{ "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": "2610.28386", "title": "VisualEvolve: A Skill Self-Evolution Framework for Visual Reasoning", "summary": "Replaces reinforcement-learning fine-tuning of tool-use behavior with an external, evolving library of executable visual skills selected and refined through a Select-Execute-Evolve-Evaluate loop, with snapshot selection and progressive evidence access to control regressions. Cost is inference-time API or GPU compute for trajectory generation, skill evaluation on a held-out validation set, and engineering effort to maintain the skill library and evolution harness rather than model training. Failures include skill regressions propagating if snapshot selection is misconfigured, evolution driven by non-representative validation tasks, and accumulated API or compute spend when many evolution cycles are needed.", "rationale": "It fits constrained infrastructure because it requires no parameter updates and works with API-based VLMs, but the gains are benchmark-specific and the framework's added pipeline complexity requires validation on the reader's own visual tasks before production adoption.", "technique": "Self-evolving visual skill library without retraining", "family": "uso_de_ferramenta", "recommendation": "testar", "published": "2026-10-20", "as_of": "2026-10-21", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.28386", "independent_implementations": 0, "total_implementations": 1, "stars": 0, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "Replaces reinforcement-learning fine-tuning of tool-use behavior with an external, evolving library of executable visual skills selected and refined through a Select-Execute-Evolve-Evaluate loop, with snapshot selection and progressive evidence access to control regressions. Cost is inference-time API or GPU compute for trajectory generation, skill evaluation on a held-out validation set, and engineering effort to maintain the skill library and evolution harness rather than model training. Failures include skill regressions propagating if snapshot selection is misconfigured, evolution driven by non-representative validation tasks, and accumulated API or compute spend when many evolution cycles are needed.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.28386", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Improves over the initial skill library in 23 of 24 backbone/benchmark settings, with a maximum gain of 15.1 percentage points on VisualProbe Hard (Qwen3.5-Plus), using over an order of magnitude fewer trajectories than representative RL-based methods.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.28386", "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": "" }