{ "schema_version": 1, "builder_plan": { "scenario": "Your coding assistant makes changes that still need substantial repair before merging.", "change": "Try the technique on held-out issues from a repository with reliable tests.", "baseline": "The same coding agent, repository snapshot, tools, and token 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 patches, regression failures, human repair time, and total cost per accepted change.", "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": "2609.37968", "title": "SelfSearch: Reward-Free Search for Self-Improving Agents", "summary": "SelfSearch replaces evaluation-guided agent search (repeated downstream benchmark scoring during optimization) with reward-free self-modification driven by records of prior improvement episodes. Cost is low, on the order of single-digit dollars per search plus the engineering effort to log and replay self-modification traces, with no extra serving memory or latency. It can fail because gains are benchmark- and model-specific (population-mean improvements with individual runs varying up to 11.2 points), self-modifications may overfit to logged experience, and unconstrained agent self-editing carries regression and safety risks in production harnesses.", "rationale": "The method is cheap enough to run on a limited cloud budget and API-only models, but reported gains are population-level and task-dependent, so the reader should validate it on their own agent workload before adopting it as standard practice.", "technique": "Reward-free self-modification search for agents", "family": "agentes_de_codigo", "recommendation": "testar", "published": "2026-09-29", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.37968", "independent_implementations": 2, "total_implementations": 3, "stars": 132, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "SelfSearch replaces evaluation-guided agent search (repeated downstream benchmark scoring during optimization) with reward-free self-modification driven by records of prior improvement episodes. Cost is low, on the order of single-digit dollars per search plus the engineering effort to log and replay self-modification traces, with no extra serving memory or latency. It can fail because gains are benchmark- and model-specific (population-mean improvements with individual runs varying up to 11.2 points), self-modifications may overfit to logged experience, and unconstrained agent self-editing carries regression and safety risks in production harnesses.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.37968", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "On SWE-bench Multilingual the evolved agent improves success by 5.0 percentage points while reducing execution cost by 38.5% on co-solved tasks; a $4.03 search produced a harness matching the top public harness (82.0% on Terminal-Bench 2.1 with DeepSeek V4 Flash).", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.37968", "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": "" }