How End-to-End Reinforcement Learning Powers the Kimi Researcher AI Agent

Introducing Kimi Researcher

Kimi Researcher is an AI Agent designed to perform genuine research rather than merely act as a search tool. It achieved a SOTA score of 26.9% on the Humanity’s Last Exam benchmark, generating reports of tens of thousands of words.

Why End‑to‑End Reinforcement Learning?

Traditional agent approaches rely on handcrafted workflows or supervised fine‑tuning (SFT), which are limited by human design and labeling capacity. By creating a virtual environment where the agent can explore, trial‑and‑error, and learn from successful outcomes, reinforcement learning (RL) enables the model to evolve its research abilities autonomously.

Flexibility: The RL‑trained agent is not bound by fixed rules; it generates actions dynamically based on the task.

Scalability: Performance improves by increasing training data and compute rather than human‑written prompts.

Continuous Improvement: When the agent struggles with a problem, the issue is added to the training set, allowing the model to learn a solution itself.

Results and Emergent Behaviors

Using RL, Kimi Researcher’s pass@4 metric reached 40.17%, meaning it solves over 40% of difficult problems within four attempts. The model also exhibits emergent behaviors such as multi‑step searching for verification and even proposing to email paper authors for clarification.

Practical Scenarios

The agent can quickly produce in‑depth reports on unfamiliar domains, assist with literature reviews, act as a research copilot, and handle specific tasks like benchmark discovery, historical knowledge structuring, legal data aggregation, and product recommendation analysis.

Conclusion

Kimi Researcher demonstrates that end‑to‑end reinforcement learning can transform AI agents from static tools into collaborative partners capable of autonomous research and continuous self‑improvement.