Event ·Multi-source

HiLo-RD:用跨预算自蒸馏提升掩码扩散语言模型小窗口推理

HiLo-RD(High-to-Low Budget Representation Distillation)针对掩码扩散语言模型(MDLM)在低生成预算下的问题,提出让大窗口教师状态与小窗口裁剪视图配对,并在已提交位置匹配隐藏表示的自蒸馏方法。论文指出模型存在预算导致的过早提交现象:正确解尚未完成中间推理时,答案片段已在窗口边界附近被固定,进而误导后续预测。在 GSM8K 上,该方法将低预算准确率提升 9.4 个百分点,平均可读长度下降约 25%。改进还迁移到代码生成任务,且无需代码域训练。

Current read

arXiv上新出的自蒸馏方法,专门修MDLM小窗口推理时过早下结论的毛病,GSM8K涨了9.4个点还更省token。

3 sources38° AI heatLast updated 2026/10/27 09:34:08

Evidence

3 sources
Primary sourcearXiv cs.LG
View source
Related source 2Noam Brown (OpenAI 推理)
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.