News
· 2026-08-25
Wolfram Research has introduced the Temporal System, a Wolfram Language paclet designed to address the issue of timeline inconsistency across different musical output formats such as Csound, MusicXML, and OSC control. By maintaining an immutable beat timeline store and backend-specific contracts, Temporal System ensures synchronized generation of multiple outputs, eliminating the traditional problem of timeline drift caused by separate authoring. The system supports real-time control and rehears
News
· 2026-08-25
The ArXiv team introduces SkillForge, a novel framework designed to address the limitations of current reinforcement learning (RL) agents in accumulating reusable knowledge across tasks. SkillForge enables the verification and refinement of skills through environment interaction and incorporates evidence-based skill verification and multi-path skill induction mechanisms, allowing the skill bank to grow continuously while maintaining quality. Experiments on ALFWorld, WebShop, and AppWorld demonst
News
· 2026-08-25
The ArXiv team introduces StepGuard, a step-level guardrail model designed to enhance the safety of LLM-based agents when invoking external tools. StepGuard audits tool actions before execution and leverages the StepGen automatic data engine to generate safe and unsafe trajectories for training. Experiments demonstrate that StepGuard achieves the highest average accuracy among open-weight guard models, reducing the mean attack success rate by 77.3% while only incurring a 2.8% drop in utility.
News
· 2026-08-25
The ArXiv team introduces CAFE (Coupled Agent--Feedback Evolution), a novel framework designed to address the limitations of existing search agents in handling intermediate errors and trajectory redirection. CAFE alternates between the roles of search agent and critic, integrating online and offline optimization techniques. It achieves feedback-conditioned recovery from failure trajectories and employs comparative feedback estimation and advantage shaping to enhance model performance. Experiment
News
· 2026-08-25
The ArXiv team introduces StarHarness, a framework designed to enhance AI system performance in complex enterprise tasks by evolving environment-specific agent harnesses while keeping model weights fixed. The framework employs task stratification, separates proposer-visible search tasks from proposer-hidden selection tasks, and reserves held-out tasks for evaluating generalization. Across multiple benchmarks, StarHarness achieves a 20-35 percentage point performance improvement by repairing inte
News
· 2026-08-25
The ArXiv team introduces a novel strictly causal streaming video anomaly detector characterized by its fixed-size state updated in O(1) time and memory per incoming frame, without lookahead or clip buffering. The temporal core employs a diagonal linear state-space recurrence with an input and state-dependent decay gate, trained through causal next embedding prediction on a frozen visual backbone. The study derives a closed-form relationship between the recurrence decay spectrum and both detecti
News
· 2026-08-25
The ArXiv team introduces GameWAM, the first closed-loop World-Action Model (WAM) designed specifically for video games. GameWAM employs parallel visual and action generative processes with block-causal conditioning and flow matching to jointly generate future visual observations and executable keyboard-mouse trajectories. It also incorporates a gameplay/GUI mode prediction mechanism to handle heterogeneous native controls and demonstrates competitive task success with fewer executed actions in
News
· 2026-08-25
ArXiv introduces SPO++, a novel method for optimizing stream-aligned policy in asynchronous agentic reinforcement learning. By introducing action-token measure normalization for terminal-outcome advantages, SPO++ addresses the mismatch between trajectory centering and token-weighted quantities in existing Single-Stream Policy Optimization (SPO). Experiments on ALFWorld and Math-TIR benchmarks demonstrate that SPO++ significantly improves online learning efficiency compared to SPO, particularly i
News
· 2026-08-25
The ArXiv team introduces SimVerity, a verdict-transfer assurance framework designed to quantify the likelihood of AI agent success in physical deployment scenarios. By replaying matched scenarios on target smart home deployments and cross-validating agent execution against independently qualified physical witnesses, SimVerity highlights that deployment success is a dynamic real-world process rather than a static property in simulation. The research demonstrates that simulated passes do not guar
News
· 2026-08-25
The ArXiv team introduces Edge Skillguard, a novel security mechanism designed to address safety vulnerabilities in agent skill invocation. Current skill generation frameworks like AutoSkills and Hermes Agent primarily focus on efficiency improvements while overlooking the security risks associated with skill invocation. Edge Skillguard embeds a typed authority layer within the skill artifact, continuously monitoring world states and sensor evidence to prevent malicious or misused skill calls fr
News
· 2026-08-25
The ArXiv team introduces the Predictive Athlete Readiness for Tennis (PART) framework, a novel approach for assessing tennis players' condition and injury risk using machine learning and deep learning techniques. PART processes multimodal data from nine collegiate tennis players, including physiological metrics, training and match data, wearable device sleep data, daily self-reported information, jump assessments, and match video motion analysis. The framework captures four key characteristics
News
· 2026-08-25
The ArXiv team introduces SkyDrive, a novel framework that leverages drone-based traffic monitoring to provide efficient supervision for autonomous driving agents in new environments. Unlike traditional vehicle-based data collection methods that are resource-intensive, SkyDrive uses drones to observe multiple road users simultaneously, significantly scaling up the amount of supervision data. Experiments demonstrate that just 30 minutes of aerial monitoring per location can effectively mitigate c
News
· 2026-08-25
ArXiv introduces a lightweight monocular vision pipeline for sidewalk path extraction in micromobility systems. The approach progresses through three design iterations, from a skeleton-graph baseline to a distance-transform corridor planner, and finally to a lightweight image-space architecture. The final recommended architecture achieves real-time performance with under 50 ms per frame on CPU, making it suitable for embedded pedestrian-speed micromobility systems. This research offers an effici
News
· 2026-08-25
The ArXiv team has developed ESSE, an intelligent tutoring system for introductory programming education that leverages a large language model (LLM) to provide immediate feedback on students' explanations of code. ESSE assesses the correctness and completeness of explanations and offers targeted suggestions for improvement. Experimental results demonstrate that the system effectively guides students to refine their explanations, fostering deeper learning, and highlights the potential of LLMs in
News
· 2026-08-25
The ArXiv team presents a framework that combines Agent-Based Modeling (ABM) and Reinforcement Learning (RL) to simulate smart freight corridors. The framework integrates a physical infrastructure layer, a V2X connectivity layer, and a decision-making layer that employs RL and multi-agent RL for platoon formation and charging coordination. Preliminary results show that the cognitive scenario outperforms the baseline in terms of throughput and congestion reduction, while the assisted scenario del
News
· 2026-08-25
ArXiv researchers introduce a novel method to control tool-call rates in LLM agents using a single linear direction extracted from the model's residual stream. This approach, derived from the model's inherent tool-use preference without additional training, allows for precise inference-time intervention. By adjusting the strength of the direction, the tool-call rate can be modulated from near 0% to over 90%, maintaining well-formed calls. This method not only suppresses unnecessary tool usage bu
News
· 2026-08-25
ArXiv has released TurnBench, a multi-domain benchmark for evaluating turn-taking dynamics in spoken dialogue. The benchmark includes a 30-hour hand-labeled corpus of human conversations and a standardized evaluation protocol for end-of-turn and interruption detection. Benchmarking 14 heterogeneous systems, TurnBench highlights the impact of conversation styles on interruption detection performance and exposes the limitations of current systems in smooth floor transitions. This release provides
News
· 2026-08-25
The ArXiv team introduces ADeptS-Bench, a dual-stream trustworthiness benchmark for Computer Use Agents (CUAs). The benchmark, grounded in the ADEPTS capability framework and extensive user studies, features Safety and Disambiguation streams. The Safety stream evaluates agent robustness against threats embedded in visual interfaces, while the Disambiguation stream tests agents' ability to seek clarification when instructions are ambiguous. Experiments reveal that current models struggle to balan
News
· 2026-08-25
The ArXiv team introduces a Bidirectional Mamba-enabled Model (BM) for long-term behavioral evaluation, aimed at optimizing collaborator selection in collaborative tasks. The model constructs short-term collaboration graphs between devices and integrates short-term representations across time intervals to produce stable and reliable behavior evaluations. Experimental results demonstrate that BM outperforms baseline methods in evaluation accuracy, enabling the selection of collaborators that maxi
News
· 2026-08-25
The ArXiv team introduces TrustFormer, a novel framework for task-specific multi-dimensional trust evaluation in dynamic collaborative systems. By leveraging cross-temporal and cross-dimensional attention mechanisms, TrustFormer synchronizes heterogeneous trust data and models the evolution of multi-dimensional trust, leading to more reliable collaborator selection. Experimental results demonstrate a 40.8% improvement in trust evaluation accuracy compared to existing methods, showcasing its pote
News
· 2026-08-25
GameWAM, developed by Yunn Cheng and colleagues, is the first World-Action Model (WAM) designed for native closed-loop gameplay and GUI control in video games. It addresses the limitations of existing game agents by explicitly modeling world dynamics. The model generates future visual observations and executable keyboard-mouse trajectories through parallel visual and action generative processes, incorporating block-causal conditioning and flow matching. It also predicts gameplay/GUI modes and ge
News
· 2026-08-25
Hugging Face's research investigates the cost-quality trade-offs of model switching in LLM agents during long-running tasks. The study identifies a 'handoff tax' when switching from a low-capability (LC) to a high-capability (HC) model, where quality recovery is partial but incurs substantial costs. Conversely, downshifting from HC to LC offers a more favorable cost-quality balance. Interestingly, the preferred interface design reverses with the direction of the switch: reducing LC model traject
News
· 2026-08-25
Hugging Face's research examines the state transmission issues in large language model (LLM) agent workflows involving multi-role and multi-stage tasks. The study identifies that LLM agents may weaken action-constraining states into non-binding considerations during handoff transformations, leading to errors or violations in downstream actions. Experiments across 1,296 controlled synthetic episodes show that direct handoffs preserve constraints, while compression and plan assimilation often resu
News
· 2026-08-25
Hugging Face researchers introduce LAWA, a novel architecture for World Action Models (WAMs) that addresses the high latency issue of future imagination in existing methods. By using compact latent actions as operational representations of future intentions, LAWA enables efficient test-time future imagination without generating future observations. On the RoboCasa benchmark, LAWA achieves state-of-the-art success rates of 65.6% and 80.8% in few-shot and full data settings, respectively, outperfo
News
· 2026-08-25
Hugging Face's research team introduces CAFE (Coupled Agent-Feedback Evolution), a novel framework designed to address the challenge of feedback and decision-making optimization for search agents in complex tasks. CAFE employs a shared-parameter model that alternates between the roles of search agent and critic, leveraging both online and offline optimization to enhance the agent's adaptability to failure trajectories and feedback utilization. Experimental results demonstrate that CAFE outperfor
News
· 2026-08-25
Gen-Verse has introduced Recuris, a recursive memory architecture designed to address the challenges of recursive self-improvement (RSI) in long-horizon agent tasks. By coupling working memory with experiential memory, Recuris tracks task progress and guides skill selection in real-time, reducing historical dependencies and improving task success rates. Across multiple long-horizon benchmarks, Recuris has demonstrated significant performance improvements, such as adding 17.8 points to GPT-5.6 So
· 2026-08-25
CaptureAgent is a platform designed to automate the entire lifecycle of federal contracting for small government contractors, from opportunity discovery and bidding to project management. By integrating APIs from SAM.gov, Anthropic, and OpenAI, the platform offers intelligent features such as opportunity matching, scoring, proposal drafting, and compliance tracking, as well as project management and fundraising preparation. CaptureAgent aims to address the inefficiencies, compliance complexities
· 2026-08-25
On August 25, 2026, Nous Research released Hermes Agent, an open-source AI agent platform designed to provide scalable and customizable AI solutions. The platform focuses on enabling AI agents to execute complex tasks efficiently while supporting continuous learning. Hermes Agent's architecture emphasizes flexibility, allowing developers to tailor agent behavior and functionality to specific use cases, thereby enhancing the intelligence and efficiency of AI applications.
· 2026-08-25
William TP has launched the AI Health Agent, an intelligent system designed to provide personalized health management services for users throughout their lives. By continuously learning and analyzing user health data, the agent offers precise health recommendations, disease predictions, and tailored treatment plans. Its key strength lies in its deep understanding of user health data and real-time responsiveness, delivering smarter and more efficient healthcare management experiences.
· 2026-08-25
Anthropic has launched Claude Tag Tracing, an AI agent tracing tool designed to enhance the efficiency and behavior tracking of AI agents in complex tasks. By tagging the behavior paths of agents, this tool enables developers to better understand decision-making processes and optimize task execution workflows. This release signifies Anthropic's continued exploration in the AI agent domain, providing developers with a more powerful tool for managing and optimizing AI agents.