Evidence bypass identified in on-policy distillation for scientific QA
Original title: Better Answers, Weaker Evidence Control: Evidence Bypass in On-Policy Distillation
On-policy distillation improves answer accuracy but weakens evidence responsiveness, with a new EP-OPD method recovering much of the lost evidence control.
On-policy distillation (OPD) raises original-document accuracy by 3.4 points but lowers counterfactual accuracy by 5.3 points on QASPER and PeerQA. This creates evidence bypass where the student retains the original answer after document changes. Controlled experiments localize the shift to answer-critical evidence. A factorized study shows supplying annotated spans to the teacher yields best results. EP-OPD combines on-policy supervision with answer-level margin to recover 54-66% of counterfactual loss while preserving 80-85% of accuracy gain.
- Standard OPD increases bypass by 15.7 points on Qwen3-4B
- EP-OPD recovers 54-66% of counterfactual accuracy loss
- Explicit teacher spans give largest joint improvement
- Bypass grows with student's closed-book preference
- Results hold on PeerQA with two student sizes
From the original article
1 min read at arxiv.orgOn-policy distillation (OPD) can improve answer accuracy while weakening responsiveness to supporting evidence. We investigate this separation in scientific-document question answering through matched document interventions that change the supported answer while preserving the question and unrelated context.
By Ruoran Xu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.