Jin Shang

Haoming Xu, Weihong Xu, Zongrui Li et al.

Long-horizon interactions require language models to manage accumulating information: when to update their state, when to preserve their state, and what to ignore. We study this challenge as \textbf{Contextual Belief Management (CBM)}: maintaining a predicted belief state aligned with formal evidence while isolating task-irrelevant noise. To make CBM measurable, we introduce BeliefTrack, a closed-world benchmark spanning Rule Discovery and Circuit Diagnosis, where a finite belief space and symbolic verifiers enable exact turn-level evaluation. BeliefTrack diagnoses three failures: Failed Stay, Failed Update, and Failed Isolation. Across multiple LLMs, vanilla models exhibit severe CBM failures, while explicit belief-tracking prompts provide limited gains. In contrast, reinforcement learning with belief-state rewards reduces failure rates by 70.9\% on average. Further probing reveals latent belief-state dynamics behind these failures, and representation-level steering reduces failure rates by 46.1\% across two tasks\footnote{Code is coming soon at https://github.com/zjunlp/CBM.

Yuan Liang, Ruobin Zhong, Haoming Xu et al.

Current AI agents can flexibly invoke tools and execute complex tasks, yet their long-term advancement is hindered by the lack of systematic accumulation and transfer of skills. Without a unified mechanism for skill consolidation, agents frequently ``reinvent the wheel'', rediscovering solutions in isolated contexts without leveraging prior strategies. To overcome this limitation, we introduce SkillNet, an open infrastructure designed to create, evaluate, and organize AI skills at scale. SkillNet structures skills within a unified ontology that supports creating skills from heterogeneous sources, establishing rich relational connections, and performing multi-dimensional evaluation across Safety, Completeness, Executability, Maintainability, and Cost-awareness. Our infrastructure integrates a repository of over 200,000 skills, an interactive platform, and a versatile Python toolkit. Experimental evaluations on ALFWorld, WebShop, and ScienceWorld demonstrate that SkillNet significantly enhances agent performance, improving average rewards by 40% and reducing execution steps by 30% across multiple backbone models. By formalizing skills as evolving, composable assets, SkillNet provides a robust foundation for agents to move from transient experience to durable mastery.

Hai Xiao, Jin Shang, Mengyuan Huang

Deep artificial neural networks, trained with labeled data sets are widely used in numerous vision and robotics applications today. In terms of AI, these are called reflex models, referring to the fact that they do not self-evolve or actively adapt to environmental changes. As demand for intelligent robot control expands to many high level tasks, reinforcement learning and state based models play an increasingly important role. Herein, in computer vision and robotics domain, we study a novel approach to add reinforcement controls onto the image recognition reflex models to attain better overall performance, specifically to a wider environment range beyond what is expected of the task reflex models. Follow a common infrastructure with environment sensing and AI based modeling of self-adaptive agents, we implement multiple types of AI control agents. To the end, we provide comparative results of these agents with baseline, and an insightful analysis of their benefit to improve overall image recognition performance in real world.

Jin Shang, Muhammad Junaid Farooq, Quanyan Zhu

The Internet of things (IoT) comprises of wireless sensors and actuators connected via access points to the Internet. Often, the sensing devices are remotely deployed with limited battery power and are equipped with energy harvesting equipment. These devices transmit real-time data to the base station (BS), which is used in applications such as anomaly detection. Under sufficient power availability, wireless transmissions from sensors can be scheduled at regular time intervals to maintain real-time data acquisition. However, once the battery is significantly depleted, the devices enter into power saving mode and need to be more selective in transmitting information to the BS. Transmitting a particular piece of sensed data consumes power while discarding it may result in loss of utility at the BS. The goal is to design an optimal dynamic policy which enables the device to decide whether to transmit or to discard a piece of sensing data particularly under the power saving mode. This will enable the sensor to prolong its operation while causing minimum loss of utility to the application. We develop an analytical framework to capture the utility of the IoT sensor transmissions and leverage dynamic programming based approach to derive an optimal real-time transmission policy that is based on the statistics of information arrival, the likelihood of harvested energy, and designed lifetime of the sensors. Numerical results show that if the statistics of future data valuation are accurately predicted, there is a significant increase in utility obtained at the BS as well as the battery lifetime.