Learning Perception to Action Rules through Reinforcement Learning for Theory-of-Mind Reasoning
This work generates concise rules that express relations between perception to action and uses Group Relative Policy Optimization (GRPO) to train Qwen2.5 and Llama3 models across four ToM benchmarks, suggesting that explicit rules help models bridge the reasoning gap in ToM and support knowledge transfer across models.
Soyoung Oh, Nursulu Sagimbayeva, N. Jennings et al.
· 0 citations