Answers are grounded in this paper and cite their sources.
Prospective protocol for experiments on answer scoring and scorer-specific recertification of SmolVLM2-2.2B-Instruct and Qwen3-VL-4B-Instruct on binary VQAv2 using an RTX 5080. The deposited files specify hypotheses, sample identifiers and SHA-256 hashes, pinned model snapshots, answer and end-token definitions, precision settings, statistical tests, calibration error allocations, and executable analysis code. The protocol was publicly deposited before the new main inference. The 256-image development checks and synthetic statistical checks are included. The primary scoring analysis uses 4,000 images disjoint from the earlier 1,024-image diagnostic. These images had previously supplied native predictions for unlabeled calibration, and public annotations had been parsed for eligibility. Scorer-specific calibration reuses 800 development, 2,000 labeled calibration, and 2,800 evaluation images; the protocol records their earlier exposure. Evaluation labels for the 4,000-image scoring analysis are extracted after main inference and policy fitting. Greedy-generation comparisons use the earlier 1,024-image diagnostic subset. PROTOCOL.txt provides the complete specification. The ZIP contains the machine-readable protocol, sample manifest, source-code snapshot, and development-validation receipts. Protocol and original derived identifiers are licensed under CC BY 4.0; original executable code is licensed under MIT. Benchmark images and pretrained weights are obtained separately from their original providers. Prior code and derived data: https://doi.org/10.5281/zenodo.24281585.
No takes yet. Share an insight, caveat, or question.
Yoo et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: