{
  "schema_version": 1,
  "builder_plan": {
    "scenario": "A narrow, repeated product task may not need your largest model.",
    "change": "Check whether a released student model fits the task before generating training data.",
    "baseline": "Your current model and an off-the-shelf small model on the same held-out tasks.",
    "steps": [
      "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.",
      "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.",
      "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.",
      "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test."
    ],
    "measure": "Student quality, serving cost, teacher calls, training cost, and failures on unusual inputs.",
    "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation."
  },
  "builder_review": null,
  "arxiv_id": "2609.39306",
  "title": "ReSAIL: Mitigating Collapse in Iterative Agent Self-Distillation",
  "summary": "ReSAIL replaces naive iterative self-distillation, where deployment and privileged-information performance collapse across cycles, with sensitivity-guided step selection, trajectory-balanced distillation losses, and KL regularization of PI-conditioned outputs toward a frozen teacher. It costs additional teacher forward passes per interaction step, per-step sensitivity computation, and retention of the frozen teacher during each training cycle, on top of repeated fine-tuning runs. It can fail when privileged information is unavailable or weakly informative at deployment, when trajectories are too few for selection statistics to be meaningful, and it is validated only on three benchmarks over three cycles, so collapse mitigation beyond that horizon is unverified.",
  "rationale": "The method directly targets a failure mode a small team would hit when recycling self-generated agent data, but adopting it presupposes an existing PI-based iterative training pipeline and workload-specific validation on tasks unlike the paper's benchmarks.",
  "technique": "Selective iterative self-distillation for PI-conditioned agents",
  "family": "destilacao",
  "recommendation": "testar",
  "published": "2026-09-30",
  "as_of": "2026-10-18",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2609.39306",
  "independent_implementations": 0,
  "total_implementations": 0,
  "stars": 0,
  "citations": 0,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "ReSAIL replaces naive iterative self-distillation, where deployment and privileged-information performance collapse across cycles, with sensitivity-guided step selection, trajectory-balanced distillation losses, and KL regularization of PI-conditioned outputs toward a frozen teacher. It costs additional teacher forward passes per interaction step, per-step sensitivity computation, and retention of the frozen teacher during each training cycle, on top of repeated fine-tuning runs. It can fail when privileged information is unavailable or weakly informative at deployment, when trajectories are too few for selection statistics to be meaningful, and it is validated only on three benchmarks over three cycles, so collapse mitigation beyond that horizon is unverified.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2609.39306",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "Average absolute gain of 22.5% in final-cycle success rates over three cycles when added to self-distillation baselines on ALFWorld and TextCraft, plus improved action prediction for multimodal GUI agents on AITZ.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2609.39306",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "exposure_map": [
    {
      "dimension": "quality",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compute",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "latency",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "operations",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compatibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "security",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "data_and_training",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "reproducibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "risks": [],
  "minimum_test": [],
  "open_questions": [],
  "independent_tests": [],
  "lab_reproductions": [],
  "report_available": false,
  "validation_tier": "unknown",
  "evidence_tier": "unknown",
  "infrastructure_basis": "unknown",
  "equations": [],
  "equations_status": "not_fetched",
  "core_kind": "",
  "equations_fetched_at": "",
  "equations_source_url": "",
  "equations_source_sha256": "",
  "exposures_reviewed_at": "",
  "exposures_source_sha256": ""
}