Researchers from Meta AI and the University of Illinois Urbana-Champaign have developed EvoHarness-RL, a trainable coordination layer enabling large language models to create, access, and manage their own external state. This framework achieved a 96.9% success rate on the ALFWorld benchmark, compared to 47.9% for the baseline model without it. The system relies on three components: Belief, Progress, and Experience, forming the BPE architecture. Training combines supervised fine-tuning with expert demonstrations and Group Relative Policy Optimization (GRPO), a cost-aware technique balancing performance and computational efficiency. The model also demonstrated strong generalization capabilities with an 86.6% success rate in unseen environments not encountered during training.
Source: Read the original article
