Big language models are great at words but waste lots of time and energy when they try random actions in non-language games like Sudoku, Sokoban, 2048, FrozenLake, and Rubik’s Cube.

#SCOUT#Reinforcement Learning#Supervised Fine-Tuning

Turn-PPO: Turn-Level Advantage Estimation with PPO for Improved Multi-Turn RL in Agentic LLMs

Intermediate

Junbo Li, Peng Zhou et al.Dec 18arXiv

Turn-PPO is a new way to train chatty AI agents that act over many steps, by judging each conversation turn as one whole action instead of judging every single token.

#Turn-PPO#multi-turn reinforcement learning#agentic LLMs

Meta-RL Induces Exploration in Language Agents

Intermediate

Yulun Jiang, Liangze Jiang et al.Dec 18arXiv

This paper introduces LAMER, a Meta-RL training framework that teaches language agents to explore first and then use what they learned to solve tasks faster.

#Meta-Reinforcement Learning#Language Agents#Exploration vs Exploitation