{ "schema_version": 1, "builder_plan": { "scenario": "An assistant chooses the wrong API or sends arguments your product cannot use.", "change": "Apply the proposed tool-selection or argument-generation change in a sandbox.", "baseline": "Your current agent with the same tool schemas and task budget.", "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": "Correct completed actions, invalid calls, retries, latency, and cost per completed task.", "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": "2608.30147", "title": "CAST: Critique-Aware Supervision for Training Reliable Long-Horizon Tool-Calling Agents", "summary": "CAST replaces prompt-based critique agents and outcome-only reinforcement signals with a trained critique model that converts sparse task outcomes into action-level rationales, which then supervise policy fine-tuning. The cost is a trajectory-analysis and rationale-synthesis pipeline plus fine-tuning of both a critique model and a policy model, which fits a single 24 GB GPU only for smaller Qwen3 variants with parameter-efficient methods and adds training-data engineering complexity. It can fail when synthesized rationales are incorrect or unfaithful under partial observability, when the target domain lacks sufficient logged trajectories, or when error modes shift after deployment, since the reported gains are benchmark-specific.", "rationale": "The method targets a real production need (intercepting irreversible tool actions) and is reproducible at small scale with LoRA fine-tuning, but its benefit depends on having domain trajectories and validating rationale quality on the reader's own workload.", "technique": "Critique-aware fine-tuning for tool-calling agents", "family": "uso_de_ferramenta", "recommendation": "testar", "published": "2026-08-31", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2608.30147", "independent_implementations": 0, "total_implementations": 1, "stars": 28, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "CAST replaces prompt-based critique agents and outcome-only reinforcement signals with a trained critique model that converts sparse task outcomes into action-level rationales, which then supervise policy fine-tuning. The cost is a trajectory-analysis and rationale-synthesis pipeline plus fine-tuning of both a critique model and a policy model, which fits a single 24 GB GPU only for smaller Qwen3 variants with parameter-efficient methods and adds training-data engineering complexity. It can fail when synthesized rationales are incorrect or unfaithful under partial observability, when the target domain lacks sufficient logged trajectories, or when error modes shift after deployment, since the reported gains are benchmark-specific.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2608.30147", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Fine-tuned Qwen3-family models outperform GPT-OSS-120B by over 10 pass^4 points on Retail tasks and gain an additional 9% on out-of-domain Telehealth tasks.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2608.30147", "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_formula", "core_kind": "algorithm", "equations_fetched_at": "2026-09-23", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }