Science & Mathematics

Spatial intelligence requires agents to actively acquire visual evidence and integrate observations into a coherent representation of the three-dimensional world. Yet spatial question answering offers limited evidence of whether models can construct accurate metric maps through autonomous exploration. We introduce SpatialScout, a benchmark for evaluating vision-language model (VLM) agents on active object-level 3D mapping. SpatialScout comprises 80 diverse indoor scenes containing 3,140 target objects. Given an object inventory, RGB observations, camera calibration, and camera poses, agents control a camera to explore each scene and estimate object centers in a shared coordinate frame. The task requires coordinating viewpoint selection, cross-view association, geometric estimation, and map revision under bounded interaction budgets. Evaluation jointly measures localization accuracy and target coverage through object-scale overlap scores that penalize missing predictions. Experiments with frontier VLMs reveal substantial limitations: the strongest evaluated model, Claude Opus 5.5, achieves a mean overlap score of 41.71%, with 47.23% of targets exceeding an intersection-over-union (IoU) threshold of 0.5 and only 3.25% exceeding 0.9. These results reveal a substantial gap between recovering coarse object layouts and constructing complete, precise metric maps. SpatialScout provides a quantitative testbed for developing agents that build reliable 3D spatial representations through active perception. Code is available at https://github.com/JesseZZZZZ/SpatialScout.

Smartphone GUI agents complete tasks in mobile apps by repeatedly reading the screen and asking a model for the next action. Mobile neural processing units (NPUs) enable local execution, keeping screen contents on the phone, but model latency accumulates with every action. Hybrid linear attention models reduce memory requirements, yet reusing their recurrent state requires checkpoints at the appropriate input positions. Agent requests retain instructions but rewrite recent history, so checkpoint policies designed for chat continuation can miss reusable prefixes. Retaining rollback states for speculative decoding can also move prompt processing off the NPU's matrix unit, adding overhead that outweighs decoding savings on short action outputs. A loaded model can further hold the NPU awake between tasks. We present Tandem, a runtime for hybrid GUI agents on mobile NPUs. Tandem saves recurrent state where consecutive prompts diverge, allowing later requests to reuse previously processed context. It keeps most prefill on the matrix unit and produces rollback states through a short per-token tail. Its verification kernel reads each weight block once for all proposed tokens. Between request bursts, Tandem releases NPU power votes without unloading the model. We implement Tandem in llama.cpp and evaluate it on recorded agent interactions across two Snapdragon generations. On a Snapdragon 8 Elite, Tandem reduces per-step latency by 30% for text-based screen descriptions and 16% for screenshots relative to llama.cpp's fastest NPU configuration, reduces energy per text step by 33%, and lowers standby power with the model loaded from 0.76 to 0.22 W.

Composed Image Retrieval (CIR) retrieves products matching a reference image modified by natural language---a core capability for eCommerce search. Existing training-free methods fuse CLIP embeddings via arithmetic, but we identify a fundamental limitation: user text modifications are attribute-sparse---a user who types ``make it red with shorter sleeves'' specifies what to change but not what to keep (the floral pattern, the V-neck, the fit-and-flare silhouette). We introduce language-space fusion: a Vision-Language Model (VLM) jointly interprets the reference image and modification to generate a visually grounded product-title caption---e.g., ``Red floral fit-and-flare dress with short sleeves and V-neck''---that fills in the missing attributes from the image. Our method, CaptionFuse, requires no task-specific training and is model-agnostic. A controlled experiment confirms that attribute content--- not linguistic style---drives the gains: imperative and descriptive captions with identical attributes achieve equal R@10, while both outperform raw modifications by +6.5 points.

LLM trading agents are evaluated almost exclusively in backtests and autonomous benchmark harnesses. Deployment, where actions are irreversible, raises a different question: who verifies the agent's actions, and which failures can each verifier structurally catch? We report on AllPath, a personal LLM trading agent deployed against a live paper-brokerage account. The system watches positions through a deterministic rules engine, and each night it reviews its own day and proposes revisions to its own trading strategies. Verification is split across three layers. Code verifies what must not depend on judgment: deterministic gates enforce risk limits and block any strategy revision that fails an integrity check. The agent verifies itself within fixed bounds: its nightly review has an iteration cap and no tools that can act. A human verifies judgment, through approval screens that show the current state of each proposal. We also compared two verification policies for the agent's self-revisions on the live system: two and a half weeks with every self-revision human-gated, then five weeks with reflection revisions auto-applied behind the same code gates, while the operator's chat channel stayed human-gated throughout. Each layer caught a different kind of failure. The code gates rejected no proposal; their role was keeping revisions consistent. The human rejected two proposals that passed every code check, both over a disagreement about the market, and later removed 11 of the 19 rules the agent had added on its own, at a median lag of eight days. The most instructive production incidents, such as strategies silently left unmonitored, passed every gate; the operator found them by hand, and the fix each time was an interface change that makes the condition visible. We contribute the layered architecture, the two-regime account, and a production incident taxonomy, none observable in a backtest.

Robotic value models assess task execution from visual observations and instructions, providing feedback for policy learning and execution monitoring. However, existing evaluations cover only selected aspects of execution understanding, leaving the reliability of value feedback insufficiently tested. We introduce RoboValue, a fine-grained sim-and-real benchmark for unified evaluation of robotic value models. Its framework evaluates four complementary capability dimensions through shared interfaces: task-state understanding, temporal progress monitoring, failure and recovery reasoning, and value consistency. Diagnostic trajectories and instruction variations test whether value judgments accurately reflect task requirements and changes throughout execution. RoboValue-Dataset covers 15 simulation and 20 real-world manipulation tasks, providing expert demonstrations for training and annotated trajectories for fine-grained evaluation. We evaluate representative robotic value models in zero-shot and one-shot settings under standard conditions and separate shifts in robot embodiment and environment. Our results reveal errors in fine-grained instruction grounding, history-dependent progress tracking, and recovery assessment, even in models with strong outcome and forward-progress scores. The same semantic subtask can also receive substantially different value gains across valid execution orders. These findings identify concrete targets for value supervision, while the dataset and public leaderboard support continued assessment of feedback reliability. The website is available at https://robovalue-benchmark.github.io/.

AI agents deployed inside organizations can pose an insider threat if they make mistakes, are compromised by an attacker, or pursue misaligned objectives. We propose honeypots as an AI control method to improve the security of internal agent deployments. Rather than only using honeypots for pre-deployment evaluations, defenders should plant decoy opportunities for harmful actions throughout the production environment to detect misbehavior, incriminate malicious agents, and deter attacks. Such honeypots can elicit misbehavior under real deployment conditions, identify malicious agents even if chain-of-thought monitorability degrades, provide reliable warnings about agent misbehavior to security teams, and scale cheaply to many agent instances. The presence of honeypots also introduces uncertainty about which opportunities are real, thus deterring malicious agents from attacking and reducing the rate of misbehavior. We discuss the limitations of honeypots and give reasons to expect them to remain effective as agent capabilities increase. Finally, we outline key research questions to measure and improve the effectiveness of honeypots: testing the robustness of honeypots, comparing their detection effectiveness with other monitoring methods, understanding their incrimination and deterrence effects, and building supporting infrastructure.

Recursive reasoning models refine persistent internal states by repeatedly applying learned operators. We investigate whether adding a persistent intermediate state that integrates repeated updates of a faster state through the same operator improves predictive performance relative to each model family's baseline at matched parameter counts. We define this organisation operationally and implement it within three architectures: the Hierarchical Reasoning Model (HRM), Tiny Recursive Model (TRM) and Universal Reasoning Model (URM). Using the official codebases on ARC-AGI-1, we compare each baseline with its parameter-matched nested variant. We evaluate held-out queries from tasks whose demonstrations were available during training. The experiments use a single-GPU budget far below the published training protocols. Under the tested conditions, nested variants show no consistent advantage in token accuracy, predictive loss or exact solving, which remains near zero for all configurations. The results therefore leave the hypothesis open. We specify a confirmatory comparison, with an informative baseline, multiple seeds, controls that isolate nesting and predefined criteria, to determine whether this organisation merits further development.

Strong thermal gradients in tungsten plasma-facing components can bias the migration of irradiation-induced defects and thereby alter their redistribution, interaction, and retention. Here, we extend an atomistic jump-statistics framework previously developed for hydrogen to quantify the thermomigration of a single vacancy and interstitial He$_1$--He$_4$ clusters in body-centered-cubic tungsten. Non-equilibrium molecular-dynamics simulations performed over several thermal gradients are used to determine the directional jump ratio, while geometrically derived projected jump lengths convert the fitted directional-bias coefficient into the heat-of-transport coefficient $Q^*=αkT^2$, where $k$ is the Boltzmann constant, $T$ the temperature, and $α$ a constant. All investigated defects exhibit negative values of $α$, corresponding to net migration toward the hot region. He$_4$ has the largest response, $α=-2.2457\times10^{-2}$ K$^{-1}$, followed closely by the vacancy, whereas He$_3$ exhibits the weakest response despite its intermediate cluster size. Independent constant-temperature simulations over 1000--1600 K show that He$_3$ is also the least mobile cluster and has the largest effective migration energy, $E_m^{\mathrm{eff}}=0.271$ eV. In contrast, He$_4$ is more mobile than He$_3$ and has the smallest effective migration energy, 0.087 eV. Agreement between the He$_3$ concentration profile predicted from the jump statistics and that obtained from a long molecular-dynamics trajectory provides an independent validation of the method. These results show that thermomigration cannot be inferred from cluster size, diffusivity, or migration energy alone and provide defect-specific heat-of-transport parameters for higher-length-scale models of tungsten under fusion-relevant thermal gradients.

Multi-agent approaches have been investigated recently as a means of improving the scalability of reinforcement learning for neural combinatorial optimization. Existing methods assume that all problem instance information is known upfront. However, real-world problems are often dynamic: relevant information is progressively revealed as the agents take actions (e.g., arriving at an edge in a routing problem only to discover it is untraversable). In such cases, existing methods fail to coordinate agents to explore and exploit the progressively revealed information. To address the arising coordination challenge, we introduce PACT (Plan-Augmented Coordination via Trace-graphs). PACT augments the environment graph with auxiliary nodes representing the plans of agents, which are connected to the resource nodes to indicate future dependencies. The proposed message-passing scheme over the augmented graph grounds inter-agent coordination in the environment and enables proactive, rather than reactive, conflict resolution. We show that PACT improves solution quality in dynamic problems relative to baselines and exhibits the best transfer to larger instances on which, remarkably, it outperforms centralized oracles with privileged access to instance uncertainty.

LLM agentic workflows increasingly employ different models for specialized roles. Users may prefer different trade-offs between task quality and deployment cost, sometimes also considering latency. Finding good assignments incurs a separate search cost from evaluating multiple candidates. We introduce Cost-Coupled Gittins (CC-Gittins), a multi-objective extension of single-objective cost-aware Gittins search for Pareto recommendation. It exploits the coupling between evaluation and deployment costs: assignments that cost more to deploy also cost more to evaluate. The search alternates between seeking higher quality and lower deployment cost, reusing each evaluation to inform both searches. After each batch, CC-Gittins recommends assignments that are nondominated under uncertainty-adjusted estimates of quality and deployment cost. We evaluate CC-Gittins on eight quality--cost matrices from workflows with two to five roles and up to 1,681 candidate assignments. We also evaluate the method with latency as a third objective. Across the eight benchmarks, CC-Gittins recovers a substantial fraction of the true Pareto frontier at roughly 10% of the exhaustive-evaluation cost, and recovers most of the frontier with few false positives by 30%.

We prove that, for every \(\varepsilon>0\), a \((\frac{k}{2k-1}+\varepsilon)\)-approximation algorithm for maximizing monotone \(k\)-submodular functions requires exponentially many value-oracle queries. This improves the \(\frac{k+1}{2k}\) oracle bound of Iwata, Tanigawa, and Yoshida. Their \(\frac{k}{2k-1}\)-approximation algorithm therefore gives the exact threshold in the monotone setting. For \(k\ge5\), our bound also matches, up to an arbitrarily small loss, Aggarwal's recent approximation guarantee for non-monotone \(k\)-submodular maximization.

Evaluating an agent's planning ability remains challenging for open-domain tasks, where verifying procedures through execution is often impractical. Existing evaluations typically assess only the final plan and do not distinguish procedural errors such as missing, incorrect, or misordered actions. We introduce STEPBENCH, a benchmark of 6,885 verified procedures crawled from wikiHow across 81 open-domain categories. From each procedure, we derive 5 diagnostic tasks: four with deterministic answers involving withheld, incorrect, or shuffled steps, and one requiring next-step generation. We additionally evaluate full-plan generation to examine how diagnostic performance correlates with end-to-end plan validity. Using the same procedures across tasks lets us directly compare different planning capabilities. For generation, STEPEVAL identifies critical failures that prevent goal completion. Across 14 configurations of 12 models, we find substantial differences across planning capabilities, with step ordering proving the most difficult: the strongest model correctly orders only 63.0% of procedures, dropping to 33.1% for those with nine or more steps. Incorrect-step detection also predicts judged plan validity across models, suggesting a lower-cost alternative for evaluating generation quality. On STEPBENCH-Val, comprising 214 procedures published in 2026, comprehension scores and rankings remain stable, suggesting limited influence from memorisation.

Human skin contains diverse mechanoreceptors, including Merkel cells and Pacinian corpuscles, each localized in specific regions. Although their spatial organization is known to be shaped by molecular signaling and neural induction during development, experimentally examining its functional significance under controlled conditions in living systems remains challenging. In particular, physically comparing alternative receptor arrangements that do not occur biologically is difficult. In this study, we developed a layered biomimetic robotic skin designed to mimic key structural features of human skin tissue by embedding receptor-analog sensors within compliant soft media. Using this platform, we implemented both biologically motivated and counterfactual sensory arrangements and quantitatively evaluated their performance across three passive tactile tasks: detection of weak contacts, a two-point discrimination-like assessment, and texture identification. Arrangements reflecting key features of the biologically motivated sensory organization exhibited higher performance than the alternative configurations examined. This study addresses a biological question concerning how spatial organization influences tactile function by presenting a biomimetic robotic approach that enables physical and empirical comparison of biologically motivated and counterfactual sensory arrangements that are difficult to investigate in living systems.

Video face swapping transfers the reference identity to a target video while preserving its other attributes and motion. Despite recent progress in reference-guided flow-based methods, jointly achieving faithful identity transfer, high fidelity, and real-time generation over long video streams remains challenging. To address this challenge, we rethink the task-specific structure of reference-guided video face swapping. First, face swapping changes only identity while preserving most visual content, yet flow-matching supervision is weakly sensitive to these changes. We therefore develop Joint Rec-ID Loss, which balances explicit identity supervision and latent reconstruction according to the noise level. Second, the target video already provides non-edited attributes and temporal dynamics, enabling generation with fewer spatial tokens and shorter temporal contexts through Face-RoI Generation and Short-Context Rollout Training. By integrating these designs, we propose StreamingSwap, a reference-guided flow-matching framework for real-time streaming video face swapping. We also introduce HumanFaceVid, comprising 1,000 hours of cinematic training data and evaluation sets of 200 short and 200 long videos. Experiments demonstrate that our method achieves competitive identity transfer and visual quality at 20 FPS on a single NVIDIA RTX 4090 GPU, with stable generation over videos up to 10 minutes long. Project webpage: https://aim-uofa.github.io/StreamingSwap/

Invisible watermarks are used to mark who owns a painting. We show that a simple attack removes the most robust one we test. The attack redraws the watermarked painting with brush strokes, using a pretrained stroke-based renderer. We call this attack BrushOff. It needs no training, no knowledge of the watermark, and no generative model. We test four watermarks on 500 paintings from five styles. VINE, a recent watermark built to survive image editing, survives all 14 conventional distortions we test, with at least 97.8% detection. Repainting lowers its detection to 1.8%, while the output keeps a no-reference quality score close to that of the clean paintings. Diffusion regeneration can also remove VINE, but at the same removal level it moves the image twice as far from the painting in DreamSim distance, and it changes what the painting shows. Because the renderer draws a countable number of strokes, we can control the attack strength step by step. Finally, we test what kind of signal survives. A simple probe that stores bits in the relative brightness of neighboring regions survives repainting, while a probe that uses absolute brightness does not. The same probe fails under diffusion regeneration, while VINE fails under repainting. Robustness to these two attacks is therefore a separate property, and benchmarks should test both.