Enriched Curriculum Improves Agent Withdrawal Recall in Tool-Using Tasks
October 16, 2026
Fine-tuned Qwen2.5-3B-Instruct agents using an enriched curriculum achieved withdrawal recall of up to 0.86 in the NAQD-Env environment. This significantly outperforms baseline instruction-tuned models, which reached a maximum recall of only 0.06 when tasked with suspending actions due to changing evidence.
HOW THIS AFFECTS YOU
●
builderYou can improve agent reliability in dynamic environments by using curriculum-based SFT focused on action withdrawal.
●
researcherThis highlights the necessity of training agents to handle authority revocation and dependency changes.