How visual evidence gets encoded: different priors interpolate into different hallucinations
debreuil · x · 2026-10-24
Sharing a structured annotation of a line drawing (a 'church with a side tower' sketch with corners, junctions, and closed boundaries — 29 spans, 28 sites), the author argues evidence is encoded through many stages from rough outline to detail, and different priors (organic vs man-made) interpolate differently, producing different hallucination patterns — 'local window repetition does not imply whole-scene symmetry.' Useful for understanding how vision models hallucinate.
Related event: Experiments explore structured visual evidence parsing of line sketches(2 posts)→
More from Research
- Google Discover's SDF filters stale articles before ranking — stale reports down 54.9% in two years — gaganghotra_ · 2026-10-24
- Google says Gemini 4 Argon found fusion counterexamples and designed 5 lab-verified enzymes in hours — Eric520CC · 2026-10-24
- MIT left 13 AI agents alone on virtual islands for 30 hours — they started to play, and it made them better — alliekmiller · 2026-10-24
- Robert Plenge: How much causal human biology we know determines AI's role in drug discovery — rbhar90 · 2026-10-24
- Bug Hunt Bench: blind-graded leaderboard pits frontier coding models against 105 real planted bugs — PawelHuryn · 2026-10-24
- MIT Team: Only 2 of 286 TB Antigen Candidates Detected on MHC-I — BrianHie · 2026-10-24