A research paper compares the performance of two hierarchical red team agent architectures, RL+RL and LLM+LLM, in different environments, revealing environment-dependent inversion and highlighting the importance of considering specific failure modes when designing hybrid planner-executor architectures.

Researchers propose using Test-Time Training (TTT) layers in a Decision Transformer to improve long-term memory in offline Reinforcement Learning (RL). They analyze the performance of the Decision Titan, a variant of the Decision Transformer with TTT layers, in the X-Maze environment.

Researchers propose Reward-Free Policy Optimization (RFPO) for large language models (LLMs), utilizing a pretrained critic to predict future outcomes and provide learning signals without requiring external rewards or completed rollouts.

A new method, PACE, is proposed for controlling staleness in asynchronous reinforcement learning (RL) for large language model post-training. PACE improves validation accuracy and reduces GPU time, matching synchronous RL performance.

Agentick is a unified benchmark for evaluating sequential decision-making agents, including RL, LLM, and hybrid approaches. It provides 37 procedurally generated tasks across various difficulty levels and observation modalities, and includes a coding API, oracle reference policies, and a live leaderboard. The evaluation of 27 configurations and over 90,000 episodes highlights the need for improvement across all agent paradigms.