← News

Models & toolsSources: arXiv · 20 Nov 2026

Keel uses a frontier model to edit a small VLM's weights and stop answer loops

The Keel paper describes an agentic weight-editing method in which a frontier multimodal LLM finds and removes looping behavior in small vision-language models without any training.

In plain words

Some small vision-language models get stuck repeating themselves when answering questions about images, using up their token budget before giving an answer. Keel is a method where a large frontier model acts as an editor, locating the cause in the smaller model's weights and changing them directly. No training is involved, and the model keeps its original size and cost.

What was announced

The authors frame Keel as the first agentic weight-editing method for vision-language models on image tasks. Given only a symptom description, a frontier multimodal LLM acting as an editor locates the cause of the looping, edits the weights, and tests the result. A failure census and instruments show the editor the failing images and where the VLM attends, and the harness keeps only edits that clearly reduce the failure without lowering accuracy on questions no editor saw. The repair requires no training, preserves the model's size and per-token cost, and can be applied to already-edited weights.

On Qwen3.5-0.8B and Qwen3.5-2B, across six image benchmarks ranging from visual mathematics to GUI grounding, the edit removes most loops on every benchmark and shortens answers, with edited models significantly more accurate on average than the thinking models. The larger model gains most on GUI grounding, a task absent from the practice questions. The editors traced the loops to the attention outputs of the middle layers, where the smaller VLM stops consulting the image during a loop. Because the repair lives in the weights, it combines with budget forcing, and the authors report the best average accuracy on both models, significantly above every training-free baseline compared on the larger one.

What it does not show

  • The evidence is the paper's abstract only; full experimental details, methodology, and peer review status are not available from this excerpt.
  • The accuracy and loop-reduction claims are the authors' own reported results and have not been independently confirmed.
  • The specific frontier multimodal LLM used as the editor is not named in the abstract.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.