Ulms ERT training makes vision nets far easier for LLMs to edit in weight space
Timo Ropinski · hf · 2026-11-20
A paper by Timo Ropinski and Sebastian Kranz (Ulm University) studies how training vision models makes them easier for LLMs to edit via weight-space programming.
Key points:
- A benchmark where five LLMs edit CNNs and ViTs on relabeling and error-correction tasks across Waterbirds, CIFAR-100, PACS, and Entity-30
- Over 2,600 edit processes reported, each scored on held-out images for success, side effects, and accuracy loss
- Compares standard training (with label smoothing), gradient-based editable training, and the novel ERT (Edit-Ready Training), which rehearses the direct closed-form weight edits LLMs perform
- Edit-ready networks substantially raise LLM edit success; ERT achieves the highest mean score for GPT-6 Astra and Claude Opus 5.5 across all baselines
The LLM testing framework, tasks, editing trajectories, and scores will be released publicly.
More from Models
- Nathan Lambert tries Gemini Neon in Antigravity: feels in distribution of Claude Code and Codex — natolambert · 2026-11-20
- Polymarket odds say Gemini will top Artificial Analysis leaderboard by end of November — LeviTurk · 2026-11-20
- Vercel: decision models hit 26% of gateway requests; Anthropic wins new teams 2:1 — airesearch12 · 2026-11-20
- Not all open weights are equal: comparing release packages across labs — WarmMycologist9549 · 2026-11-20
- Leaked Gemini 4 lineage: neon scales hydrogen's base, beats carbon at lower cost — haider1 · 2026-11-20
- Hobbyist Converts Qwen3.5-2B Dense to MoE, With Full Benchmarks — Bayliner1980 · 2026-11-20