What Does Privileged Information Add to On-Policy Self-Distillation?
- 类型：arxiv
- 标识：2609.20612
- 链接：https://arxiv.org/abs/2609.20612
- 主分类：llm-infra
- 形态：method
- 被引：3
- 被引来源：Semantic Scholar
- S2被引：3
- OpenAlex被引：0
- 影响力被引：0
- TLDR：AMPLE-Math, a reusable suite of 5,319 mathematical problems with six reasoning views that share the same answer, is constructed and AMPLE-Math, a reusable suite of 5,319 mathematical problems with six reasoning views that share the same answer, is compared, suggesting that OPSD can improve access to existing reasoning capabilities through parameters shared by direct-response and thinking-enabled inference.
- OpenAlex ID：W7213581439
- OpenAlex DOI：10.48550/arxiv.2609.20612
- DOI：10.48550/arxiv.2609.20612
- DOI来源：OpenAlex
- 开放获取：green
- 开放获取链接：https://doi.org/10.48550/arxiv.2609.20612
- OpenAlex更新：2026-09-20
- 待LLM分类：否
- 成熟度：research
- 场景：self-distillation、language-model-training、teacher-student
- 标题中文：特权信息为 On-Policy 自蒸馏带来了什么？
- TLDR中文：构建了 AMPLE-Math，一个包含 5,319 道数学问题、每题对应六种共享同一答案的推理视角的可复用套件，并通过对比表明 OPSD 能借助直答推理与带思维链推理所共享的参数，提升对已有推理能力的调用效率。
- 来源文件：
- /inbox/tom/_candidates/2026-09-18-agent-rag-longcontext-candidates.json
- [S2 enrich]
- [OpenAlex backfill]