{ "schema_version": 1, "builder_plan": { "scenario": "An agent looks good in demos, but you cannot tell which product tasks it fails.", "change": "Use the study's evaluation idea to inspect a small set of tasks with human-checked answers.", "baseline": "Your current evaluator and a manually reviewed reference set.", "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": "Agreement with human review, missed failures, false alarms, and evaluation cost per 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": "2609.37515", "title": "Hierarchical Compression of Vision-Language Model Benchmarks", "summary": "PRIMEBench replaces running full vision-language benchmark suites by hierarchically cleaning non-visual or saturated items, selecting one representative benchmark per capability category, and pruning items via a vision-aware variance score computed from multimodal embeddings. It costs a one-time item-selection pipeline requiring embeddings and a panel of models, plus acceptance that rankings are preserved only statistically rather than exactly. It can fail when evaluating models far outside the selection panel, when capability categories drift as new benchmarks and models appear, and when fine-grained capability diagnosis (not just ranking) is needed, since pruning discards per-item signal.", "rationale": "A small team evaluating VLMs on one GPU or via paid APIs can cut benchmark compute roughly 20x by adopting the released pruned suite, but should first validate that the preserved rankings hold for the specific models and capabilities they care about.", "technique": "Hierarchical benchmark compression for VLM evaluation", "family": "avaliacao_de_agente", "recommendation": "testar", "published": "2026-09-29", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.37515", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "PRIMEBench replaces running full vision-language benchmark suites by hierarchically cleaning non-visual or saturated items, selecting one representative benchmark per capability category, and pruning items via a vision-aware variance score computed from multimodal embeddings. It costs a one-time item-selection pipeline requiring embeddings and a panel of models, plus acceptance that rankings are preserved only statistically rather than exactly. It can fail when evaluating models far outside the selection panel, when capability categories drift as new benchmarks and models appear, and when fine-grained capability diagnosis (not just ranking) is needed, since pruning discards per-item signal.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.37515", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "The released suite retains 5% of items (and removes over 97% in the most aggressive configuration) while preserving model rankings, with the highest mean fidelity at 5% retention on models held out from item selection.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.37515", "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": "" }