Kimi-Researcher:端到端 RL 训练实现新兴智能体能力

End-to-End RL Training for Emerging Agentic Capabilities

June 20, 2025 • 10 min read

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.

Kimi-Researcher has also achieved strong performance across several complex and challenging real-world benchmarks. On xbench, a new, dynamic, professionally-aligned suite designed to bridge AI capabilities with real-world productivity, Kimi-Researcher achieved 69% pass@1 (averaged on 4 runs) on xbench-DeepSearch, outperforming models such as o3 with search tools. On benchmark tests for multi-turn search reasoning (FRAMES, Seal-0) and factual information (SimpleQA), Kimi-Researcher also achieved strong performance.

Comparison of Kimi-Researcher and other models

Figure 1: Potential fluctuations in tools, such as search engines, may affect performance. The results are tested on: HLE on June 17, 2025; and xbench-DeepSearch, Seal-0, Frames, and SimpleQA on June 18, 2025.

All Kimi-Researcher results were evaluated using o3-mini. Scores of other models are referenced from the relevant papers or leaderboards. For benchmarks with fewer than 200 test samples (xbench, Seal-0), we performed four runs and reported the average result (avg@4). We do not compare multi-agent workflows based on multiple frontier models here, as our focus is on evaluating model capabilities.

End-to-end agentic RL is promising but challenging

Kimi-Researcher is an autonomous agentic and thinking model designed to solve complex problems through multi-step planning, reasoning, and tool use. It leverages three main tools: a parallel, real-time internal search tool; a text-based browser tool for interactive web tasks; and a coding tool for automated code execution.

Formally, given the state observation \(s_t\) (for instance, \(s_0\) includes system prompt, tool declarations, and user query), Kimi-Researcher generates \(think_t\) and \(action_t\). An action can either be a tool call or an indication to terminate the trajectory.

Traditional agent development has key limitations:

  • Workflow-Based Systems: Multi-agent workflows assign roles to specialized agents and coordinate using prompt-based workflows. While effective, they are tied to specific LLM versions and need frequent manual updates.
  • Imitation Learning with Supervised Finetuning (SFT): Aligns models well with human demonstrations but struggles with data labeling — especially for long-horizon, agentic tasks in dynamic environments.

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. However, it introduces new challenges:

  • Dynamic Environments: Agents must adapt to constantly changing conditions.
  • Long-Horizon Tasks: Kimi-Researcher can run 70+ search queries per trajectory, with context windows reaching hundreds of thousands of tokens.
  • Data Scarcity: High-quality RL datasets for agentic QA are rare.
  • Rollout Efficiency: Multi-turn reasoning and heavy tool use can slow training.

Approach

Kimi-Researcher is trained via end-to-end reinforcement learning. We observe a consistent improvement in agent performance across different domains.

Training accuracy throughout RL process Model performance on internal datasets

Training data

To address the scarcity of high-quality agentic datasets, we engineered our training corpus with two complementary objectives.

First, we developed a suite of challenging, tool-centric tasks designed to promote robust tool-use learning. These prompts are deliberately constructed such that solving the task requires invoking specific tools — making naive approaches either infeasible or substantially less efficient.

Tool invocation rates

Second, we curated and synthesized reasoning-intensive tasks to reinforce the agent's core cognitive abilities:

  • Math and Code Reasoning: Tasks that target logical inference, algorithmic problem-solving, and sequential computation.
  • Hard Search: Scenarios where the agent must iteratively search, synthesize, and reason within context constraints.

Effectiveness of synthetic tasks

RL training

The model is primarily trained using the REINFORCE algorithm. Key factors contributing to stable training:

  • On-policy Training: Critical to generate strict on-policy data. During training, we disable LLM engine mechanisms like toolcall format enforcers.
  • Negative Sample Control: Strategically discarding some negative samples to prevent entropy collapse during RL training.

Kimi-Researcher uses outcome rewards for training:

  • Format Reward: Penalized for trajectories with invalid tool calls or exceeding maximum context/iteration limits.
  • Correctness Reward: Rewards based on comparison between model's answer and ground truth.

A gamma-decay factor is applied to correct trajectories, encouraging the model to discover shorter, more efficient exploration.

Context management

A long-horizon research trajectory may involve massive observation contexts. We designed a context-management mechanism that allows the model to retain important information while discarding unnecessary documents, extending a single rollout trajectory to over 50 iterations. An early ablation study shows that a model trained with context management uses 30% more iterations.

Large-scale agent RL infrastructure

Large-scale agent RL infrastructure

To address efficiency and stability challenges, we developed infrastructure with:

  • Fully asynchronous rollout: Efficiently orchestrates actor rollouts, environmental interactions, and reward calculations in parallel.
  • Turn-level partial rollout: Tasks exceeding time budget are saved to replay buffer, with remaining turns executed with updated model weights (1.5x acceleration).
  • Robust sandbox environment: Unified sandbox architecture with Kubernetes-based hybrid cloud for dynamic resource allocation.

Emerging agentic capacities

During end-to-end reinforcement learning, we observed several notable emergent abilities:

Resolving conflicting information: Kimi-Researcher resolves inconsistencies through iterative hypothesis refinement and self-correction.

Caution and rigor: Even for seemingly straightforward questions, it deliberately performs additional searches and cross-validates information before answering.

Use cases

  • Academic research
  • Legal & regulatory insights
  • Obscure information retrieval
  • Clinical evidence review
  • Corporate financial analysis

ESC

输入关键词开始搜索

支持搜索标题、内容、标签