NeurIPS 2026 Papers

Flow-based vision-language-action (VLA) models provide expressive action generation for embodied control, but supervised fine-tuning (SFT) often confines them to narrow expert behaviors. Online reinforcement learning (RL) can improve beyond demonstrations through environment interaction, yet flow-based VLAs struggle with sparse long-horizon rewards and intractable action likelihoods. We propose \textbf{\emph{LiFT}}, a likelihood-free tree policy optimization framework for flow-based VLAs. LiFT first addresses sparse-reward credit assignment by creating sibling continuations from shared rollout histories and using their final outcomes to compute branch level relative advantages. LiFT further uses local action-flow instability as an expansion criterion, focusing rollout budget on ambiguous action-generation regions that are most likely to reveal meaningful outcome differences. Finally, LiFT applies branch-level advantages through a decision-level surrogate ratio on executed action chunks, matching policy updates to the granularity of tree-based credit assignment. This yields a likelihood-free policy update without attributing chunk-level credit to individual denoising steps. Evaluations on in-distribution and out-of-distribution benchmarks show that LiFT improves task success and robustness to distribution shifts.

High-resolution 3D medical image generation remains challenging because fully volumetric models are computationally expensive, while efficient 2D slice generators often fail to preserve anatomical consistency across the third dimension. We propose LiFT, a framework for Lifted Inter-slice Feature Trajectories that factorizes 3D volume synthesis into per-slice image generation and inter-slice trajectory learning. Rather than modeling the volumetric distribution end-to-end, LiFT treats a volume as an ordered trajectory in feature space, capturing how anatomical structures appear, transform, and disappear across depth. A tri-planar drifting loss aligns the trajectory of generated slices with the trajectories of real volumes, enabling distributional learning over inter-slice progressions in unconditional generation; in paired translation, a bidirectional $z$-context mixer trained against the registered target supplies through-plane coherence while preserving per-slice fidelity. We evaluate LiFT on BraTS 2023 (unconditional and missing-modality MR) and SynthRAD2023 (MR-to-CT). Across these settings, LiFT preserves per-slice quality, approaches the reported cWDM missing-MR reconstruction quality, and improves through-plane coherence on MR-to-CT relative to a no-mapper ablation, demonstrating that lightweight inter-slice trajectory learning is a viable route to high-resolution 3D medical synthesis.

Robust and accurate neural decoders are integral to the development of neurotechnological applications such as brain-computer interfaces and closed-loop experiments. Recent work has shown that tokenizing neural data at the resolution of individual spikes facilitates multi-session pretraining and delivers state-of-the-art decoding performance. However, these spike-based models are currently restricted to supervised learning (SL) regimes, limiting both pretraining and finetuning to datasets with paired behavioural labels. To address this, we introduce MOJO ($\textbf{M}$asked aut$\textbf{O}$encoder-based $\textbf{JO}$int training), a training framework designed for spike-tokenizing models that jointly leverages self-supervised learning (SSL) via masked autoencoding and SL objectives for better decoding performance and interpretability. We evaluate MOJO on three intracortical spiking datasets—monkey motor cortex across various reaching tasks, multi-regional mouse recordings during vision and cognitive decision making tasks—demonstrating superior performance to purely SL-trained models. This improvement is especially notable when training with limited labelled data, specifically in the few-shot finetuning regime where only a small amount of labelled data is available to adapt a model to a new recording session. The incorporation of SSL also yields more interpretable neuronal representations, improving performance on analyses such as brain region classification and spike-statistics prediction despite the lack of explicit optimization for these tasks. We then show that MOJO generalizes beyond spiking data using human electrocorticography (ECoG) during speech articulation. We show that MOJO continues to outperform purely SL-trained models and achieves performance comparable to neuro-foundation models (NFMs) built specifically for continuous signals. Overall, augmenting spike-tokenizing models with SSL improves their performance in label-impoverished settings and enables the use of unlabelled data across various tasks and species, and while generalizing to other neural modalities. These results suggest a path towards more flexible and scalable data usage when training NFMs.

Latent visual reasoning involves visual evidence more directly in multimodal reasoning by inserting continuous latent tokens before textual generation. However, the necessity of these latent tokens at inference remains ambiguous. We show that replacing latent tokens with random noise or removing them completely causes little performance degradation across spatial reasoning benchmarks. Reinforcement learning further diminishes the latent generation behavior after post-training. These observations raise a central question: $\textit{Is latent visual reasoning still meaningful}$? We argue that its value should be measured by how effectively latent tokens guide learning, rather than whether they persist as an inference-time format. Our analysis shows that latent reasoning is unevenly favorable across question types, yet hard task-level routing for applying latent generation is brittle. Motivated by these findings, we propose an attention-based reward that encourages generated latent tokens to interact with later text tokens during RL. This reward promotes latent utilization when the latent mode is activated while preserving the flexibility to use pure-text reasoning. Experiments show that our method improves performance across perception and visual reasoning benchmarks, even when latent tokens are rarely generated after post-training. Our results highlight that, without explicit expression at inference, latent visual reasoning can shape better visual grounding and more accurate textual reasoning $\textbf{in silence}$. Our code and trained models are publicly available at Hugging Face https://huggingface.co/collections/doubleblindsubmission/submission3847.

Vision-language-action models (VLA) promise human-steerable autonomous driving, but their training is bottlenecked by the scarcity of frames paired with natural-language instructions: while camera streams and expert trajectories are logged at scale, language annotations (e.g., "turn left at the intersection") remain scarce and expensive to acquire. To address this challenge, we introduce Latent Action Driving Annotations (LADA), a three-stage pipeline that transforms abundant unlabelled observation–trajectory pairs into a substrate for language-conditioned control. First, we train a latent action model with a vector-quantised bottleneck, producing a compact codebook of high-level vehicle intents. Second, a small language-annotated subset is used to train a vision-language translator to map observations and language instructions into this codebook. Third, we train a driving VLA on observation–latent-action pairs over the full unlabelled corpus. Using fewer than 5% of language annotations and without leveraging any auxiliary chain-of-thought reasoning or visual question answering streams, LADA achieves a Driving Score of 87.98 and a Success Rate of 70.46% on the closed-loop Bench2Drive benchmark, matching or surpassing fully supervised baselines.

Vision Language Models (VLMs) offer the exciting possibility of processing text as rendered images, bypassing the need for tokenizing the text into long token sequences. Since VLM image encoders map fixed-size images to a fixed number of visual tokens, varying rendering resolution provides a fine-grained compression knob. However, accuracy deteriorates quickly as compression increases: characters shrink below the vision encoder's effective resolution, making them indistinguishable. To address this, we propose LensVLM, an inference framework and post-training recipe that enables VLMs to scan compressed images, then selectively expand only the relevant images to their uncompressed form via learned tools. Building on Qwen3.5-9B-Base, LensVLM maintains accuracy comparable to the full-text upper bound at 4.3$\times$ effective compression and outperforms retrieval-based, text- and visual-compression baselines up to 10.1$\times$ effective compression across seven text QA benchmarks. LensVLM also generalizes to multimodal document and code understanding tasks, with the accuracy gain over baselines growing as compression increases. Our analysis validates this approach: training makes visual compression robust to rendering choices, and as compression grows the model increasingly relies on expanded content rather than unreliable visual reading. The analysis also yields practical tool-choice guidance: text expansion is preferable for rendered text, while high-resolution image expansion suits native documents whose layout cues carry task-relevant information.

Web agents promise to automate consequential operations on enterprise web systems, yet the benchmarks used to evaluate them score correctness on surface-level proxies: DOM/URL predicates, action-trace matchers, or vision-language verdicts on a closing screen. These proxies are inadequate for enterprise resource planning (ERP) systems, where every action commits to a transactional ledger and the criterion of correctness is what the ledger now records, not what the page now renders. We argue that ERP web operations should be evaluated against what a run actually wrote to the relational backend, with the server-side operation trace recorded alongside the outcome to provide the audit provenance that enterprise governance requires. We introduce Ledger, a path-validated, database-grounded benchmark for browser agents on a production-grade open-source ERP: 424 oracle-certified task instances across 53 functional modules, scored by an Atomic Edit Set validator (field-level $F_1$ against a PostgreSQL oracle, under a typed normalization function) paired with a Canonical Operation Tracer (server-side instrumentation of the ERP’s JSON-RPC dispatch). Each episode runs under pg_restore snapshot isolation; no language-model judge is invoked; three formal guarantees (score determinism, episode isolation, tracer completeness) make scores auditable and exactly replayable. A cross-paradigm evaluation of seven contemporary browser agents shows Ledger discriminates architectures sharply, with task-success rate spanning 81.6% down to 13.7% across the roster while long-horizon enterprise workflows remain far from saturated.

Boolean satisfiability (SAT) provides a foundation tool for cryptographic key recovery by encoding ciphers into CNF or ANF representations. However, heuristic solvers and their neural-enhanced variants often degrade significantly on cryptographic instances due to pronounced structural symmetry, a large number of intermediate variables, and complex algebraic dependencies. Existing neural approaches—ranging from end-to-end prediction to solver-integrated heuristics—face challenges in scalability, assignment accuracy, or computational efficiency. In this paper, we propose a unified neural-guided enumerative SAT framework for cryptographic key recovery. It performs a single forward neural guidance to identify a subset of $k$ critical variables, followed by enumeration over their assignments before invoking a heuristic SAT solver. This design effectively reduces the combinatorial search space while keeping neural overhead minimal. We further introduce a taxonomy of neural guidance across data and model capability regimes, including supervised selection with cryptographic priors, uncertainty-driven selection with assignment prediction, and robust fallback strategies using the model trained on general SAT datasets. Experiments on 10 SAT solvers and 2 cube-and-conquer strategies over SAT4CryptoBench and SAT Competition benchmarks demonstrate up to $5\times$ speedup and approximately 2× average improvement on cryptographic instances, while maintaining better performance on general SAT datasets. These results highlight the effectiveness and generalization of this framework, illustrating its potential to bridge machine learning and SAT solving in cryptanalysis tasks.

Experience-driven learning has emerged as a promising paradigm for enabling agents to improve from interaction trajectories by accumulating and reusing past experience. However, existing approaches are predominantly developed in textual settings and rely on manually designed memory schemas, limiting their applicability to multimodal environments. In real-world scenarios, experience is inherently multimodal, involving heterogeneous signals across perception, reasoning, and action, which makes effective memory design significantly more challenging. In particular, the optimal way to structure and utilize multimodal experience is highly task-dependent and evolves over time, rendering fixed memory designs insufficient. In this work, we propose a new paradigm, learning to learn from multimodal experience, which shifts memory design from a predefined component to an adaptive and learnable process. Our framework enables agents to dynamically construct, organize, and utilize memory based on task requirements and interaction history, effectively learning how to structure experience for improved performance. Experiments demonstrate that adaptive memory design substantially enhances agent performance and generalization across multimodal tasks, highlighting the critical role of learning memory mechanisms in experience-driven learning.

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

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

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

Embodied LLMs endow robots with high-level task reasoning, but they cannot reflect on what went wrong or why, turning deployment into a sequence of independent trials where mistakes repeat rather than accumulate into experience. Drawing upon human reflective practitioners, we introduce Reflective Test-Time Planning, which integrates two modes of reflection: \textit{reflection-in-action}, where the agent uses test-time scaling to generate and score multiple candidate actions using internal reflections before execution; and \textit{reflection-on-action}, which uses test-time training to update both its internal reflection model and its action policy based on external reflections after execution. We also include retrospective reflection, allowing the agent to re-evaluate earlier decisions and perform model updates with hindsight for proper long-horizon credit assignment. Experiments on our newly-designed Long-Horizon Household benchmark and MuJoCo Cupboard Fitting benchmark show significant gains over baseline models, with zero-shot generalization to photorealistic HM3D environments and real-robot experiments on a Franka Panda arm. Ablations confirm that reflection-in-action and reflection-on-action are mutually dependent, and that retrospective reflection achieves better credit assignment than step-wise external feedback at lower computational overhead. Qualitative analyses further highlight behavioral correction through reflection.

Real-world degraded images often contain multiple co-occurring degradation types, making composite image restoration fundamentally different from the single-degradation setting assumed by most existing all-in-one methods. These methods typically apply uniform spatial computation and single-label task conditioning, limiting both spatial adaptivity and explicit modeling of degradation mixtures. We propose **CART** (**C**omposite-**A**daptive **R**outing and **T**ask Conditioning), a unified framework that jointly decides *where* to spend computation and *what* degradation cues should guide restoration: a spatial mixer routes patches to graded-capacity experts by local restoration difficulty, and a channel mixer routes each image through degradation-specific experts, conditioned on a global task feature with multi-hot classification supervision. Empirically, CART achieves state-of-the-art performance for composite degradation restoration on CDD-11 and delivers the best results to date on the conventional 3-task and 5-task all-in-one benchmarks.

Reasoning traces improve vision-language models on complex multimodal tasks, but forcing every input to follow a long reasoning process leads to overthinking and unnecessary inference cost. We study adaptive fast/slow thinking for VLMs, where a single model should answer directly for simple inputs and reason step by step only when needed. The key observation is that offline adaptive thinking is not merely a length-control problem: fast direct answers and slow reasoning traces are typically generated by distinct behavior policies. This dual-reference structure makes standard offline reward optimization, such as Decoupled Generation and Optimization (DGO), mismatched because it assumes a single reference policy and can misweight one mode. We propose \textsc{Dual-Reference DGO}, which models fast and slow responses under a fast/slow mixture reference and derives a reward-weighted offline objective for adaptive VLM reasoning. Our formulation shows that router-free fast/slow switching emerges from competition between fast and slow partition functions, and yields an efficient mode-wise correction for behavior-reference mismatch. Experiments on multiple datasets show that \textsc{Dual-Reference DGO} improves the accuracy--length trade-off over baselines. Ablations show that single-reference DGO tends to collapse to almost always-fast or always-slow behavior, while our dual-reference formulation preserves both modes and learns difficulty-dependent routing.

Multi-agent systems have emerged as a promising paradigm for multimodal medical reasoning, but existing methods rely on two assumptions that hinder practical deployment: (i) always-active dense collaboration among multiple agents, and (ii) dependence on large cloud-based models that raise privacy and efficiency concerns. In this paper, we challenge both assumptions and propose SMART-Med, a fully local multi-agent method that treats collaboration as a learnable decision rather than a default behavior. Our key insight is that not every medical query requires costly multi-agent deliberation: many can be reliably solved by a single well-trained agent, while only uncertain cases benefit from expert collaboration. Based on this insight, SMART-Med introduces an uncertainty-driven routing mechanism that adaptively decides when to answer directly and when to recruit a small subset of expert agents, effectively converting dense multi-agent reasoning into selective, query-adaptive coordination. To make this routing reliable, we curate a new multimodal medical dataset MedVR-3K with high-quality reasoning traces and a difficulty hierarchy, and then fine-tune local agents using GRPO with two complementary rewards: a certainty-aware accuracy reward and a vision-language reward model for correctness and reasoning quality respectively. Experiments on four medical visual question answering benchmarks show that SMART-Med achieves competitive performance compared to existing MAS approaches while substantially reducing inference cost (reducing inference time by more than 3.8$\times$ and token usage by 72\%). Moreover, after fine-tuning on MedVR-3K, SMART-Med yields an average improvement of 6.7\% across Slake, PathVQA, VQA-RAD, and PMC-VQA. Our results suggest that selective collaboration can be an efficient yet effective alternative to dense collaboration and enable practical, privacy-preserving deployment of multi-agent medical AI.

Open-vocabulary dense perception requires grounding language-specified concepts in local regions for recognition, localization, and segmentation beyond closed category sets. CLIP learns image-level vision-language alignment, whereas dense prediction requires spatially precise and semantically reliable local features. Although recent dense CLIP adaptation methods reduce this gap through region-level semantic transfer and spatial correlation guidance, maintaining stable local semantics under contextual variation remains challenging. Specifically, the global contextual modeling in CLIP may introduce undesirable dependencies into local features. Changes in background, scale, viewpoint, or scene layout may cause identical object regions to yield inconsistent dense representations, leading to unstable correspondences and semantic drift. Moreover, spatial correlation guidance from visual foundation models captures visual relatedness between patches, but not whether coherent regions correspond to the queried textual concept, leaving concept-level dense alignment under-constrained. To address these issues, we introduce Semantic Spatial Consistency Learning $\textbf{(SSCL)}$, an adaptation method with two complementary objectives: 1) Context-Invariant Spatiotemporal Consistency (CISC) enforces semantic consistency for the same anchor region under varying contexts to mitigate context-induced feature drift; 2) Structure-Guided Semantic Refinement (SGSR) leverages spatial affinities from a frozen visual foundation model to refine CLIP text-patch responses, producing text-conditioned dense supervision with improved spatial coherence and concept-level alignment. Experiments on video instance segmentation, region classification, semantic segmentation, and object detection show consistent gains over the baseline methods, demonstrating the effectiveness and generalization capability of the proposed SSCL. Our code and models will be made publicly available.

Human videos contain rich manipulation priors, but using them for robot learning remains difficult because raw observations entangle scene understanding, human motion, and embodiment-specific action. We introduce MoT-HRA, a hierarchical vision-language-action framework that learns human-intention priors from large-scale human demonstrations. We first curate HA-2.2M, a 2.2M-episode action-language dataset reconstructed from heterogeneous human videos through hand-centric filtering, spatial reconstruction, temporal segmentation, and language alignment. On top of this dataset, MoT-HRA factorizes manipulation into three coupled experts: a vision-language expert predicts an embodiment-agnostic 3D trajectory, an intention expert models MANO-style hand motion as a latent human-motion prior, and a fine expert maps the intention-aware representation to robot action chunks. A shared-attention trunk and read-only key-value transfer allow downstream control to use human priors while limiting interference with upstream representations. Experiments on hand motion generation, simulated manipulation, and real-world robot tasks show that MoT-HRA improves motion plausibility and robust control under distribution shift.

Despite the effectiveness of Cube-and-Conquer (C&C) for solving challenging Boolean Satisfiability (SAT) problems, no prior work has shown that transformer-based models can learn effective cubing heuristics. We introduce a neuro-symbolic post-training framework for this task. We design an MCTS-based data curation pipeline that uses symbolic heuristics to explore splitting decisions over SAT competition formulas, producing preference data grounded in solver statistics and augmented with reasoning traces from a teacher model. Our two-stage post-training, supervised fine-tuning (SFT) followed by direct preference optimization (DPO), enables a 4B-parameter model to achieve a pass@5 score of 53 on 100 SAT competition benchmarks, surpassing frontier LLMs such as Claude-Sonnet-4 (50) and matching the best symbolic heuristic (53). Ablations show that SFT alone improves pass@5 from 46 to 51, with DPO adding 2 additional benchmarks; an entropy/agreement ablation on realized first-cube decisions further shows that SFT, not DPO, accounts for the root-level decision diversity that produces complementary per-run coverage over deterministic symbolic methods. This demonstrates that transformers can be trained to make effective cubing decisions in a domain traditionally dominated by symbolic methods.