments across benchmarks show that removing the identified elements consistently reduces the teacher's uncertainty, validating the faithfulness of our attributions.