SD-Search: On-Policy Hindsight Self-Distillation for Search-Augmented Reasoning
The paper titled 'SD-Search: On-Policy Hindsight Self-Distillation for Search-Augmented Reasoning' introduces a novel approach for improving search-augmented reasoning agents. This method utilizes on-policy hindsight self-distillation to provide step-level supervision without the need for external models or annotations. The authors propose a dual-role model that enhances decision-making by leveraging insights from previous rollouts.
- ▪The proposed method, SD-Search, derives supervision from the policy itself.
- ▪It eliminates the need for external teacher models or additional annotations.
- ▪The model functions as both a student and a teacher, conditioning on different aspects of the data.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.18299 (cs) [Submitted on 18 May 2026] Title:SD-Search: On-Policy Hindsight Self-Distillation for Search-Augmented Reasoning Authors:Yufei Ma, Zihan Liang, Ben Chen, Zhipeng Qian, Huangyu Dai, Lingtao Mao, Xuxin Zhang, Chenyi Lei, Wenwu Ou View a PDF of the paper titled SD-Search: On-Policy Hindsight Self-Distillation for Search-Augmented Reasoning, by Yufei Ma and 8 other authors View PDF HTML (experimental) Abstract:Search-augmented reasoning agents interleave internal reasoning with calls to an external retriever, and their performance relies on the quality of each issued query.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
