{ "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": "2610.19820", "title": "ActionOPD: Distilling Decisions, Not Tokens", "summary": "ActionOPD replaces offline imitation and token-level on-policy distillation with action-level on-policy distillation, rewarding student actions by semantic agreement with sampled teacher actions, so teacher token probabilities and shared tokenization are no longer required. It costs an on-policy RL training loop with sampling from both student and teacher at each state, which is substantially more compute- and orchestration-intensive than offline supervised fine-tuning, though the teacher can be a black-box API. It can fail if the sampled teacher actions are unrepresentative or the semantic-agreement reward is noisy, and the gains are demonstrated only on curated tool-use tasks, so transfer to other agent workloads is unverified.", "rationale": "The method fits constrained infrastructure by allowing a black-box API teacher and small students, but the on-policy RL training cost and task-specific evidence base justify a small-scale pilot rather than direct adoption.", "technique": "Action-level on-policy distillation for agents", "family": "destilacao", "recommendation": "testar", "published": "2026-10-14", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.19820", "independent_implementations": 2, "total_implementations": 3, "stars": 7, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "ActionOPD replaces offline imitation and token-level on-policy distillation with action-level on-policy distillation, rewarding student actions by semantic agreement with sampled teacher actions, so teacher token probabilities and shared tokenization are no longer required. It costs an on-policy RL training loop with sampling from both student and teacher at each state, which is substantially more compute- and orchestration-intensive than offline supervised fine-tuning, though the teacher can be a black-box API. It can fail if the sampled teacher actions are unrepresentative or the semantic-agreement reward is noisy, and the gains are demonstrated only on curated tool-use tasks, so transfer to other agent workloads is unverified.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.19820", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Average gains of 3.3-6.9 percentage points across four out-of-distribution benchmarks for Qwen3-4B/8B students; competitive with offline distillation and token-level OPD on τ²-bench.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.19820", "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": "" }