{
  "schema_version": 1,
  "id": "evt-2017-deep-network-guided-proof-search",
  "slug": "deep-network-guided-proof-search",
  "title": {
    "en": "Deep networks are integrated directly into automated theorem-prover search",
    "zh-CN": "深度网络被直接接入自动定理证明器的内部搜索"
  },
  "dates": {
    "event": "2017-01-24",
    "first_public_claim": "2017-01-24",
    "paper_release": "2017-01-24",
    "artifact_release": null,
    "independently_verified": null,
    "added_to_chronicle": "2026-09-15",
    "last_updated": "2026-09-15"
  },
  "event_types": [
    "system"
  ],
  "significance": {
    "tier": "H2",
    "rationale": {
      "en": "Deep Network Guided Proof Search was an important early step from learned premise selection toward learned control inside the proof search itself. By training neural models on Mizar-derived proof traces and inserting their predictions into E prover's processed-clause selection, the work established a concrete deep-learning route for internal automated-theorem-prover guidance before later reinforcement-learning and proof-assistant-native systems.",
      "zh-CN": "Deep Network Guided Proof Search 是从“学习式前提选择”进一步走向“直接学习证明器内部搜索控制”的重要早期节点。该工作从 Mizar 衍生的证明轨迹训练神经模型，并把预测直接接入 E prover 的 processed-clause selection，在后来的强化学习与证明助手原生系统出现之前，建立了深度学习引导自动定理证明内部搜索的具体路线。"
    }
  },
  "mathematical_novelty": {
    "type": "tooling_only"
  },
  "interfaces": [
    "formal",
    "symbolic"
  ],
  "claim": {
    "en": "Sarah Loos, Geoffrey Irving, Christian Szegedy and Cezary Kaliszyk trained deep neural networks on existing automated proofs of Mizar statements and used them to guide processed-clause selection inside E prover. The paper reported finding first-order proofs for a subset of translated Mizar theorems that previously lacked ATP-generated proofs, increasing the corpus coverage with ATP proofs from about 56% to 59%.",
    "zh-CN": "Sarah Loos、Geoffrey Irving、Christian Szegedy 与 Cezary Kaliszyk 使用既有 Mizar 自动证明轨迹训练深度神经网络，并将其用于引导 E prover 内部的 processed-clause selection。论文报告称，该方法为一部分此前没有 ATP 证明的 Mizar 翻译定理找到了新的 first-order proofs，使语料中具有 ATP 证明的覆盖率从约 56% 提升到 59%。"
  },
  "summary": {
    "en": "The system learns from clause-level traces of successful automated proofs and ranks generated clauses during saturation-style proof search. Because neural evaluation is slower than hand-engineered heuristics, the authors used hybrid guidance and a two-phase search strategy that lets the learned model improve the search trajectory before faster conventional heuristics finish the proof attempt.",
    "zh-CN": "该系统从成功自动证明的 clause-level 轨迹中学习，并在 saturation-style proof search 中对生成的 clauses 进行排序。由于神经网络评估慢于人工设计启发式，作者采用混合引导与 two-phase search：先利用学习模型改善搜索轨迹，再由更快的传统启发式完成后续证明尝试。"
  },
  "why_it_matters": {
    "en": "The work demonstrated that neural models could influence the inner loop of a mature automated theorem prover rather than acting only as an upstream filter. That distinction became important in later work on learned clause guidance, reinforcement-learning proof search and search-in-the-loop theorem proving.",
    "zh-CN": "该工作展示了神经模型不仅可以作为上游过滤器，还可以直接影响成熟自动定理证明器的内部搜索循环。这一区别后来成为学习式 clause guidance、强化学习证明搜索以及 search-in-the-loop theorem proving 的重要技术前史。"
  },
  "ai_role": {
    "level": "autonomous_primary",
    "description": {
      "en": "Learned neural models ranked processed clauses during automated proof search, while E prover carried out the symbolic inference and constructed successful first-order proofs without human step-by-step intervention.",
      "zh-CN": "学习式神经模型在自动证明搜索中排序 processed clauses，由 E prover 执行符号推理并构造成功的一阶逻辑证明，不需要人类逐步干预证明动作。"
    }
  },
  "human_contribution": {
    "description": {
      "en": "Researchers designed the neural clause-guidance models, proof-trace training data, hybrid and two-phase search strategy, E-prover integration and evaluation over Mizar-derived problems.",
      "zh-CN": "研究者设计神经 clause-guidance 模型、证明轨迹训练数据、混合与两阶段搜索策略、E prover 集成以及基于 Mizar 衍生问题的评测方案。"
    }
  },
  "systems": [
    "e-prover",
    "deep-network-guided-proof-search"
  ],
  "organizations": [
    "google-research",
    "university-of-innsbruck"
  ],
  "people": [
    "sarah-loos",
    "geoffrey-irving",
    "christian-szegedy",
    "cezary-kaliszyk"
  ],
  "problems": [
    "automated-theorem-proving",
    "mizar-theorem-proving"
  ],
  "methods": [
    "deep-learning",
    "clause-selection",
    "proof-search",
    "saturation-based-theorem-proving",
    "supervised-learning"
  ],
  "verification": {
    "evidence_level": "E2",
    "status": "paper_released",
    "formal_assurance": "none",
    "history": [
      {
        "date": "2017-01-24",
        "status": "paper_released",
        "evidence_level": "E2",
        "note": {
          "en": "The arXiv and LPAR research records establish the neural proof-search guidance method and the reported Mizar-derived theorem-proving results. The Chronicle does not record a directly inspectable proof artifact or independent replay for this event, so formal assurance remains `none`.",
          "zh-CN": "arXiv 与 LPAR 研究记录确认了神经证明搜索引导方法及论文报告的 Mizar 衍生定理证明结果。Chronicle 当前未记录可直接检查的证明 artifact 或独立 replay，因此 formal assurance 保持为 `none`。"
        }
      }
    ]
  },
  "sources": [
    {
      "id": "arxiv-deep-network-proof-search-1701.06972",
      "tier": "S1",
      "type": "paper",
      "title": "Deep Network Guided Proof Search",
      "url": "https://arxiv.org/abs/1701.06972",
      "primary": true,
      "published_date": "2017-01-24",
      "accessed_at": "2026-09-15"
    },
    {
      "id": "easychair-deep-network-proof-search-lpar2017",
      "tier": "S1",
      "type": "paper",
      "title": "Deep Network Guided Proof Search — LPAR-21",
      "url": "https://easychair.org/publications/paper/ND13",
      "primary": true,
      "published_date": "2017-05-04",
      "accessed_at": "2026-09-15"
    },
    {
      "id": "google-research-deep-network-proof-search",
      "tier": "S2",
      "type": "official_announcement",
      "title": "Deep Network Guided Proof Search — Google Research publication record",
      "url": "https://research.google/pubs/deep-network-guided-proof-search-2/",
      "primary": false,
      "published_date": null,
      "accessed_at": "2026-09-15"
    }
  ],
  "artifacts": [],
  "relationships": {
    "predecessors": [
      "evt-2016-deepmath"
    ],
    "successors": [
      "evt-2018-rl-theorem-proving"
    ],
    "related": []
  },
  "timeline": {
    "featured": true,
    "significance": "primary"
  },
  "tags": [
    "theorem-proving",
    "deep-learning",
    "e-prover",
    "mizar",
    "clause-selection",
    "proof-search",
    "automated-theorem-proving"
  ],
  "editorial": {
    "created_at": "2026-09-15",
    "last_reviewed": "2026-09-15",
    "reviewers": [
      "Charlie-Wang-03"
    ]
  },
  "corrections": []
}