Three research papers published on arXiv on August 10, 2026 advance different aspects of long-horizon AI agent systems.
According to the first paper, researchers introduced a benchmark and annotation framework for “trajectory attribution” in large language model (LLM) agents. The benchmark provides annotations covering “task-aligned actions, unsafe actions, and safety refusals” across more than 1,300 trajectories from AgentDojo and Agent3Sigma settings. The framework defines two evaluation tasks: “primary attribution localization and attribution-chain recovery,” with project resources available on GitHub.
A second paper presents Recursive Synthetic Terminal Tasks (RST), a framework for generating long-horizon training data for terminal agents. According to the authors, RST produced 37,484 synthesized tasks at approximately $0.05 per task across fifteen recursive rounds. The paper reports that task difficulty increased substantially, with median reference solutions growing from 67 to 374 lines of code, and DeepSeek-V4-Pro pass@4 performance dropping from 90% at round 1 to 2.5% in later rounds.
The third paper introduces CEDAR, an “autonomous method that uses Large Language Model (LLM) agents to discover complex systems satisfying user-specified behavioral goals,” according to the abstract. CEDAR combines LLM-driven Monte Carlo Tree Search with complex systems modeling and was accepted as a talk at the 2026 Conference on Artificial Life.