Models over-edit code written by other models; CROCODIL training framework fixes it
omarsar0 · x · 2026-09-06
Researchers found a weird behavior: coding models over-edit code written by other models. Different training data produces different stylistic preferences, so models make more—often excessive—edits on foreign code than on their own, and multi-model commit history now changes how each model edits.
CROCODIL is a post-training framework that reduces this: a similarity reward penalizes large changes, an execution reward scores build/test success, and the two are multiplied rather than added, preventing the policy from shrinking edits by simply failing the task.
More from coding & agent
- Grok launches Imagine Video 1.5 agent powered by Image 2.0 for multi-shot storytelling — belce_dogru · 2026-09-06
- 82% of enterprises run AI pilots, but over half never deploy: production, not models, is the bottleneck — worldofjaved · 2026-09-06
- Pamela Fox: I like LLM-generated code, but give READMEs a human pass — DanWahlin · 2026-09-06
- Taking a Break Is Hard When Your Astra Agent Keeps Running /goal — AIandDesign · 2026-09-06
- Dev Powers Filesystem Simulator VSH With Monty to Track Side Effects Before Execution — samuelcolvin · 2026-09-06
- Cheating Agents Answer Faster: A Missed Red Flag for Reward Hacking — zainhas · 2026-09-06