Determine whether procedural fine-tuning framing has a reproducible advantage

Determine whether naming or appending the per-action procedural reminder in the system prompt produces a reproducible training advantage over the corresponding unnamed or absent framing for fine-tuned Qwen2.5-3B-Instruct agents evaluated in NAQD-Env.

Background

The three enriched-curriculum arms use identical user views and assistant targets but differ in system-prompt framing: named procedure, unnamed checklist, or no appended procedure. The observed ordering varies by metric and evaluation split, while each framing has only one training run. Consequently, the study does not resolve whether any framing yields a reproducible advantage across training runs.

References

The appropriate conclusion is that the present study does not establish a reproducible advantage for naming or appending the procedure.

— Learning When to Withdraw: Curriculum Coverage and Rule-Specific Failures in Fine-Tuned Language Agents  (2610.21045 - Abouzahra, 15 Oct 2026) in Section 5.3, “Framing Differences Remain Inconclusive”