Second Thought Parallel Reasoning Framework for LLM Agents

Second Thought is a training-free inference framework that forks four auxiliary reasoning branches during an agent's action-observation idle window. By decoding these branches concurrently, the method reduces average turn counts across three agentic benchmarks and three reasoning LLMs.