Prospective Protocol for Vision–Language Model Score Decomposition and Scorer-Specific Recertification
Year: 2026
Type: article
Abstract: Protocol for an exploratory extension of answer and end-token score decomposition to Qwen3-VL-2B-Instruct and InternVL3.5-4B-Instruct, using pinned open-weight snapshots on one RTX 5080. The specification fixes 1,024 reused diagnostic images, 256 separate development images for numerical checks, BF16 and language-layer NF4, A/B and yes/no prompts, same-forward full/body scores, greedy token records, paired statistical tests, multiplicity correction, and executable code hashes. The extension is exploratory and uses previously studied pilot models and previously exposed images. The protocol is deposited before extension inference. The earlier protocol and its hypotheses remain available unchanged at 10.5281/zenodo.24294124. Earlier code and derived data are available at 10.5281/zenodo.24294502. PROTOCOL_SUPPLEMENT.txt gives the specification. The archive contains machine-readable protocol, sample identifiers, source code, and synthetic validation. Protocol and original identifiers are licensed under CC BY 4.0; original executable code under MIT. Images and pretrained weights are obtained from their official providers.
Authors: Yoo, Nam-Hyun, Park, Hyojin, Yang, Jinhong
Institutions: Kyungnam University, Inje University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7233834734
