Weiguang Zhang, Linhua Dong, Bin Dong, Shanshan Jiang, Maizhen Ning, Xiaoqiang Kang, Kun Zhang, Hao Xu 等 12 位作者
原文摘要
Hanjian manuscripts written on Han Dynasty bamboo and wooden slips are invaluable heritage of ancient Chinese writing but suffer from damaged ink and incomplete strokes. Existing image-to-image restoration methods may alter observed strokes and offer limited control over stroke geometry. Inspired by experts' manual redrawing practices, we address these issues by introducing \textbf{ReDraw}, a two-stage vision-language model(VLM) that reformulates character restoration as Scalable Vector Graphics(SVG) code generation and editing. Stage one targets faithful vectorization of observed strokes geometry or conservative tracing of mildly degraded strokes. Guided by skeleton chain-of-thought reasoning and reinforcement learning with structure-aware rewards, stage two predicts skeletons before contours to redraw severely incomplete strokes by appending paths while keeping existing SVG paths unchanged. To support training, we develop a multi-layer data synthesis pipeline that disentangles ink, material, and degradation to generate 1.2M samples, and derive up to 6,400 adaptation pairs from 320 annotated real glyphs. Experiments demonstrate synthetic restoration performance and adaptability to real Hanjian glyphs. Code and demo will be available upon acceptance.
On-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. We identify evidence bypass: a student retains an originally correct answer after the document changes to support a different one. On QASPER with a Qwen3-4B student, standard OPD raises original-document accuracy by 3.4 percentage points, lowers counterfactual accuracy by 5.3 points, and increases bypass by 15.7 points; the same directional pattern appears on PeerQA. Controlled perturbations localize the shift to answer-critical evidence, and bypass grows with the student's pre-distillation closed-book preference for the original answer. A factorized study of teacher targets, privileged information, divergence, and supervision density finds that explicitly supplying the teacher with annotated answer-critical spans yields the largest joint improvement in evidence response and claim support. These findings motivate Evidence-Paired OPD (EP-OPD), which combines on-policy supervision on matched documents supporting different answers to the same question with an answer-level margin that favors each document's supported answer over its paired alternative. Across both datasets and two student sizes, EP-OPD recovers 54-66% of standard OPD's counterfactual-accuracy loss while preserving 80-85% of its original-document accuracy gain. Our results establish evidence responsiveness as a distinct training and evaluation target alongside answer accuracy. Our code is publicly available at: https://github.com/RuoranXu/EP-OPD.