Fine-tuned language agents learn when to pause actions effectively
Learning When to Withdraw: Curriculum Coverage and Rule-Specific Failures in Fine-Tuned Language Agents
Artificial Intelligence
Summary
When language agents use tools to complete tasks, they sometimes need to stop actions if new information shows those actions are wrong. This paper studies how well fine-tuned models learn to stop only the wrong parts of their plans without undoing everything. The authors test this in a controlled environment where rules and dependencies are clear, showing fine-tuned models withdraw actions more accurately than baseline models. However, some challenges remain, like handling new rule combinations and preserving independent actions.
What this means in practice
- •For chatbot developers: Improve agents’ ability to selectively pause actions when task requirements change, reducing unnecessary interruptions.
- •For workflow automation teams: Use fine-tuning techniques to help automated systems better manage dependencies and rule violations during task execution.
Authors
Mohamed Abouzahra
Abstract
Tool-using language agents must revise prepared actions when evidence changes or authority is revoked. Selective withdrawal means suspending affected actions, preserving supported work, and resuming only after sufficient repair. We study supervised fine-tuning in NAQD-Env, a synthetic environment with explicit dependencies and a deterministic reference policy. Four LoRA checkpoints of Qwen2.5-3B-Instruct compare an original curriculum with an enriched one, the latter under three system-prompt framings with identical views and targets. Three instruction-tuned baselines and all four checkpoints are evaluated under three inference prompts on fresh validation cases, held-out dependency families, and held-out structural combinations. Under the naqd prompt, SFT checkpoints reach decision agreement of 0.77-0.92 and withdrawal recall of 0.46-0.86, versus 0.45-0.63 and at most 0.06 for the baselines. Rule-level diagnostics qualify these gains. The original-curriculum checkpoint wrongly withdraws 26 of 34 actions with surviving alternative support; enriched checkpoints withdraw 0-1. Conversely, all SFT checkpoints narrow prohibited combinations less often than the 7B and 8B baselines, a case absent from both curricula. Delegation results show more withdrawal but poorer preservation of independently authorized actions, and all four checkpoints omit valid event reports. Framing contrasts vary across metrics and splits and do not establish a reproducible advantage or equivalence. One training run per arm, jointly changed curriculum properties, prior evaluation exposure, and baseline-SFT quantization differences limit causal attribution. The contribution is a method for detecting rule-specific regressions that aggregate accuracy can conceal, not evidence that fine-tuning alone ensures agent safety.