NeurIPS 2026 Papers

Existing underwater image enhancement (UIE) methods predominantly rely on unidirectional mapping, which lacks often results in over-enhancement or under-enhancement, without user-friendly controllability. To address this, we propose a paradigm of unsupervised trajectory learning for Omnidirectional Controllable UIE that treats restoration as a omnidirectional dynamical system, namely OC-UIE. The main purpose of our work is the realization of a consistent omnidirectional mapping across the representation space, which allows the model to master complex underwater dynamics beyond traditional unidirectional constraints. To implement this, we propose an Omnidirectional Training that optimizes over arbitrary source-target positions sampled from the data distribution. We decompose this purpose into two synergistic stages: first, Representation Trajectory Flattening is employed to organize intricate degradations into a unified adaptation axis; second, Trajectory Omnidirectional Integration is introduced to model the enhancement as a path integral over the resulting manifold. By optimizing only the relative shifts, OC-UIE isolates structural preservation from fluid degradation dynamics. This approach provides a controllable space for intensity-parameterized enhancement with state-of-the-art performance.

As large language models (LLMs) become more capable and agentic, the requirement for trust in their outputs grows significantly, yet at the same time concerns have been mounting that models may learn to lie in pursuit of their goals. To address these concerns, a body of work has emerged around the notion of "honesty" in LLMs, along with interventions aimed at mitigating deceptive behaviors. However, some benchmarks claiming to measure honesty in fact simply measure accuracy—the correctness of a model's beliefs—in disguise. Moreover, no benchmarks currently exist for directly measuring whether language models overtly lie. In this work, we introduce a realistic, large-scale human-curated dataset for directly measuring lies of commission, allowing us to disentangle accuracy from honesty. Across a diverse set of LLMs, we find that while larger models obtain higher accuracy on our benchmark, they do not become more honest. Surprisingly, most frontier LLMs obtain high scores on truthfulness benchmarks yet exhibit a substantial propensity to lie under pressure, resulting in low scores on our benchmark. We find that simple methods, such as representation engineering interventions, can improve honesty. These results underscore the growing need for robust evaluations and effective interventions to ensure LLMs remain trustworthy.

We argue that cooperative multi-agent reinforcement learning (MARL) should treat test-time coordination as part of the problem definition rather than a deployment artifact. Our focus is on long-horizon, centrally monitored cooperative systems such as warehouse robotics, automated storage-and-retrieval systems, and cloud-resource scheduling, where teams share a reward, operate under shifting objectives, and already run with global telemetry at deployment. The dominant paradigm in this regime, centralized training with decentralized execution (CTDE), uses global information only during training and provides no coordination channel at execution. This is sufficient when agents are weakly coupled or the environment is stationary, but it fails when the team must realign after conditions change. The limitation is structural rather than algorithmic, and more training cannot add a channel that does not exist at execution time. We call this the recoverability gap. To address it, we name the underexplored design space of intermittent test-time coordination as centralized training with mixed execution (CTME), in which a coordinator observes global state and emits high-level signals while agents act locally. CTME is not a particular learning algorithm, but a cooperative MARL problem setting that makes test-time information structure explicit. It specifies team-level information, high-level signals, and update frequency after training, with CTDE recovered as a limit case. We do not argue that CTDE is obsolete. We argue that removing the test-time channel is a modeling choice, not a necessity, and the wrong default when global telemetry is available and teams must recover from drift.

Cognitive primitives such as planning, exploration, and backtracking are widely regarded as core to competent web agents. Yet existing studies are largely retrospective: agents are evaluated on domain-based tasks, and trajectories are inspected post hoc for evidence of these behaviors. Because such tasks are not designed to require specific primitives, observed behaviors do not reliably reflect true capability demands. This raises a key question: Can we design realistic tasks where success provably depends on a chosen cognitive primitive? WEBSTRESS addresses this by casting capability evaluation as a controlled comparison. For each task, we construct a paired variant in which a target capability becomes necessary (e.g., transforming a direct path into an obstructed one requiring exploration). The performance gap between the pair measures the agent’s reliance on that capability. Across seven environments and seven capabilities, this yields 519 paired tasks. Evaluating six frontier agents reveals consistent weaknesses: all suffer 18-28\% drops on capability-demanding variants, with 75\% of failures being belief failures, where agents incorrectly declare success. In contrast, humans exhibit only a 5.7\% performance drop under the same paired evaluation. These results highlight that, despite strong overall performance, substantial gaps remain between human and agent cognitive behavior, and that our approach provides a principled way to measure them.

Long-horizon agents increasingly operate across many steps, tools, and observations. In this setting, the relevant oversight question is not only whether each action is locally valid, but whether the evolving trajectory still corresponds to the task the user authorized. Drift can accumulate quietly: an agent may call the right tool with plausible arguments at every step, while its prefix moves toward a broader role, an adjacent objective, or evidence the user never supplied. Existing monitors mostly check local compliance, deliver final-trace verdicts, or score generic risk; they do not directly estimate this prefix-level relation. We introduce ontological trust, a task-conditioned property of trajectory prefixes, and instantiate it as RGE, an online monitor that decomposes trust along Role, Goal, and Evidence. RGE uses LLMs only to derive structured representations. State updates, projections, and intervention decisions are deterministic, so the output is a replayable and auditable trust trajectory rather than a single end-to-end judge verdict. We construct a cross-domain trajectory corpus from OSWorld, FinanceBench, and EICU-AC, covering benign executions, prefix-paired drift, and pseudo-consistency failures. On this corpus, RGE outperforms adapted rule-, judge-, and shield-style baselines on prefix-paired drift detection. With the two larger estimator models, it exceeds 93\% Drift F1 on every benchmark while keeping benign coverage at or above 95.8\%. Pseudo-consistency is harder: detection depends on whether task completion is externally visible, a structural limit we characterize empirically.

Offline goal-conditioned reinforcement learning (GCRL) is typically benchmarked by the best tuned success rate of each method. This score measures attainable performance, but it does not reveal how reliably a learned goal-conditioned signal can be extracted into a policy: a method could succeed across many value-learning and extraction settings, or only at a narrow, hard-to-find configuration. We study this gap across four methods, GCIQL, GCIVL, QRL, and CRL, under a shared advantage-weighted regression (AWR) extractor. For each method, we construct trainability landscapes over the optimizer learning rate, which affects value learning and actor optimization, and AWR temperature, which controls how selectively the actor imitates high-advantage transitions. Across AntMaze, Cube, and Scene, we observe distinct regimes: high-scoring methods may be broadly accessible or brittle, while broad relative basins may still sit below low absolute ceilings. To interpret these differences, we pair landscapes with post-hoc diagnostics of future-vs-random goal discrimination and AWR weight concentration. % whether the learned advantage function separates matched future goals from random goals, and whether AWR spreads weight across many behavior-supported transitions or concentrates on a small tail. Their relationship to downstream success is task-dependent. On AntMaze, where future goals align with path-like progress, these diagnostics explain landscape regimes. On Cube and Scene, goal ranking and manipulation control decouple: methods can rank goals well while failing downstream, or succeed through action-conditioned advantages despite weak future-vs-random separation. These results show that peak tuned success alone does not establish broadly extractable goal-conditioned behavior. Trainability landscapes expose this gap, while extraction diagnostics offer a lower-cost lens on how learned signals become policies.

Activation steering has emerged as a promising alternative for controlling language-model behavior at inference time by modifying intermediate representations while keeping model parameters frozen. However, large-scale evaluations such as AxBench show that existing steering methods are often outperformed by simple in-context prompting and generalize poorly to unseen concepts. We hypothesize that these limitations arise from unvalidated simplifying assumptions shared across prior methods, which typically restrict steering interventions to fixed, single-step, position-invariant transforms. We propose FLAS (Flow-based Activation Steering), which learns a general, concept-conditioned velocity field $v_t(h,t,c)$ that transports unsteered activations to steered ones without relying on these assumptions. On AxBench, FLAS is the first learned method to consistently outperform prompting, reaching held-out harmonic means of $1.015$ on Gemma-2-2B-IT and $1.113$ on Gemma-2-9B-IT without per-concept tuning. Analysis of the learned flow shows curved, multi-step, token-varying trajectories, which suggests that previous hypotheses on activation space geometry might be incomplete. Our code is available at https://anonymous.4open.science/r/FLAS.

Semi-supervised medical image segmentation has gained increasing attention for its outstanding performance with limited annotations. Most methods follow the Mean Teacher framework, where Exponential Moving Average (EMA) is primarily used to stabilize predictions, while its potential roles beyond stabilization are largely overlooked. In this work, we interpret EMA as a first-order infinite impulse response (IIR) low-pass filter with inherent frequency selectivity, which enables the preservation of semantic information at different levels. Motivated by this insight, we propose a dual-EMA teachers framework, including a long-term teacher focusing on global structures and a short-term teacher emphasizing local details. Furthermore, to help the student better exploit the guidance from both teachers, a First-In-First-Out replay queue is designed to reuse historical high-confidence pseudo-labeled pairs, leveraging pseudo labels provided at different training stages and reducing the sensitivity to sampling bias. Meanwhile, we impose feature-level global semantic and local relation consistency constraints to alleviate the impact of noisy pseudo labels, facilitating robust transfer of complementary representations from dual teachers to the student model. Experiments on three public medical image datasets demonstrate that our method outperforms current SoTA methods, verifying its effectiveness.

Dense correspondence matching has historically been bounded by simplifying spatio-temporal priors, such as smooth motion and rigid geometry. While effective for classical tasks, these foundations collapse at the frontier of Vision-Language-guided Image Editing and Generation (VL-IEG), where transformations yield correspondences that are perceptually obvious yet physically discontinuous. To transcend spatio-temporal priors, we draw inspiration from human cognition: learning a generalized model of identity-preserving visual consistency rather than relying on physical constraints. To achieve this, we introduce FreeMatching, a unified framework that extracts identity-preserving knowledge from foundation models and diverse datasets. The model's capability is forged through a two-stage training paradigm: supervised pre-training on diverse annotated datasets, followed by weakly-supervised refinement on data without dense annotations. Experimentally, FreeMatching not only achieves performance competitive with SOTA methods on classical benchmarks but also establishes the first strong baseline for the challenging VL-IEG task. Furthermore, we demonstrate its utility as a quantitative metric for evaluating identity-preserving consistency in generative editing, aligning closely with human judgment.

Conditional diffusion models have emerged as a powerful paradigm for image restoration, leveraging pre-trained generative priors to recover high-fidelity details from degraded inputs. However, existing methods typically follow a static inference protocol, treating the degraded observation as a fixed control signal throughout sampling. This overlooks the potential of inference-time scaling, where guidance can be refined on the fly to improve output quality. Recent test-time optimization approaches that manipulate noise variables or denoising trajectories often compromise structural fidelity, while those targeting text embeddings lack the spatial precision needed for restoration. To address this, we propose Condition-Aware Test-Time Optimization (CATTO), a training-free method that iteratively refines the visual conditioning signal at inference time. CATTO performs reward-aligned condition refinement under the pre-trained prior to enhance perceptual quality. To avoid complex backpropagation, we design a gradient-free optimization strategy guided by a joint objective: maximizing a perceptual reward while enforcing trajectory- and condition-level consistency to explicitly preserve structural fidelity. This process is accelerated by optimizing within a low-dimensional frequency subspace and reusing reliable updates across nearby denoising steps. Experiments on image restoration benchmarks show that CATTO improves perceptual quality over strong diffusion-based baselines without updating model parameters.

Multimodal Large Language Models (MLLMs) have demonstrated substantial promise in spatial understanding. Existing works typically incorporate prior knowledge extracted from a pre-trained foundation model to further enhance the spatial awareness of MLLMs. In this paper, we first reveal that when integrating diverse foundation models into MLLMs, each model provides complementary spatial priors that benefit different tasks. Motivated by this, we propose $\textbf{ViPS}$, a novel multi-model prior framework designed to fully unleash the potential of incorporating multiple $\textbf{Vi}$sual $\textbf{P}$riors from diverse models into MLLMs for $\textbf{S}$patial understanding. Specifically, ViPS introduces an Efficient Prior Proxy to generate multiple foundational priors with minimal inference overhead, and a Dynamic Prior Fusion mechanism to achieve harmonious and context-aware prior fusion and injection from the prior proxies. Extensive experiments demonstrate that ViPS successfully harmonizes diverse visual priors, establishing new state-of-the-art performance across multiple complex spatial reasoning and 3D spatial understanding benchmarks.

Many decision-level fusion methods implicitly assume that semantically aligned soft outputs can be directly aggregated once they share the same label simplex. However, this assumption can fail even under identical label semantics, as oracle-specific differences in confidence sharpness, boundary uncertainty, and tail-mass allocation may place soft outputs in incompatible probability geometries and bias direct soft fusion. To address this issue, we formalize this failure mode as geometric incomparability and propose a pre-fusion comparability recovery framework, which learns a shared consensus through constrained oracle-specific rectifiers on the probability simplex. The framework includes a closed-form variant with frozen geometry descriptors and an alternating-refinement variant with consensus-dependent descriptors. Theoretically, we analyze the rectifier family, prove exact recovery under matched radial distortions, characterize the bias of direct fusion, and establish stable recovery under approximate mismatch. Experiments on controlled synthetic stress tests and real black-box decision-level settings show that comparability recovery consistently reduces fusion bias and improves reliability over direct aggregation. These results indicate that geometric comparability is a necessary condition for reliable soft-output fusion. The code is available at \url{https://anonymous.4open.science/r/CFCR-ACR-B8F1}.

Vision encoders for retrieval are typically trained with class-label supervision: each training pair reduces to a scalar that uniformly pushes the embedding apart or pulls it together, as if every visual attribute either differed or matched. A multimodal large language model (MLLM), shown the same pair, can articulate those attributes and use them to predict whether the images share a class. We propose \textbf{SAGA}, a framework that turns this language-grounded, attribute-aware perception into a training signal for the encoder itself. Specifically, we use Group Relative Policy Optimization (GRPO) to reward the MLLM for correct predictions on the vision encoder's tokens. Since correct predictions require those tokens to expose the specific attributes that differ or match between the pair, the gradient pushes the encoder to encode them, replacing the uniform pair-level scalar with attribute-resolved supervision. An auxiliary attention-distillation loss anchors the encoder's embedding to tokens the MLLM attended to, and a standard metric-learning loss shapes the embedding geometry for nearest-neighbour retrieval. The MLLM is frozen throughout and discarded at inference, matching the deployment cost of a metric-learning baseline. SAGA improves Recall@1 by 3 to 6 points over state-of-the-art baselines on CUB-200-2011, Cars-196, FGVC-Aircraft, and iNaturalist Aves on zero-shot image retrieval.

The standard post-training recipe for large multimodal models (LMMs) applies supervised fine-tuning (SFT) on curated demonstrations followed by reinforcement learning with verifiable rewards (RLVR). However, SFT introduces distributional drift that neither preserves the model's original capabilities nor faithfully matches the supervision distribution. This problem is further amplified in multimodal reasoning, where perception errors and reasoning failures follow distinct drift patterns that compound during subsequent RL. We introduce \textbf{PRISM}, a three-stage pipeline that mitigates this drift by inserting an explicit distribution-alignment stage between SFT and RLVR. Building on the principle of on-policy distillation (OPD), PRISM casts alignment as a black-box, response-level adversarial game between the policy and a Mixture-of-Experts (MoE) discriminator with dedicated perception and reasoning experts, providing disentangled corrective signals that steer the policy toward the supervision distribution without requiring access to teacher logits. While 1.26M public demonstrations suffice for broad SFT initialization, distribution alignment demands higher-fidelity supervision; we therefore curate 113K additional demonstrations from Gemini~3 Flash, featuring dense visual grounding and step-by-step reasoning on the hardest unsolved problems. Experiments on Qwen3-VL show that PRISM consistently improves downstream RLVR performance across multiple RL algorithms (GRPO, DAPO, GSPO) and diverse multimodal benchmarks, improving average accuracy by +4.4 and +6.0 points over the SFT$\rightarrow$RLVR baseline on 4B and 8B, respectively.

Stereo matchers estimate depth by comparing the left and right images from a stereo camera pair, but they assume that the two cameras are accurately calibrated and rectified. In real deployments, small physical shifts after calibration, caused by camera motion, mechanical drift, vibration, or installation changes, can break this assumption and degrade depth estimation. Online stereo rectification is therefore needed to correct the image pair during operation and recover a stereo geometry where standard stereo matching can work reliably. Recent online rectification methods mainly optimize row alignment, making the left and right views of the same scene point fall on the same image row. We show that this target, while necessary, is insufficient for deployable stereo depth estimation. A rectifier may make a small set of matched points look well row-aligned, but still distort the full images in ways that hurt stereo matching: useful image regions may be cut off, large blank areas may appear, objects may be stretched or squeezed, or image scale may change unevenly across the view. We propose VirtualRect, an online stereo rectification method designed for reliable real-world deployment. Rather than treating rectification as simply making the left and right image rows line up, this method estimates a rectified virtual stereo camera. This makes the rectified images and the camera parameters come from the same geometry, so the images used for matching and the model used for depth computation remain consistent. To improve stability, it further constrains the rectification behavior over the whole image and reduces the influence of geometrically unreliable matches. Experiments on multiple datasets show that, compared with the prior state-of-the-art online rectifier, VirtualRect reduces vertical flow error by 17\% on Carla-Flowguided and 7\% on Semi-Truck Highway.

Forensic verification has converged on a binary "real vs. fake" label that conflates fully synthetic, tampered, and AI-retouched images despite their very different consequences. We argue that GenAI manipulations decompose along two complementary channels: a camera channel of sensor fingerprints (destroyed by neural processing), and a semantic channel of scene-level coherence (disrupted by content edits). Each manipulation type leaves a distinctive two-channel signature. We instantiate this taxonomy in 2CAP (2-Channel Authenticity Protocol), pairing a contrastively-trained camera encoder with a frozen semantic encoder to serve two applications on a shared backbone: (i) four-class authenticity classification via cross-attention fusion with learnable reliability weights; and (ii) evidence generation, where an authentic image serves as a reference for a query and the two encoders produce per-channel alignment scores together with a patch-level saliency map. These signals are supplied as privileged context to a frozen Vision Language Model (VLM) through an Observe-Generate-Refine loop, yielding explanations with localization without forensic fine-tuning. On a multi-source benchmark, 2CAP attains the best classification performance, including strong retouching detection where existing methods fail, and improves explanation quality.

Modern end-to-end autonomous driving systems often suffer from fragile out-of-distribution (OOD) generalization and limited decision transparency, largely due to their reliance on raw sensory observations and susceptibility to spurious visual correlations. Although memory-augmented approaches have been explored to improve generalization, existing methods typically store raw observations, symbolic rules, or entangled latent features, making retrieval sensitive to surface-level similarity rather than decision-relevant driving structure. To address this limitation, we propose DriveMem, a plug-and-play framework that moves beyond raw observations by distilling and storing invariant driving memories for generalizable driving systems. DriveMem integrates two core components: Invariant Feature Abstraction that first distills perturbation-stable, decision-relevant representations from visual tokens, and Prototype Memory Bank that then stores and retrieves prototypical driving experiences grounded in invariant representations. Extensive open-loop and closed-loop experiments demonstrate that DriveMem improves OOD generalization and safety-related planning metrics across diverse unseen scenarios. Notably, under severe sensor disruptions, DriveMem achieves up to a 39.5\% relative reduction in Collision Rate compared with a strong domain-generalization baseline. These results suggest that building prototype memory in an invariant representation space is a promising approach for improving robustness, while offering prototype-grounded traceability for model behavior.

Retrieval-augmented generation (RAG) improves the factuality of large language models (LLMs) and vision-language models (VLMs) by grounding generation in external knowledge. However, most existing RAG frameworks assume a centralized retrieval document pool, which is often impractical in sensitive domains such as healthcare, where data are inherently distributed and subject to strict privacy constraints. Recent efforts on decentralized RAG primarily follow prompt-based paradigms that exchange raw, human-readable retrieved content, leading to substantial communication overhead and potential privacy risks. To address these limitations, we propose Representation-based Federated RAG (FedRepRAG), a decentralized RAG framework that avoids raw-data exchange during retrieval. FedRepRAG retrieves information from private local document stores and exchanges only compressed latent representations. To integrate the retrieved knowledge, we introduce a lightweight collaboratively trained projector that maps retrieval embeddings into the feature space of a frozen LLM/VLM backbone. This design makes the exchanged information interpretable only within the federated system, thereby enhancing privacy while substantially reducing communication cost. Experiments on decentralized visual question answering (VQA) and question answering (QA) benchmarks show that FedRepRAG consistently outperforms direct inference and local retrieval baselines, while substantially reducing retrieval tokens, computational cost, and inference-time communication overhead compared with raw-context transfer. Overall, FedRepRAG establishes an effective, efficient, and privacy-preserving framework for federated RAG.

Standard Retrieval Augmented Generation (RAG) is poorly matched to agent memory. Unlike large heterogeneous corpora, agent memory forms a bounded and coherent interaction stream in which many spans are highly correlated or near duplicates. As a result, flat top-$k$ similarity retrieval often returns redundant context, while summary-centric hierarchies can blur the subtle details that distinguish one candidate from another. We argue that agent memory should follow the principle of decoupling before aggregation: the system should first isolate reusable facts, updates, and distinguishing details from similar histories, and only then organise them for efficient retrieval. Based on this principle, we propose xMemory, which constructs a revisable hierarchical memory structure from original messages to segments, memory components, and groups. xMemory segments interaction history into local events, decouples each segment into memory components, aggregates related components into high-level groups using a sparsity--semantic faithfulness objective, and maintains this structure incrementally as memory evolves. At inference time, xMemory retrieves top-down, first selecting a compact backbone of complementary groups and components, and then expanding to segments and raw messages only when additional evidence reduces the reader's uncertainty. Experiments on LoCoMo and PerLTQA across diverse open source and closed source LLMs show consistent gains in answer quality and inference token efficiency, supported by analyses of redundancy, evidence density, and coverage.