{ "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": "2609.30906", "title": "ToolSearcher: Optimizing Tool Selection at Scale via Reinforcement Learning", "summary": "It replaces prompting an off-the-shelf LLM to pick tools from a large repository by RL fine-tuning a policy for multi-turn search with category-constrained discrimination, event-level search rewards, and trajectory-aligned credit assignment. The cost is a full RL training pipeline with reward design, multi-turn rollout collection, and curated tool-selection data, which is expensive to build and reproduce on a 24 GB GPU or API budget. It can fail through reward hacking on proxy search metrics, overfitting to the benchmark's tool repository, and poor transfer to a production tool catalog with different categories and APIs.", "rationale": "The method requires RL infrastructure, reward engineering, and training data that exceed the reader's constrained setup, so it is worth tracking but not adopting until pretrained checkpoints or a lightweight adaptation appear.", "technique": "RL-trained multi-turn tool selection at scale", "family": "uso_de_ferramenta", "recommendation": "observar", "published": "2026-09-25", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.30906", "independent_implementations": 1, "total_implementations": 2, "stars": 131, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "It replaces prompting an off-the-shelf LLM to pick tools from a large repository by RL fine-tuning a policy for multi-turn search with category-constrained discrimination, event-level search rewards, and trajectory-aligned credit assignment. The cost is a full RL training pipeline with reward design, multi-turn rollout collection, and curated tool-selection data, which is expensive to build and reproduce on a 24 GB GPU or API budget. It can fail through reward hacking on proxy search metrics, overfitting to the benchmark's tool repository, and poor transfer to a production tool catalog with different categories and APIs.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.30906", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Consistently outperforms strong baselines on large-scale tool selection benchmarks involving iterative search and tool composition; no multiplicative factor or absolute number is stated in the abstract.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.30906", "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": "" }