benchmarks evals
The Attack Changed the Representation. The Caption Stayed Right.
A controlled vision-language experiment separates corrupted internal signals from the answer a model freely generates.
Summary
A controlled vision-language experiment separates corrupted internal signals from the answer a model freely generates.
On Qwen2.5-VL-7B-Instruct and two hundred held-out COCO images, a targeted perturbation could lower training loss for a chosen caption while free generation still produced the correct description. The study traces the disagreement through the decoder and reports that susceptible and resistant outcomes diverge there despite similar visual-encoder corruption. It also flags circularity in one probe estimate. The result concerns a particular model and attack setup, and argues for measuring generated behavior alongside internal proxies.
Why it matters
A controlled vision-language experiment separates corrupted internal signals from the answer a model freely generates.
Limits and context
No additional limitation was separately recorded.
Key claims
A controlled vision-language experiment separates corrupted internal signals from the answer a model freely generates.
Evidence: source-2026-10-04-015
Sources
- arXiv preprint 2610.03445, original v1arXiv / submitting research authors · primary research
Corrections
: Publication note: This October 4 edition was reconstructed on October 5 after the workstation was offline. It uses original October 2 submissions from the standard Sunday-evening arXiv release, rather than claiming those papers were available for the missed 9 AM run. The date and 9 AM publication slot preserve archive chronology; October 5 developments are excluded. The September 27 Invention Desk is carried over because its replacement was generated October 5.