Enabling Self-Improving Agents to Learn at Test Time With Human-In-The-Loop Guidance

Large language model (LLM) agents often struggle in environments where rules and required domain knowledge frequently change, such as regulatory compliance and user risk screening. Current approaches, like offline fine-tuning and standard prompting, are insufficient because they cannot effectively adapt to new knowledge during actual operation. To address this limitation, we propose the Adaptive Reflective Interactive Agent (ARIA), an LLM agent framework designed specifically to continuously learn updated domain knowledge at test time. ARIA assesses its own uncertainty through structured self-dialogue, proactively identifying knowledge gaps and requesting targeted explanations or corrections from human experts. It then systematically updates an internal, timestamped knowledge repository with provided human guidance, detecting and resolving conflicting or outdated knowledge through comparisons and clarification queries. We evaluate ARIA on the realistic customer due diligence name screening task on TikTok Pay, alongside publicly available dynamic knowledge tasks. Results demonstrate significant improvements in adaptability and accuracy compared to baselines using standard offline fine-tuning and existing self-improving agents. ARIA is deployed within TikTok Pay serving over 150 million monthly active users, confirming its practicality and effectiveness for operational use in rapidly evolving environments.

Inverse Delayed Reinforcement Learning

Inverse Reinforcement Learning (IRL) has demonstrated effectiveness in a variety of imitation tasks. In this paper, we introduce an IRL framework designed to extract rewarding features from expert trajectories affected by delayed disturbances. Instead of relying on direct observations, our approach employs an efficient off-policy adversarial training framework to derive expert features and recover optimal policies from augmented delayed observations. Empirical evaluations in the MuJoCo environment under diverse delay settings validate the effectiveness of our method. Furthermore, we provide a theoretical analysis showing that recovering expert policies from augmented delayed observations outperforms using direct delayed observations.

Case Study: Runtime Safety Verification of Neural Network Controlled System

Neural networks are increasingly used in safety-critical applications such as robotics and autonomous vehicles. However, the deployment of neural-network-controlled systems (NNCSs) raises significant safety concerns. Many recent advances overlook critical aspects of verifying control and ensuring safety in real-time scenarios. This paper presents a case study on using POLAR-Express, a state-of-the-art NNCS reachability analysis tool, for runtime safety verification in a Turtlebot navigation system using LiDAR. The Turtlebot, equipped with a neural network controller for steering, operates in a complex environment with obstacles. We developed a safe online controller switching strategy that switches between the original NNCS controller and an obstacle avoidance controller based on the verification results. Our experiments, conducted in a ROS2 Flatland simulation environment, explore the capabilities and limitations of using POLAR-Express for runtime verification and demonstrate the effectiveness of our switching strategy.