VLTest: Finding Joint Image-Text Failures in Vision-Language Models

No. 223 of 245 · arXiv · Bin Duan, Jintao Lin, Guowei Yang · published 23 Oct 2026, 04:00 UTC · 64 min read

AI & machine learning
Paper

2 judges. Scored by two of the three judges. The third reads only the day’s strongest candidates. How far stories are checked

VLTest is a black-box testing approach for vision-language models that addresses two problems: an output change is not necessarily a failure, since mutations can alter the correct answer, and joint failures require matched single-modality controls to be attributed to the combination.

Read the original at arxiv.org. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Score

0.56total under rubric v2.0

Significance0.50
Reader interest0.47
Audience breadth0.45
Practical relevance0.56
Substantiated0.89
Technical depth0.90
New information0.76
Counts against: Promotional0.04
Counts against: Overclaim0.26

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

Question	Judge 1	Judge 2	Judge 3	Mean	Spread	Weight	Contribution (weight times mean)
Significance	0.59	0.40	—	0.50	0.20	+0.25	+0.124
Reader interest	0.49	0.45	—	0.47	0.04	+0.20	+0.094
Audience breadth	0.55	0.35	—	0.45	0.20	+0.10	+0.045
Practical relevance	0.67	0.45	—	0.56	0.21	+0.15	+0.084
Substantiated	0.99	0.80	—	0.89	0.19	+0.10	+0.089
Technical depth	0.99	0.80	—	0.90	0.19	+0.10	+0.090
New information	0.92	0.60	—	0.76	0.32	+0.10	+0.076
Promotional	0.04	0.05	—	0.04	0.01	−0.15	−0.007
Overclaim	0.22	0.30	—	0.26	0.08	−0.15	−0.039
Sum of contributions							+0.556
Claim-check penalty							0.000
Total, as published							0.556

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested. Judge 3 did not score this story; its mean uses the other two.

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN —

Recorded for context only, checked 24 Oct 2026, 08:54 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition and listed in its JSON. This one was scored under rubric v2.0; How ranking works describes the current rubric, v2.2.