Fine-tuning helps small agents stop steps but not on all rules
One checkpoint incorrectly stopped 26 of 34 actions that had other support.
Researchers fine-tuned Qwen2.5-3B-Instruct to stop actions when evidence is different. All of them were less good than the 7B and 8B baselines on prohibited combinations. The authors write that aggregate accuracy does not show the failures of one rule.
Sources
Posted
