{
  "schema_version": 1,
  "id": "evt-2025-deepseek-prover-v2",
  "slug": "deepseek-prover-v2-subgoal-rl",
  "title": {
    "en": "DeepSeek-Prover-V2 combines informal subgoal decomposition with Lean reinforcement learning",
    "zh-CN": "DeepSeek-Prover-V2 将非形式子目标分解与 Lean 强化学习结合"
  },
  "dates": {
    "event": "2025-04-30",
    "first_public_claim": "2025-04-30",
    "paper_release": "2025-04-30",
    "artifact_release": null,
    "independently_verified": null,
    "added_to_chronicle": "2026-09-10",
    "last_updated": "2026-09-11"
  },
  "event_types": [
    "proof",
    "system"
  ],
  "significance": {
    "tier": "H2",
    "rationale": {
      "en": "DeepSeek-Prover-V2 narrowed the gap between natural-language reasoning and formal proof generation at much larger scale.",
      "zh-CN": "DeepSeek-Prover-V2 在更大规模上连接自然语言推理与形式证明生成，缩小两者能力差距。"
    }
  },
  "mathematical_novelty": {
    "type": "benchmark_result"
  },
  "interfaces": [
    "informal",
    "formal",
    "hybrid"
  ],
  "claim": {
    "en": "DeepSeek-Prover-V2-671B reported 88.9% on miniF2F-test and solved 49 of 658 PutnamBench problems.",
    "zh-CN": "DeepSeek-Prover-V2-671B 报告 miniF2F-test 88.9%，并解出 PutnamBench 658 题中的 49 题。"
  },
  "summary": {
    "en": "DeepSeek-V3 first decomposes problems into subgoals; successful subgoal proofs become cold-start reasoning traces for reinforcement learning.",
    "zh-CN": "DeepSeek-V3 先把问题分解为子目标，成功的子目标证明被整理为强化学习的冷启动推理轨迹。"
  },
  "why_it_matters": {
    "en": "The design made informal planning an explicit training scaffold for high-capacity formal theorem provers.",
    "zh-CN": "该设计把非形式规划显式变成高容量形式定理证明器的训练脚手架。"
  },
  "ai_role": {
    "level": "ai_primary_human_verified",
    "description": {
      "en": "Models decompose problems, generate Lean proofs and improve through verifier-backed reinforcement learning.",
      "zh-CN": "模型分解问题、生成 Lean 证明，并通过验证器反馈强化学习提升能力。"
    }
  },
  "human_contribution": {
    "description": {
      "en": "Researchers designed recursive data generation, subgoal decomposition and RL training.",
      "zh-CN": "研究者设计递归数据生成、子目标分解与强化学习训练。"
    }
  },
  "systems": [
    "deepseek-prover-v2",
    "lean"
  ],
  "organizations": [
    "deepseek"
  ],
  "people": [],
  "problems": [
    "minif2f",
    "putnambench",
    "proverbench"
  ],
  "methods": [
    "subgoal-decomposition",
    "reinforcement-learning",
    "formal-proof"
  ],
  "verification": {
    "evidence_level": "E3",
    "status": "paper_released",
    "formal_assurance": "machine_checked",
    "history": [
      {
        "date": "2025-04-30",
        "status": "paper_released",
        "evidence_level": "E2",
        "note": {
          "en": "The Chronicle records this event from the cited primary or official source; the status does not imply independent Chronicle verification.",
          "zh-CN": "Chronicle 依据所引 primary / official source 记录该事件；该状态不表示 Chronicle 已完成独立验证。"
        }
      },
      {
        "date": "2026-09-11",
        "status": "paper_released",
        "evidence_level": "E3",
        "note": {
          "en": "This corpus refresh added a directly inspectable public artifact maintained by the authors or releasing organization. E3 records artifact-level inspectability and does not imply independent verification of the underlying scientific or mathematical claim.",
          "zh-CN": "本轮 corpus 刷新增补了由作者或发布机构维护、可直接检查的公开产物。E3 表示证据已达到可检查 artifact 层级，并不等同于底层科学或数学主张已获得独立验证。"
        }
      }
    ]
  },
  "sources": [
    {
      "id": "arxiv-deepseek-prover-v2-2504.21801",
      "tier": "S1",
      "type": "paper",
      "title": "DeepSeek-Prover-V2: Advancing Formal Mathematical Reasoning via Reinforcement Learning for Subgoal Decomposition",
      "url": "https://arxiv.org/abs/2504.21801",
      "primary": true,
      "published_date": "2025-04-30",
      "accessed_at": "2026-09-10"
    },
    {
      "id": "github-deepseek-prover-v2",
      "tier": "S1",
      "type": "repository",
      "title": "DeepSeek-Prover-V2 models, benchmark and solution artifacts",
      "url": "https://github.com/deepseek-ai/DeepSeek-Prover-V2",
      "primary": true,
      "published_date": null,
      "accessed_at": "2026-09-11"
    }
  ],
  "artifacts": [
    {
      "id": "deepseek-prover-v2-repository",
      "kind": "code",
      "title": "DeepSeek-Prover-V2 models, benchmark and proof artifacts",
      "url": "https://github.com/deepseek-ai/DeepSeek-Prover-V2"
    }
  ],
  "relationships": {
    "predecessors": [
      "evt-2024-deepseek-prover-v1-5",
      "evt-2025-deepseek-r1"
    ],
    "successors": [],
    "related": []
  },
  "timeline": {
    "featured": true,
    "significance": "primary"
  },
  "tags": [
    "deepseek",
    "lean",
    "reinforcement-learning",
    "subgoal-decomposition"
  ],
  "editorial": {
    "created_at": "2026-09-10",
    "last_reviewed": "2026-09-11",
    "reviewers": [
      "Charlie-Wang-03"
    ]
  },
  "corrections": []
}