Interleaved Thinking
Interleaved thinking is the agentic-model ability to reason, act, observe, and then reason again after receiving tool or environment feedback. In 当我们在讨论 Harness 的时候，我们在讨论什么 | 深度对谈: MiniMax × Hermes Agent, the MiniMax guests distinguish this from chatbot behavior: a chatbot can answer the present prompt, while an agentic model must explore, correct paths, and update plans as the environment changes.
Vol.114 AI的2025和DeepSeek们的未来 | 对谈复旦张奇教授 adds 张奇’s o1/o3 interpretation. He treats reflection and multi-path revision as the important change: a model no longer has to produce the final answer in one shot, and the same pattern can improve Retrieval-Augmented Generation or search when an agent notices a bad query and rewrites it.
Key Claims
- Agentic models need to revise plans after tool calls rather than simply execute the first plan.
- Benchmarks that require cross-source search and multi-condition answers test this behavior better than one-shot chat tasks.
- Agent Harness matters because the model can only interleave thought and action when it has tools, observable state, and feedback.
- The concept is a model-side complement to Agentic Workflow and Model Harness Co-Evolution.
- Reflection loops are one reason 2025 agents can be more than RPA-style workflows, because failed steps can become new evidence for self-correction.
Connections
- MiniMax — company context for the term in the source.
- Agent Harness — runtime environment that exposes observations and feedback.
- Agentic Workflow — task pattern enabled by interleaved reasoning and action.
- Multi-Agent Collaboration — peer checking can help when a single agent’s interleaved path drifts.
- AI Coding Verification — software tasks require agents to observe tests, errors, diffs, and reviews before deciding next steps.
- 张奇, Model Post-Training Bottleneck, Retrieval-Augmented Generation, and Long-Horizon AI — vol.114’s reflection, query-rewrite, and agent outlook.