Researchers propose ReTerm, a self-improving terminal agent that can recover from errors during long-running execution. ReTerm is trained to learn from its own failed tasks and uses a customized data pipeline and online RL algorithm to improve its recovery capabilities. The model achieves a 14.6 point improvement over the base model on Terminal-Bench 2.0.

ClawLoop is a framework that enables self-improving AI agents without an external teacher model. It consists of three phases: Task Synthesis, Training, and Diagnosis and Feedback, which form a self-sustaining cycle.

Third Moment Research runs 10,000 agents with AI models GPT-6 Astra, GPT-6.1 Sol, Codex, and Daybreak for 36+ hours, demonstrating their ability to build and improve AI agents over extended periods.

This research proposes Matched-Context Group Policy Optimization (MCGPO), a method for training long-horizon language model agents using sparse terminal feedback. MCGPO constructs weighted action preferences from collected trajectories, improving performance over GraphGPO in various environment-backbone settings.

A new AI framework called MENTOR is proposed to address the issue of 'zero-reward collapse' in Reinforcement Learning from Verifiable Rewards (RLVR), a key driver for Large Reasoning Models. MENTOR dynamically generates targeted, instructional prompts to guide the learning agent's next attempt, improving efficiency and reducing training failures.

Researchers propose Branching On-Policy Distillation, an improvement to on-policy distillation for AI agents. It selects informative decision points and applies teacher supervision to both main trajectories and branches, reducing student-generated token costs by 19-31%. The approach outperforms existing methods across various benchmarks.

LiteGUI is a new framework for building lightweight GUI agents through a two-stage post-training process. It addresses the challenges of GUI interaction with a guided distillation and dual-level reinforcement learning approach. The framework includes a data generation pipeline and a corresponding dataset, Lite-Dataset, to support training and evaluation of GUI agents.

StoreBench is a live-commerce environment for evaluating and training autonomous operator agents. It simulates a production-grade online store, allowing agents to act through 29 merchant tools under a windowed operation budget. The environment is designed to test long-horizon planning and economic judgment under uncertainty.