Behavior Best Of N Bbon Sampling With Rollout Based Refinement

via “behavior best-of-n (bbon) sampling with rollout-based refinement”

Agent S: an open agentic framework that uses computers like a human

Unique: Implements in-context reinforcement learning through parallel rollout sampling and LMM-based trajectory evaluation, achieving 72.60% OSWorld accuracy without model fine-tuning by leveraging the LMM's reasoning capability to select high-quality action sequences

vs others: Outperforms single-shot planning by 10-15% on complex benchmarks through best-of-N selection, while avoiding the infrastructure complexity of external RL training or reward models