{ "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": "2610.25964", "title": "Token-Set Choice Confounds POPE: A Systematic Audit of Yes/No Extraction in VLM Hallucination Evaluation", "summary": "It replaces hard-coded vocabulary-ID yes/no comparisons in POPE-style logit readouts with a tokenizer-derived token set covering the surface forms the model actually emits (e.g., ' Yes' vs 'Yes'). The cost is a one-time audit of generated tokens per model/tokenizer plus a slightly larger lookup set; no additional memory, latency, or compute in production. It can fail if adopted naively: tokenizer readouts must still be re-derived per model family, and cross-paper score comparisons remain invalid until readout and quantization precision are matched, since measured baseline shifts exceeded most published mitigation gains.", "rationale": "The fix is a zero-cost evaluation-protocol correction that prevents misattributing benchmark gains to methods, which is essential for any team making adopt/not-adopt decisions from VLM hallucination papers.", "technique": "Token-set-aware yes/no logit readout audit", "family": "avaliacao_de_agente", "recommendation": "adotar", "published": "2026-10-19", "as_of": "2026-10-20", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.25964", "independent_implementations": 4, "total_implementations": 4, "stars": 2, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "It replaces hard-coded vocabulary-ID yes/no comparisons in POPE-style logit readouts with a tokenizer-derived token set covering the surface forms the model actually emits (e.g., ' Yes' vs 'Yes'). The cost is a one-time audit of generated tokens per model/tokenizer plus a slightly larger lookup set; no additional memory, latency, or compute in production. It can fail if adopted naively: tokenizer readouts must still be re-derived per model family, and cross-paper score comparisons remain invalid until readout and quantization precision are matched, since measured baseline shifts exceeded most published mitigation gains.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.25964", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Switching from a two-token readout (IDs 3582/1217) to an eight-token readout covering space-prefixed surface forms raises LLaVA-1.5-7B adversarial POPE F1 by 6.13 points (to 0.8221); porting the LLaVA-1.5 IDs verbatim drops F1 to 0.00 on disjoint-vocabulary models, while a tokenizer-derived readout holds between 0.80 and 0.92.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.25964", "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": "" }