Second Thought Enables Parallel Reasoning for LLM Agents

Second Thought is a training-free inference framework that utilizes the reasoning idle window during an agent's action/observation phase. It forks four auxiliary reasoning branches to decode concurrently, merging thoughts back once environment observations arrive to increase agent intelligence without sequential delays.