NeurIPS 2026 Papers

Visual reasoning, often interleaved with intermediate visual states, has emerged as a promising direction in the field. A straightforward approach is to directly generate images via unified models during reasoning, but this is computationally expensive and architecturally non-trivial. Recent alternatives include agentic reasoning through code or tool calls, and latent reasoning with learnable hidden embeddings. However, agentic methods incur context-switching latency from external execution, while latent methods lack task generalization and are difficult to train with autoregressive parallelization. To combine their strengths while mitigating their limitations, we propose ATLAS, a framework in which a single discrete "word", termed as a functional token, serves both as an agentic operation and a latent visual reasoning unit. Each functional token is associated with an internalized visual operation, yet remains a standard token in the tokenizer vocabulary and can be generated via next-token prediction. This design avoids verbose intermediate visual content generation, while preserving compatibility with standard SFT and RL training, without architectural or methodological modifications. To further address the sparsity of functional tokens during RL, we introduce Latent-Anchored GRPO (LA-GRPO), which stabilizes the training by anchoring functional tokens with a statically weighted auxiliary objective, providing stronger gradient updates. Extensive experiments and analyses demonstrate that ATLAS achieves superior performance on challenging benchmarks while maintaining clear interpretability. We hope ATLAS offers a new paradigm inspiring future visual reasoning research.

Graph-structured optimization with linear constraints is fundamental to critical infrastructure but faces scalability limits due to massive strict hard constraints and high dimensionality. While recent projection-based methods such as Trainable Sampling Kaczmarz-Motzkin Net (T-SKM-Net) guarantee feasibility, they face high computational costs in dynamic environments by processing the entire constraint set and requiring expensive matrix factorizations. To bridge this gap, we propose the Accelerated Trainable-SKM (AT-SKM) Net framework. To concentrate computation on the active constraints and eliminate redundant calculations, we introduce a hybrid sampling strategy guided by a topology-aware heterogeneous GNN model. To efficiently handle topological shifts in graph-based constraints, we employ a Cholesky Update mechanism that theoretically reduces the equality projection complexity from $\mathcal{O}(N^3)$ to $\mathcal{O}(N^2)$ under low-rank perturbations. Experiments on random geometric graphs, N-1 Security-Constrained DC-OPF, and minimum-cost gas transport problem demonstrate that AT-SKM reduces iteration counts by up to 85% and achieves 2.95$\times$-7.29$\times$ SKM layer speedups, while maintaining zero constraint violations. Our code is available at [https://anonymous.4open.science/status/Anonymous_Submission-CFC7](https://anonymous.4open.science/status/Anonymous_Submission-CFC7)

Addressing codebook collapse in vector quantization models is crucial for efficient discrete representation learning. Recent solutions increasingly rely on expressive codebook reparameterizations, improving adaptability at the cost of additional optimization and computational overhead. This motivates a simple question: what reparameterization scheme is just enough for efficient codebook learning? In this work, we propose Adaptive-Scale Vector Quantization (ASVQ), a lightweight reparameterization method for learning the codebook in a decoupled way. Experiments across image and audio tokenization tasks demonstrate that ASVQ consistently improves codebook utilization, reconstruction quality, and training stability, while remaining competitive with more complex reparameterization-based methods. These results suggest that ASVQ provides a favorable trade-off between codebook expressiveness and efficiency.

Dynamic topology reconfiguration is central to the reliability and efficiency of large satellite constellations, yet many existing approaches rely on idealized assumptions such as full constellation deployment or uniform orbital spacing. We present Adaptive Satellite Topology via Regret-Aware learning (ASTRA), a theoretically-grounded framework for dynamic satellite topology reconfiguration that builds on an online learning formulation and makes it computationally practical. ASTRA combines an ADMM-based offline solver with efficient online updates for both online gradient descent and online conditional gradient, yielding markedly cheaper constrained updates than generic optimization pipelines. On the theory side, we show that for a relevant class of entry-wise nonzero utility matrices, the objective is strongly convex, which yields logarithmic static regret for online gradient descent, and we further instantiate known dynamic-regret guarantees under inexact ADMM inner loops. Empirically, ASTRA matches or improves topology quality, presenting a good trade-off with computational time on synthetic constellations, and it remains effective on real Starlink data under partial deployment and non-uniform spacing, where idealized structural assumptions break down. These results position ASTRA as an efficient and theoretically grounded approach to topology reconfiguration in realistic Low Earth Orbit networks.

Reinforcement learning (RL) with verifiable rewards has proven effective at post-training LLMs for coding, yet deploying separate task-specific specialists incurs costs that scale with the number of tasks, motivating a unified multi-task RL (MTRL) approach. However, existing MTRL methods treat all coding tasks uniformly, relying on fixed data curricula under a shared optimization strategy, ultimately limiting the effectiveness of multi-task training. To address these limitations, we propose \textbf{{\tool}}, a multi-t\textbf{AS}k code reinforcement learning framework via u\textbf{T}ility-driven co\textbf{OR}dination. Centered on \textit{task utility}, a signal capturing each task's learning potential and cross-task synergy, {\tool} comprises two coupled modules: \textit{1) \moduleone} module hierarchically allocates training budget and prioritizes informative prompts, steering training toward the most valuable data; and \textit{2) \moduletwo} module dynamically scales per-task KL regularization, matching update constraints to each task's current training state. Experiments on two widely-used LLMs across four representative coding tasks demonstrate that {\tool} consistently improves a single model across all tasks, outperforming the best task-specific specialist by 9.0\%--9.5\% and surpassing the strongest MTRL baseline by 7.5\%--12.8\%.

Post-training quantization (PTQ) compresses large language models (LLMs) but remains largely semantic-agnostic, relying on global activation magnitude to guide precision allocation. This overlooks the functional heterogeneity of Transformer representations, where semantically meaningful channels are sparsely activated while structurally dominant channels exhibit large but context-insensitive activations. We propose Agent-guided Semantic-aware Quantization (ASQ), a PTQ framework that conditions on semantically informative positions identified by an auxiliary LLM. A dual-pathway scoring mechanism captures both activation selectivity and attention routing toward these positions, producing channel-wise importance scores for scaling. This yields a closer approximation to a semantically conditioned objective and consistently improves low-bit (e.g., W4A16) quantization performance across LLMs. Extensive experiments show consistent gains over state-of-the-art PTQ, preserving general reasoning and even surpassing FP16 performance on several benchmarks in specialized domains, while enabling training-free, privacy-preserving self-quantization with zero inference overhead.

Clarification-seeking behavior is widely regarded as a desirable property of LLM agents, enabling them to resolve ambiguity before acting on underspecified tasks. However, the security implications of this interaction pattern remain unexplored. We investigate whether the transition from standard execution to a clarification-seeking state increases an agent's susceptibility to prompt injection attacks. We introduce **ASPI** (**A**mbiguous-**S**tate **P**rompt **I**njection), a benchmark of 728 task-attack scenarios that isolates clarification as a distinct agent state and measures how this state transition affects vulnerability under controlled conditions. Each benchmark instance is evaluated under matched execution and clarification settings: in the execution setting, the agent acts on a fully specified instruction and encounters adversarial content only through tool-returned data; in the clarification setting, the agent must first request and incorporate additional user input before acting. We evaluate ten frontier LLMs and find that clarification-seeking consistently and substantially amplifies vulnerability. For instance, attack success rises from 1.8\% to 34.0\% for o3 and from 2.2\% to 35.7\% for Gemini-3-Flash. A decomposition analysis reveals that this gap reflects both a state-dependent shift in how models process incoming content and a channel-specific effect arising from the agent-solicited clarification interface. These findings demonstrate that standard execution-time security evaluation systematically underestimates the attack surface of interactive agents, and that robustness under fully specified tasks does not translate to robustness under ambiguity. For reproducibility our data and source code is available at https://anonymous.4open.science/r/aspi-5074.

Predictive models have the potential to transform drug development by increasing success rates across all stages of development pipelines. Evaluating this potential needs to include the whole pipeline, from initial screens of thousands of compounds to the expensive assays, such as toxicity assessment, focused on a few compounds in the later stages. Genetic medicines, such as antisense oligonucleotides (ASOs), could benefit millions affected by individually rare genetic disorders, if economic limitations imposed by the need to screen hundreds of sequences at considerable expense can be overcome. To enable systematic analysis of this problem, we present ASO Atlas 2.0, the largest public multi-endpoint ASO preclinical dataset, spanning 295,007 data points, including both in vitro efficacy and in vivo toxicology, across 430. Using this resource, we introduce a cost-based evaluation framework that translates per-stage classifier enrichment into projected cost savings under a representative preclinical pipeline model. We find that incorporating model predictions across all stages reduces projected pipeline cost by 71%. This saving is largely attributable to modest enrichment at costly late-stage in vivo studies rather than strong enrichment at lower-cost in vitro assays, reversing the model ranking suggested by accuracy alone and highlighting the need for improved toxicology modelling.

Long-horizon embodied tasks remain a fundamental challenge in AI, as current methods rely on hand-engineered rewards or action-labeled demonstrations, neither of which scales. We introduce ASH, an agentic system that learns an embodied policy from unlabeled, noisy internet video, without reward shaping or expert annotation. ASH follows a self-improvement loop; when it gets stuck, ASH learns an Inverse Dynamics Model (IDM) from its own trajectories, and uses its IDM to extract supervision from relevant internet video. ASH uses unsupervised learning to identify key moments from large-scale internet video and retains them as long-term memory --- allowing it to tackle long-horizon problems. We evaluate ASH on two complementary environments demanding multi-hour planning: Pokemon Emerald, a turn-based RPG, and The Legend of Zelda: The Minish Cap, a real-time action-adventure game. In both games, behavioral cloning, retrieval-augmented and zero-shot foundation-model baselines plateau, while ASH sustains progression across our 8-hour evaluation. ASH reaches an average of $11.2/12$ milestones in Pokemon Emerald and $9.9/12$ in Legend of Zelda, while the strongest baseline gets stuck in both environments at an average of $6.5/12$ and $6.0/12$ milestones, respectively. We demonstrate that self-improving agents are a scalable recipe for long-horizon embodied learning.

Federated Learning (FL) is vulnerable to model poisoning attacks, where malicious clients manipulate uploaded model updates to corrupt the global training process. Existing attacks are typically evaluated by their eventual damage, such as final accuracy degradation or convergence to random-guess performance, where the learned model becomes unusable and may ultimately lead to denial-of-service (DoS). However, such metrics overlook attack round complexity, defined as the number of communication rounds required for the poisoned global model to reach and remain near a desired degradation objective. In realistic FL systems, requiring more communication rounds increases the adversary's exposure to client participation, robust aggregation, filtering, and detection. We propose ASAP (Adaptive Sliding Agnostic Poisoning), a fast aggregation-rule-agnostic (AGR-agnostic) attack that treats the whole FL training process as an uncertain dynamical system and formulates poisoning as a feedback-control problem for steering the model toward a desired attack objective. ASAP combines adaptive sliding mode control (ASMC) with finite Fourier-basis uncertainty estimation, treating the unknown effects of benign training and aggregation as a time-varying uncertainty and using the resulting estimate to design the malicious updates. We theoretically prove that the sliding variable reaches the sliding surface in finite time and the tracking error subsequently converges exponentially. Experiments across multiple datasets, models, and aggregation rules demonstrate that ASAP reaches specified degradation objectives in fewer communication rounds while maintaining smaller target deviation than existing model poisoning attacks.

Vision Transformers (ViTs) face severe computational bottlenecks due to the quadratic complexity of self-attention at high resolutions. Existing token reduction methods rely on local metrics—such as single-layer attention scores—that are inherently vulnerable to the \emph{attention sink} phenomenon, where uninformative tokens are paradoxically preserved over salient foreground objects. We propose \textbf{ASAP} (Attention Sink Anchored Pruning), a training-free framework that recasts this sink as a feature. Modeling ViT information flow as a Lazy Random Walk, ASAP identifies the sink as a dominant accumulator of probability mass. By computing the \emph{diffusion distance} to the sink within the cumulative transition matrix, ASAP partitions tokens via \emph{Radial Diffusion Clustering} and compresses background redundancy through \emph{Transition Weight Pooling} in a single shot. Extensive experiments across image, video, and vision-language tasks demonstrate ASAP outperforms state-of-the-art methods, accelerating throughput by up to 48\% while maintaining—or even exceeding—baseline accuracy.

Large language models (LLMs) are increasingly used in binary decompilation to refine the C-like pseudocode produced by traditional rule-based decompilers. While this pseudocode is useful, it is a heuristic and lossy abstraction rather than a faithful copy of the source code. It often contains decompiler errors, especially for aggressively optimized binaries where critical low-level details are obscured. We present ASAP, an assembly-source aligned pseudocode refinement framework for binary decompilation. ASAP learns source-aligned assembly representations from paired source and binary functions using joint function-level and snippet-level contrastive alignment. A Q-Former then compresses chunk-level assembly features into a fixed number of assembly tokens that condition the decompilation LLM alongside the decompiler-produced pseudocode. During refinement, we use stochastic pseudocode masking and a relative assembly-advantage loss to reduce the model's tendency to ignore assembly features and rely only on pseudocode refining. On two decompilation benchmarks across multiple compiler optimization levels, ASAP improves the average re-execution rate from 64.7% to 71.9% and the average recompilation rate from 91.6% to 96.6% compared with the strongest baseline, offering both a new perspective and a practical solution to binary decompilation.

Current test-time scaling (TTS) techniques enhance large language model (LLM) performance by allocating additional computation at inference time, yet they remain insufficient for agentic settings, where actions directly interact with external environments and their effects can be irreversible and costly. We propose ARTIS, Agentic Risk-Aware Test-Time Scaling via Iterative Simulation, a framework that decouples exploration from commitment by enabling test-time exploration through simulated interactions prior to real-world execution. This design allows extending inference-time computation to improve action-level reliability and robustness without incurring environmental risk. We further show that naive LLM-based simulators struggle to capture rare but high-impact failure modes, substantially limiting their effectiveness for agentic decision making. To address this limitation, we introduce a risk-aware tool simulator that emphasizes fidelity on failure-inducing actions via targeted data generation and rebalanced training. Experiments on multi-turn and multi-step agentic benchmarks demonstrate that iterative simulation substantially improves agent reliability, and that risk-aware simulation is essential for consistently realizing these gains across models and tasks.

We consider time discretization for score-based diffusion models to generate samples from a learned reverse-time dynamic on a finite grid. Uniform and hand-crafted grids can be suboptimal given a budget on the number of time steps. We introduce Adaptive Reparameterized Time (ART), which controls the clock speed of a reparameterized time variable to redistribute computation along the sampling trajectory while preserving the terminal time, with the objective of minimizing the aggregate Euler discretization error. We derive a randomized companion ART-RL that recasts ART as a continuous-time reinforcement learning problem with Gaussian policies, and prove a two-directional bridge between the two: the deterministic ART optimum lifts to an optimal Gaussian policy, and conversely any optimal Gaussian policy must recover the ART control through its mean. This bridge turns continuous-time actor--critic learning into a principled, rather than heuristic, route to the deterministic timestep optimum. Within the official EDM pipeline, ART-RL improves FID on CIFAR--10 across a wide range of budgets; after one-time offline training, the distilled deterministic schedule transfers without retraining to AFHQv2, FFHQ, and ImageNet at no extra inference cost.

Existing multimodal retrieval benchmarks largely emphasize semantic matching on daily-life images and offer limited diagnostics of professional knowledge and complex reasoning. To address this gap, we introduce ARK, a benchmark designed to analyze multimodal retrieval from two complementary perspectives: (i) knowledge domains (five domains with 17 subtypes), which characterize the content and expertise retrieval relies on, and (ii) reasoning skills (six categories), which characterize the type of inference over multimodal evidence required to identify the correct candidate. Specifically, ARK evaluates retrieval with both unimodal and multimodal queries and candidates, covering 16 heterogeneous visual data types. To avoid shortcut matching during evaluation, most queries are paired with targeted hard negatives that require multi-step reasoning. We evaluate 26 representative text-based and multimodal retrievers and observe a pronounced gap between knowledge- and reasoning-intensive retrieval, with fine-grained visual and spatial reasoning as persistent bottlenecks. We further show that enhancements such as re-ranking, rewriting, and agentic retrieval yield consistent gains, but substantial headroom remains. The dataset and evaluation pipeline will be released soon.

Zeroth-order (ZO) optimization enables large-language-model fine-tuning without storing backpropagation activations, while LoRA supplies compact trainable adapters. Combining them creates a rank paradox: increasing LoRA rank improves adapter capacity, but standard two-point ZO either perturbs a rank-dependent number of coordinates or, under atomwise updates, can make the finite-difference signal unobservable. This paper shows that the bottleneck is a measurement-topology problem rather than a need for an external subspace. LoRA already decomposes into matched rank-$1$ atoms, each a complete factor-coordinate block of dimension $d_\text{out}+d_\text{in}$. Querying one atom per step keeps the stored adapter rank $r$ while removing $r$ from the single-query perturbation dimension. The naive atomwise query is still miscalibrated: if it inherits canonical LoRA scaling $\alpha/r$, the active finite-difference signal shrinks as $1/r$ and the active finite-difference signal-to-noise ratio (FD-SNR) as $1/r^2$, producing directional collapse under a fixed residual evaluation-noise floor. AR1-ZO pairs alternating rank-$1$ atom queries with topology-aware scaling $\gamma=\alpha r$, restoring rank-invariant active signal without auxiliary bases, activation hooks, curvature estimates, or extra forward queries. Theory proves atom minimality, rank-independent active query dimension, directional collapse and restoration, and the remaining rank dependence as an amortized coverage cost. Experiments on OPT and Qwen3 models validate the signal mechanism and show that AR1-ZO makes high-rank LoRA effective among matched-budget ZO methods under the standard two-forward-pass query budget.

Editing a video stream in real time, without access to future frames, is essential for interactive applications. Existing training-free video editing methods, however, assume full offline access to the entire sequence, while streaming approaches either rely on additional training or are limited to specific editing techniques. We show that this limitation can be removed by analyzing the self-attention dynamics of DMD-distilled autoregressive video diffusion models. We uncover a key asymmetry: queries and keys remain stable and aligned with their source counterparts across denoising timesteps, while values diverge and encode the evolving content. This observation leads to a simple mechanism: a single forward pass suffices to capture source identity in the value features, which can be reused throughout generation. Starting from pure Gaussian noise and injecting only these source values reconstructs the input with fidelity matching inversion-based methods, effectively eliminating inversion. Based on this insight, we introduce AR-Edit, a training-free, inversion-free method that caches source values once and selectively injects them during autoregressive denoising, enabling real-time, structure-preserving streaming video editing with minimal overhead. Code will be released.

Neural surrogate benchmarks for spatiotemporal PDEs are dominated by idealized test beds with periodic boundaries and synthetic dynamics. Architecture rankings established on these benchmarks do not transfer to real atmospheric chemistry, and standard regression metrics conceal the failure modes that determine operational utility. AQBench evaluates ten architectures across five model families on high-resolution WRF-Chem simulations of $PM_2.5$, $NO_2$, and $CO$ over the Indian subcontinent, against five operationally-grounded objectives: long-horizon stability, exceedance detection, extreme-episode bias, advection-diffusion residuals, and multi-pollutant forecasting. Three findings emerge. Spectral neural operators, the strongest family on canonical PDE benchmarks, rank in the bottom four on 168 hour RMSE here. Models that lead on aggregate regression underestimate during extreme pollution episodes and miss the majority of true exceedance events. Joint training on $PM_2.5$, $NO_2$, and $CO$ improves $PM_2.5$ accuracy across most architectures with no parameter increase, with Transolver gaining 21\% at 168th hour. The dataset, evaluation framework, and trained baselines are released.

Typical LLM responses tend to follow a default style, even though users often have distinct preferences regarding tone, verbosity, and formality that they do not explicitly state in their prompts. Evaluating whether personalization methods can adapt to these implicit preferences is challenging, since users typically provide prompts rather than reference responses, style preferences are not factually verifiable, and reference-free LLM judges may conflate personalization with general response quality. To address these challenges, we introduce the Arbitrary Preference Mapping (APM) benchmark, which decouples user attributes (e.g. enthusiastic) from response principles (e.g. persuasive) via a hidden, randomized mapping $\mathbf{C}$ that maps user attributes to preferences about response traits. Because $\mathbf{C}$ carries no semantic content and is resampled across runs, models cannot exploit stereotypical associations and must infer preferences from conversation history. Using this unbiased evaluation methodology, we adapt retrieval-augmented, prompt-optimization, and routing personalization methods and evaluate them on Llama-3.1-8B and Qwen-3.5-27B. Our results show that routing is the most reliable approach, while RAG benefits only from the stronger base LLM, and soft prompt optimization fails to improve significantly over a non-personalized baseline. Our extensive evaluation reveals that in this realistic setting, personalization remains challenging, but our adapted methods show promise.