r2l-e-0004-yue-rlvr-boundary.json

{
  "schema_version": "1.0.0",
  "record_id": "r2l-e-0004-yue-rlvr-boundary",
  "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": "yue2025rlcapacity",
      "title": "Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?",
      "persistent_identifier": {
        "type": "arxiv",
        "value": "2504.13837",
        "url": "https://arxiv.org/abs/2504.13837"
      },
      "publication_version": "source version cited by the manuscript",
      "publication_date": "2025",
      "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": "RLVR-trained models across source-reported model families",
    "version_or_checkpoint": "source-reported paper configuration",
    "lineage": "rl_for_deliberation",
    "implementation_reference": null
  },
  "task": {
    "domain": "language model reasoning or tool use",
    "name": "language and visual reasoning benchmarks",
    "version": "source-reported version",
    "split": "source-reported benchmark and model-family combinations",
    "environment_or_dataset": "language and visual reasoning benchmarks",
    "unit_of_analysis": "one evaluation item",
    "evaluation_protocol": "Compare base and RLVR-trained pass@k curves at small and large k."
  },
  "configuration": {
    "configuration_id": "yue_rlvr_large_k_comparison",
    "training_summary": "Source-reported RLVR configurations compared with their corresponding base models.",
    "inference_summary": "Repeated sampling at increasing k to estimate pass@k and the sampled capability boundary.",
    "parameters": []
  },
  "r2l_interface": {
    "reasoning_object": {
      "type": "generated reasoning trajectories",
      "description": "generated reasoning trajectories",
      "origin": "learned",
      "explicit": true
    },
    "functional_coupling": {
      "operations": [
        "generated",
        "verified",
        "consumed"
      ],
      "affects": [
        "policy_optimization",
        "action_selection"
      ]
    },
    "learning_signals": [
      "verifiable outcome rewards"
    ],
    "integration_points": [
      "reward"
    ],
    "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": "samples_per_problem",
      "type": "samples",
      "phase": "evaluation",
      "reporting_status": "not_reported",
      "quantity": null,
      "unit": null,
      "accounting_scope": "Pass@k sampling budget",
      "notes": "The manuscript extraction does not assert this quantity; verify it from the source."
    }
  ],
  "controls": {
    "comparators": [
      {
        "comparator_id": "base_model",
        "name": "Corresponding base model",
        "configuration_summary": "Base checkpoint sampled under the source's comparison protocol.",
        "resource_matching": [
          {
            "resource_id": "samples_per_problem",
            "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": "large_k_boundary",
      "metric": "pass@k capability-boundary comparison",
      "evaluation_layer": "outcome",
      "population_or_split": "multiple language and visual reasoning benchmarks",
      "estimate": {
        "kind": "text",
        "value": "At large k, base models match or exceed RLVR variants in the reported comparisons.",
        "unit": null,
        "denominator": {
          "status": "not_reported",
          "value": null,
          "description": "The source-local extraction does not assert a denominator."
        },
        "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": "base_model",
      "source_locator": "Yue et al., abstract, Figure 1, and Sections 3--4 on pass@k across base and RLVR-trained models",
      "permitted_interpretation": "The study reports improved low-k sampling efficiency without uniform expansion of the large-k sampled capability boundary."
    }
  ],
  "diagnostics": [],
  "claim_scope": {
    "supported_interpretations": [
      "The reported comparisons challenge a universal claim that RLVR expands the base model's sampled reasoning boundary."
    ],
    "unsupported_inferences": [
      "The record does not show that RLVR has no value for pass@1, efficiency, or other model families."
    ]
  },
  "limitations": [
    {
      "limitation_id": "yue_boundary_proxy",
      "source": "extractor_identified",
      "category": "external_validity",
      "description": "Large-k pass@k is an empirical proxy for sampled capability, not proof of the model's complete reasoning support.",
      "affected_measurement_ids": [
        "large_k_boundary"
      ]
    }
  ],
  "extensions": {
    "review_study_id": "R2L-S-0060",
    "evidence_record_id": "R2L-E-0004",
    "claim_direction": "counterevidence"
  }
}