Coding agent repairs vision-language model weights without gradients, arXiv study reports
A weight-space programming study planted faults in two small vision-language models and had a coding agent undo them, with mixed success across fault types.
In plain words
Researchers deliberately damaged the stored weights of two small vision-language models and let a coding agent fix them by editing the weight files directly, with no retraining. The agent often restored the original values and accuracy, but only when the damage left a hidden numerical clue. When that clue was removed, repairs largely failed.
What was announced
The arXiv paper, whose abstract is the available source, describes weight-space programming as a language model editing a network's weights directly without gradients. The authors planted faults in SmolVLM-256M and LFM2-VL-450M and had a coding agent compare each edit against the original weights. Without being told where to look, the agent located rescaled projector rows and noise in the vision encoder, inferred the unknown scale factors, and returned most or all rescaled weights to their exact original values, with accuracy near its original level.
The repairs depended on a trace of how the faults were computed: the original weights use 16-bit precision, so faults computed in 32-bit arithmetic left low bits carried only by the damaged weights. When the authors rounded those bits away, both models still answered test questions as before, but the agent no longer repaired either model unaided; on LFM2-VL accuracy recovered by about 4 points versus 17 with unrounded weights. Restoring the low-bit markers brought the edits back. Swapped rows or columns were not found unaided on either model within the study's budgets, and on LFM2-VL no session found the injected noise. The authors recommend that weight-repair studies report the weights' number format and include a rounded control.
What it does not show
- Only the paper's abstract is available, so methods, budgets and full results cannot be verified from the supplied evidence.
- The agent's success partly relied on a numerical artifact of 32-bit fault computation, which the authors note a real failure need not leave.
- Swapped rows or columns and, on one model, encoder noise were not repaired unaided, so coverage of fault types is partial.
For security and IT teams
- Scope
- Findings cover two specific small models, SmolVLM-256M and LFM2-VL-450M, under the authors' planted-fault test setup, not general production repair outcomes.
Sources
Source-support review completed 20 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
20 Nov 2026 · Updated wording and source-support review.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.