r2l-e-0001-lightman-prm.json

{
  "schema_version": "1.0.0",
  "record_id": "r2l-e-0001-lightman-prm",
  "record_status": "draft",
  "disclaimer": "DRAFT SOURCE EXTRACTION: values are transcribed from the cited source and remain ineligible for synthesis until two named authors independently verify them.",
  "provenance": {
    "evidence_class": "reported_by_source",
    "source": {
      "citation_key": "lightman2023let",
      "title": "Let's Verify Step by Step",
      "persistent_identifier": {
        "type": "arxiv",
        "value": "2305.20050",
        "url": "https://arxiv.org/abs/2305.20050"
      },
      "publication_version": "source version cited by the manuscript",
      "publication_date": "2023",
      "accessed_at": "2026-07-23"
    },
    "extraction": {
      "extracted_at": "2026-07-23",
      "extracted_by": "Codex-assisted manuscript revision",
      "verification_status": "unverified",
      "verification_checks": []
    },
    "synthesis_eligibility": {
      "status": "pending",
      "reason": "Pending independent verification by two named manuscript authors."
    },
    "evidence_details": {
      "input_record_ids": [],
      "formula": null,
      "normalization": null,
      "uncertainty_procedure": null,
      "assumptions": [],
      "theorem_locator": null,
      "supporting_record_ids": [],
      "synthesis_rationale": null
    },
    "change_log": [
      {
        "changed_at": "2026-07-23",
        "changed_by": "Codex-assisted manuscript revision",
        "reason": "Created a source-reported draft for a manuscript quantitative claim."
      }
    ]
  },
  "system": {
    "name": "Process-supervised reward model",
    "version_or_checkpoint": "source-reported paper configuration",
    "lineage": "rl_for_deliberation",
    "implementation_reference": null
  },
  "task": {
    "domain": "language model reasoning or tool use",
    "name": "MATH",
    "version": "source-reported version",
    "split": "representative subset of 500 MATH test problems",
    "environment_or_dataset": "MATH",
    "unit_of_analysis": "one evaluation item",
    "evaluation_protocol": "Best-of-1,860 selection using the process reward model."
  },
  "configuration": {
    "configuration_id": "lightman_prm_best_of_1860",
    "training_summary": "The source trains a reward model from human process labels; this record does not treat the generator as RL-trained.",
    "inference_summary": "Select the highest-scored candidate among 1,860 generated solutions.",
    "parameters": []
  },
  "r2l_interface": {
    "reasoning_object": {
      "type": "generated mathematical solution with step boundaries",
      "description": "generated mathematical solution with step boundaries",
      "origin": "learned",
      "explicit": true
    },
    "functional_coupling": {
      "operations": [
        "generated",
        "verified",
        "consumed"
      ],
      "affects": [
        "policy_optimization",
        "action_selection"
      ]
    },
    "learning_signals": [
      "human process labels"
    ],
    "integration_points": [
      "reward",
      "search"
    ],
    "grounding": {
      "target": "The task inputs, executable checks, or reference answers declared by the source.",
      "mechanism": "The source-specific reward or verifier connects the explicit trace to task feedback.",
      "diagnostic_ids": []
    }
  },
  "resources": [
    {
      "resource_id": "candidate_solutions",
      "type": "samples",
      "phase": "inference",
      "reporting_status": "reported",
      "quantity": 1860,
      "unit": "candidate solutions per problem",
      "accounting_scope": "Candidate pool used for each best-of-1,860 selection.",
      "notes": "Source-local selection budget."
    }
  ],
  "controls": {
    "comparators": [
      {
        "comparator_id": "outcome_reward_model",
        "name": "Outcome-supervised reward model",
        "configuration_summary": "Outcome-supervised model evaluated with the same selection protocol.",
        "resource_matching": [
          {
            "resource_id": "candidate_solutions",
            "status": "not_reported",
            "explanation": "Resource parity requires confirmation during human verification."
          }
        ]
      }
    ],
    "no_comparator_reason": null,
    "ablations": [],
    "leakage_and_contamination_checks": [
      "Use only source-reported contamination checks; none are inferred in this extraction."
    ],
    "privileged_information_checks": [
      "Verifier, reference-answer, and tool access must be checked during human verification."
    ]
  },
  "outcomes": [
    {
      "measurement_id": "math_best_of_1860",
      "metric": "MATH solved percent",
      "evaluation_layer": "outcome",
      "population_or_split": "representative subset of 500 MATH test problems",
      "estimate": {
        "kind": "numeric",
        "value": 78.2,
        "unit": "percent",
        "denominator": {
          "status": "reported",
          "value": 500,
          "description": "500 evaluated cases"
        },
        "aggregation": "Source-reported aggregate for the stated configuration."
      },
      "uncertainty": {
        "status": "not_reported",
        "description": "The cited result does not report an uncertainty interval for this value."
      },
      "comparator_id": "outcome_reward_model",
      "source_locator": "Lightman et al., Figure 3, Best-of-1860 test performance, product reduction with neutral labels treated as positive",
      "permitted_interpretation": "The PRM-selected solution is correct for 78.2 percent of the stated 500-problem subset under best-of-1860 selection."
    }
  ],
  "diagnostics": [],
  "claim_scope": {
    "supported_interpretations": [
      "Process-supervised selection reaches the reported source-local result under the stated sampling budget."
    ],
    "unsupported_inferences": [
      "The record does not establish universal superiority of process supervision or an RL-trained generator."
    ]
  },
  "limitations": [
    {
      "limitation_id": "lightman_component_boundary",
      "source": "extractor_identified",
      "category": "external_validity",
      "description": "The result evaluates verifier-based selection and does not show that the generator was trained with RL.",
      "affected_measurement_ids": [
        "math_best_of_1860"
      ]
    },
    {
      "limitation_id": "lightman_uncertainty",
      "source": "extractor_identified",
      "category": "missing_uncertainty",
      "description": "No uncertainty interval is retained for the reported percentage.",
      "affected_measurement_ids": [
        "math_best_of_1860"
      ]
    }
  ],
  "extensions": {
    "review_study_id": "R2L-S-0023",
    "evidence_record_id": "R2L-E-0001",
    "claim_direction": "positive_local"
  }
}