Source. https://moonshotai.github.io/Kimi-Researcher/ (June 20, 2025)
A related academic paper is *Kimi k1.5: Scaling Reinforcement Learning with LLMs* (arXiv:2501.12599, Submitted 22 Jan 2025, last revised 3 Jun 2025, v4). Its title is not “Kimi-Researcher.” The quotations below are taken from the report.
Abstract
Meet Kimi-Researcher, an autonomous agent that excels at multi-turn search and reasoning. It performs an average of 23 reasoning steps and explores over 200 URLs per task. Built on an internal version of the Kimi k-series model and trained entirely through end-to-end agentic reinforcement learning (RL), it achieved a Pass@1 score of 26.9%—a state-of-the-art result—on Humanity's Last Exam, and Pass@4 accuracy of 40.17%. Starting from an initial HLE score of 8.6%, Kimi-Researcher reached 26.9% almost entirely through end-to-end RL training, providing compelling evidence that end-to-end agentic RL can significantly advance agent intelligence.
Summary
The report describes a closed-loop research agent rather than a single search followed by a summary. Given a question, the policy reads available evidence, selects search, browsing, or code execution, retains useful observations, and stops when it produces an answer. Training is characterized as end-to-end agentic RL: the policy that controls the trajectory is learned, using outcome rewards, on-policy rollouts, context management, and asynchronous infrastructure. Humanity’s Last Exam is reported as 26.9% Pass@1 and 40.17% Pass@4, rising from an 8.6% starting point. These figures describe the training outcome on that benchmark. They are not product accuracy. The report does not release full training data, complete reward details, weights, or the entire evaluation pipeline.
Detailed reading
What “end-to-end” denotes
End-to-end agentic reinforcement learning trains a single model to solve problems holistically: given a query, the agent explores a large number of possible strategies, receives rewards for correct solutions, and learns from the full trajectory.
The phrase does not mean that tools are absent. It means that the policy controlling the trajectory is learned, rather than assembled from a fixed workflow. Planning, perception, and tool use are optimized together. Training data are synthesized and filtered: items that require tools, and items that require multi-step reasoning. Ambiguous, trivial, and incorrectly keyed examples are removed. Pass@N is used as a difficulty filter.
The model is trained primarily with REINFORCE. Format penalties apply to invalid tool calls and to trajectories that exceed context or iteration limits. Correctness compares a valid final answer with a verified key. Successful trajectories receive a gamma decay so that a shorter correct path is preferred to a longer one.
On-policy generation and memory
During training, we disable LLM engine mechanisms like toolcall format enforcers to ensure each trajectory is generated entirely based on the model's own probability distribution.
Rollouts therefore come from the current model. If training used only well-formed tool calls, the policy would not encounter the errors it actually produces. Excess negative samples can reduce token probabilities and collapse entropy; the report discards some negatives to keep exploration available.
To address this, we design a context-management mechanism that allows the model to retain important information while discarding unnecessary documents, thereby extending a single rollout trajectory to over 50 iterations.
A naive agent that appends every document may exceed the context limit within about 10 iterations. An early ablation reports about 30% more iterations with context management, together with higher performance. A large context window is storage. Agent memory is a decision about which claims, evidence, and source pointers remain visible for the next step.
Infrastructure and interpretation of results
We implement a fully asynchronous rollout system with extensible Gym-like interfaces.
Task duration varies widely. Synchronous batches leave accelerators idle while waiting for the slowest trajectory. The report uses fully asynchronous rollouts, turn-level partial rollout (a timed-out task is saved and later resumed with updated weights), and a sandbox with stateful tool sessions.
Combined with adapted algorithms, this mechanism delivers substantial rollout acceleration (at least 1.5x).
xbench-DeepSearch is reported as 69% Pass@1, averaged over four runs. FRAMES, Seal-0, and SimpleQA are described as strong. The report notes that search tools vary over time, that some small sets were averaged, that answers were graded with o3-mini, and that comparison numbers came from papers or leaderboards. Behaviors such as reconciling conflicting sources and seeking additional verification are described as becoming more frequent. A cautious reading is that the environment rewarded those strategies. It does not imply consciousness.
Evidence boundary
The report does not release full training data, complete reward details, weights, or the entire evaluation pipeline. The base model is an internal k-series checkpoint. Additional search can increase latency without increasing reliability. Outcome rewards may not identify which intermediate statement first failed. Context compression may discard decisive evidence. The training direction may be summarized as: verified difficult tasks, on-policy exploration, outcome rewards, managed context, and asynchronous infrastructure. A research agent, on this account, is not a model that merely stores more facts. It is a system trained to detect missing evidence, collect it, compare alternatives, and stop when an answer can be supported.