Science & Mathematics

Visual tool use enables vision-language models (VLMs) to acquire task-relevant evidence, but its effectiveness can be limited by predefined visual operations and the tool-use strategies learned during training. Representing these operations and strategies as reusable external skills allows them to be refined through self-evolution without updating model parameters. However, skill refinement can introduce performance regressions. Meanwhile, inspecting accumulated task execution trajectories together with skill evolution history may lead to evolution evidence overload. To address these challenges, we propose VisualEvolve, a self-evolving framework that evolves executable visual operations and tool-use strategies for VLMs through a Select--Execute--Evolve--Evaluate cycle. Skill-Library Snapshot Selection uses performance on a fixed validation set disjoint from execution tasks to select an effective historical snapshot for subsequent evolution, limiting the propagation of performance regressions. Progressive Evidence Access (PEA) guides capability-gap diagnosis through a coarse-to-fine hierarchy comprising the Skill-Performance Report, Evolution Log, Trajectory Summaries, and relevant raw trajectory segments. Across four VLM backbones and six evaluation sets from three benchmarks, VisualEvolve improves over the Initial Skill Library in 23 of 24 settings, with a maximum gain of 15.1 percentage points on VisualProbe Hard using Qwen3.5-Plus. Ablation studies support the contributions of Snapshot Selection and PEA, while further experiments show that skills can be improved using experience from different data settings and directly reused across backbones. Moreover, VisualEvolve enables effective skill self-evolution with over an order of magnitude fewer task-solving trajectories than estimated for representative RL-based methods.

LLM agents are commonly trained with RLVR, but the sparse nature of typical RLVR rewards makes credit assignment challenging, especially for long-horizon tasks. We propose Self-Credit, an end-to-end agentic RL method that creates dense supervision signals from sparse outcome rewards using self-generated credit assignment via intra-group rescaling. Self-Credit builds on group-based RL with three components. First, during its rollouts, the policy model generates summaries with citations to earlier turns containing critical evidence, which serve as self-generated credit. Second, we refine this model-generated credit with intra-group rescaling, which compares cited actions across successful and failed rollouts for the same prompt. This approach uses group-level agreement and disagreement to remove undeserved credit and further emphasize credit assigned to critical actions. Third, advantage reweighting uses the rescaled credit weights to update the advantages of the cited turns. On BrowseComp-Plus, Self-Credit improves over the base models by 37.0 and 36.2 points on Qwen3 8B and 32B respectively, and outperforms the strongest agentic RL baselines by 8.5 and 5.6 points, respectively. Self-Credit also improves over the strongest baselines by 4.9 points on GAIA and 4.7 points on xBench-DeepSearch. Analyses show that self-generated credit provides useful signals for selecting critical actions, and that intra-group rescaling, especially on failed rollouts, is important across model scales.

LLM-driven evolutionary search is a promising paradigm for scientific discovery, but existing methods remain costly and slow, often applying expensive reasoning uniformly while underusing evidence from failed, redundant, and saturated trials. We introduce Speculative Evolution (SpecEvo), a multi-LLM framework that coordinates lightweight and frontier models through continual consolidation of search evidence. Lightweight models power both parallel Speculators, which generate and refine candidate programs, and an Advisor, which distills accumulated successes, failures, and saturated directions into persistent guidance for subsequent exploration. At sparse checkpoints, a frontier-model Navigator inspects the evolving population and intervenes through synthesis, surgical refinement, or reframing according to search stagnation. This speculate-then-consolidate workflow delegates routine exploration and evidence consolidation to inexpensive models while reserving frontier reasoning for strategic interventions whose benefits extend across many subsequent low-cost evaluations. Across 176 tasks spanning four scientific and algorithmic discovery domains, SpecEvo matches or outperforms frontier-model baselines while reducing API cost by up to 80%, achieves up to 1.5x greater evaluation efficiency than cost-efficient baselines on equation discovery, and improves over the strongest CO-Bench baseline by 3.8%. These results support trajectory-conditioned model coordination as an effective approach to cost-efficient scientific discovery.

A cluster admission gate can receive a capacity record that no longer matches current schedulability. We study one exogenous binary schedulability unit under a bound on the joint probability of admission and unavailability. For a symmetric Markov state, we derive the stationary admission frontier for a fixed observation delay and apply finite likelihood allocation to a delay-tagged delivered record. A one-sided exact binomial endpoint calibrates each fitted policy, and exact enumeration confirms the per-policy confidence target in every tested parameter cell. A separate deterministic guard charges every unresolved admission against the realized allowance and bounds unavailable-state admissions at every prefix under truthful outcome feedback. An executable admission procedure composes the allocation and guard. The controlled system binds source sampling, NATS delivery, randomized admission, Kubernetes objects, and API-derived outcomes in one append-only record. Independent audit finds no missing event, identity mismatch, or state mismatch in the completed functional block. A separate CPU request-accounting check preserves Job and Pod identity across an unschedulable-to-complete transition.

Precisely reproducing a target span verbatim within a long context is a fundamental capability test for LLMs: even when the target string is present in context, models often still drift within the span or fail to stop correctly, revealing limitations in source grounding and boundary control. This capability has broad practical relevance. For instance, code generation often requires embedding sensitive strings such as API keys exactly as they appear in context. In this work, we construct a benchmark to test exact span reproduction ability across contracts, scientific literature, code, and random sequences, and find that current models still make surprisingly frequent errors. To address this, we propose CopyGen, a lightweight copy-constrained plugin for frozen LLMs that turns copying from a byproduct of next-token prediction into a controlled process. Our key idea is to decompose copying into three distinct decisions: when to copy, how far to copy, and when to stop. Concretely, CopyGen first determines whether to enter copy mode, then predicts a tentative copy length, and finally verifies the generated span token by token. During verification, we adaptively adjust the logits of copy continuation tokens to incentivize or suppress continuation, and emit only the longest valid prefix. Across three LLM backbones and four benchmark tasks, CopyGen improves average exact match by 35% relative without model finetuning, and integrates seamlessly with SFT to further improve it by 4%, while substantially accelerating generation. These results suggest that reliable exact copying is not something LLMs obtain for free, but a decoding-time control problem that requires explicit modeling. Code is available at https://github.com/mbzuai-nlp/CopyGen.

Agentic AI extends large language models (LLMs) from generating responses to planning and executing complex workflows, substantially expanding their utility while exposing applications to unsafe actions, information leakage, and adversarial manipulation. Multi-agent systems further expand these capabilities through specialization and collaboration, but distributed authority, heterogeneous roles, and opaque or potentially compromised participants make safeguarding even more challenging. Existing defenses address important individual threats but provide fragmented coverage across local execution, system-level interactions, and enforcement mechanisms. We introduce MAS-Guard, a unified framework that closes this gap for safeguarding single- and multi-agent AI applications containing both instrumentable and opaque agents. Within instrumentable agents, the proposed local Agent Guardrail (A-Guard) layer applies task-specific experts to assess agentic events. At the system level, developer-owned Agent Charters and a Control Plane combine deterministic structural checks of routes, agent capabilities, and message formats with semantic assessments of roles and message content. These complementary layers enable enforcement through a shared set of interventions: allow, reflection, confirmation, and block. Evaluated across six benchmarks covering diverse single- and multi-agent scenarios with three agent LLMs from different providers, MAS-Guard reduces the overall attack success rate by 71.5-86.3% relative to unguarded agents, while meeting the utility reference criterion on benign tasks.

We present a decentralized multi-view tracking system for scalable camera networks. Single-view detections are projected to bird's-eye-view (BEV) coordinates and fused by independent multi-view trackers, each managing a dynamic spatial zone. To support scalability across spatially partitioned camera networks, we introduce a decentralized layer for trajectory unification, utilizing a locking protocol and periodic track-merging operations termed global collate. This process is supported by a delayed-output buffering mechanism that retroactively applies history corrections. The configurable delay window absorbs asynchronous coordination latency and helps stabilize cross-zone identity consistency. We detail the design of dynamic zone polygons, consensus-based coordination, and the validation framework, which together maintain global ID consistency with minimal accuracy loss compared to a single-instance baseline in the evaluated configurations.

We present SPECTER, a post-compromise covert channel that uses host-side USB power control to exfiltrate information from air-gapped and network-segregated systems. A transmitter with administrative privileges encodes bits through USB power-management or device-state operations without placing the payload in USB data transfers. The resulting signals are recovered through three alternative receiver paths based on smartphone charging states, visible peripheral behavior, and near-field RF activity. Experiments on Windows 10 and Ubuntu 22.04 demonstrate the channel on commodity hardware, with charging-state signaling at 0.2-0.25 bit/sec across Android and iOS devices and peripheral-state signaling at 1.0 bit/sec using a USB mouse on Windows. RF reception is demonstrated using a software-defined radio with a loop antenna in direct contact with the USB cable. Channel availability and observable side effects depend on the host configuration, connected device, and receiver path. These results identify host-side USB power control as a mechanism for low-rate exfiltration of short secrets and motivate treating USB power-management and device-state events as security-relevant telemetry.

We introduce StreamEval with four key features: (1) First interactive streaming video generation benchmark. (2) Fully open source. Benchmark data, evaluation code, protocols, and scoring rules are publicly available. (3) Comprehensive coverage. Five tasks cover avatar dialogue, avatar performance, world generation, instruction editing, and reference editing, with 300 scripted cases lasting 15 to 300 seconds. Ten metrics assess task execution, state consistency, visual fidelity, and audiovisual coordination. (4) Monthly updates. We will evaluate new and updated models and publish leaderboard results monthly. Results from 37 models show that high visual quality and identity retention do not necessarily coincide with reliable request execution or sustained state consistency. The leaderboard is available at https://streameval.ai.

Low Earth orbit (LEO) satellites increasingly perform time-critical onboard processing, yet heterogeneous computing capabilities, rapidly changing inter-satellite links (ISLs) and uncertain neighbor behavior make decentralized task offloading difficult under partial observations. This paper presents ORCA, an architecture of Onboard Reflective Collaborative Agents for autonomous onboard task offloading. ORCA organizes offloading decisions into a four-Agent \emph{Perceive}$\rightarrow$\emph{Reason}$\rightarrow$ \emph{Act}$\rightarrow$\emph{Learn} loop. It aims to improve efficiency and reliability in dynamic, heterogeneous, and unreliable LEO satellite networks. The simulation results of a heterogeneous 60-satellite Walker-delta constellation demonstrates that ORCA delivers competitive system efficiency in cooperative settings while achieving superior task-success robustness and offloading reliability under previously unseen malicious node behavior, indicating that outcome-grounded reflection can improve the reliability of decentralized onboard offloading without centralized runtime coordination.

Recent advances in multimodal large language models (MLLMs) have substantially improved spoken interaction and video understanding. Existing online and streaming benchmarks primarily assess causal multimodal perception through timestamped question answering, whereas real-world interaction requires models to continuously identify task-relevant multimodal cues and decide whether, when, and how to respond. This leaves open how well current models can sustain real-time interaction throughout an evolving multimodal stream. To address this gap, we introduce DIVE, a benchmark for evaluating full-duplex interaction with visual engagement under continuous and realistic multimodal streams. It organizes six tasks into three complementary interaction scenarios. (1) Proactive Intervention evaluates whether models can identify emerging visual or semantic cues and intervene at the appropriate moment. (2) Continuous Engagement evaluates sustained participation through continuous visual tracking and visually grounded query interaction. (3) Interaction Coordination evaluates how models coordinate listening and speaking under overlapping and multi-party interactions. DIVE contains 525 audio/video samples with fine-grained human annotations and precise temporal metadata. Its evaluation protocol separates response quality, participation decisions, and response timing. We evaluate representative real-time models using response quality, window response rates, and reaction latency. Experiments show that leading real-time models achieve sub-second average response latency, yet their average window response rates remain at or below 53.77%. Response quality varies substantially across models and tasks, with adapting ongoing responses to newly arriving information remaining particularly challenging. We hope DIVE facilitates progress toward more natural and capable real-time multimodal interactive systems.

We present our solution to the AKBC LM-KBC 2026 Shared Task, where we achieved a score of 0.6836 on the official leaderboard, more than double the baseline of 0.2964. This task requires generating all possible objects for subject-relation pairs using a language model with up to 32 billion parameters, with no retrieval or fine-tuning involved. We show that the prompting methodology is as important as the model architecture: while the instruction-tuned model version achieves 0.5681 with its chat interface, the base model achieves 0.6836 when prompted using text completion. Thus, instruction tuning decreases factual recall; we refer to this effect as the instruction tax, which is caused by the model's weights rather than the prompting method. Our final solution is fairly straightforward: we prompt the base model with some examples from the training data, generate 64--256 outputs, and retain only the repeated ones. Two specialized methods are then applied for each relation: one improves the recall of large answer sets, and the other verifies a person's vitality before predicting their place of death.

Large language model agents are often given an inference budget and asked to maximise success within it. A multi-turn agent has to allocate this budget across many environment interactions. An external harness can enforce the budget, but it remains unclear whether the agent itself can learn when and how much to spend. We ask whether agents can learn this allocation across dynamically specified budgets and which rewards best teach this behaviour. We adapt existing length penalties to budget constraints spanning multiple resource types, introduce new reward formulations, and compare eleven objectives in a shared training pipeline. We train Qwen3.5 models at 4B and 9B on AppWorld with budgets over tokens, API calls, and environment steps, and on the BabyAI gridworld with turn budgets. We train a separate policy for each environment and resource type, with each policy trained across budget values. Training with rewards that penalise overspending substantially improves within-budget success and natural compliance over untrained models and over training on task success alone, with agents learning to adapt their resource use to stated budgets and task difficulty. These gains extend to unseen budgets and, zero-shot, to a shifted task distribution and a new benchmark. Reward objectives trade off success within budget and natural compliance. A sigmoid penalty for overspending most often achieves the highest within-budget success at an individual budget, while a fixed negative reward for overspending generally yields the strongest natural compliance. Combining penalties for overspend frequency and magnitude gives the smallest average gap to the best method at each budget. Overall, we show that a single agent can learn to allocate resources across turns and adapt its behaviour to user-specified budgets using budget feedback without external budget-allocation controllers.

Deep learning has become the default tool for medical image analysis, driven by the open-source movement. However, achieving the best performance on a given task remains difficult, with architecture selection and hyperparameter optimization (HPO) still considered "more of an art than a science". Automated machine learning (AutoML) automates parts of model training, yet helps little in medical imaging: general-purpose approaches lack medical architectures and preprocessing, and medical-focused ones demand engineering expertise and remain confined to narrow tasks such as 3D segmentation. We introduce MedicalMLOmni, a multi-agent AutoML framework that automates model training for diverse single- and multi-modal medical tasks through a natural language interface. Specifically, MedicalMLOmni performs evidence-based planning by retrieving published results to guide model selection and configuration, and employs parallel HPO where multiple agents explore adjustment strategies concurrently. MedicalMLOmni achieves the best score on 19 of 20 benchmarks; PMC-VQA is the exception, where a vision-language model with broader pre-training leads. Backbone-matched controls attribute the gains to the framework, not the backbone. By removing coding requirements, MedicalMLOmni lowers the barrier for domain experts with rich datasets but limited engineering expertise.

Deep learning models have become essential tools in medical imaging, yet training them end-to-end for diverse tasks remains challenging due to limited labeled data and computational constraints. To address this, we introduce Brain2Vec, a foundation model pre-trained on 74,425 T1-weighted Magnetic Resonance Imaging (MRI) scans pooled from multiple studies. Brain2Vec can serve both as a weight initialization backbone for fine-tuning downstream tasks and as a standalone feature extractor. We evaluated Brain2Vec in three tasks: downstream predictive performance, convergence speed, and embedding quality. Using Brain2Vec's pretrained weights as initialization and fine-tuning on labeled data for a brain age prediction task, results demonstrate improved final accuracy and faster convergence compared to random initialization. Additionally, its frozen embeddings outperform anatomically defined volumetric features on four of five downstream classification tasks and perform competitively with a state-of-the-art foundation model pre-trained on substantially more data. Lastly, we show that Brain2Vec encodes information primarily from a spatially localized and neuroanatomically meaningful set of brain regions, and that manipulating a subset of latent features produces monotonic, region-specific effects on the reconstructed anatomy. Brain2Vec code and trained models are available as an open-source package on https://github.com/CBICA/Brain2Vec. Users can apply Brain2Vec locally by installing the package, or directly through the NiChart cloud platform https://cloud.neuroimagingchart.com, where Brain2Vec features can be derived without requiring local installation or specialized infrastructure.