Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

SD-Search: On-Policy Hindsight Self-Distillation for Search-Augmented Reasoning

About

Search-augmented reasoning agents interleave internal reasoning with calls to an external retriever, and their performance relies on the quality of each issued query. However, under outcome-reward reinforcement learning, every search decision in a rollout shares the same trajectory-level reward, leaving individual queries without step-specific credit. Recent process-supervision approaches address this gap by drawing step-level signals from outside the policy, relying either on a much larger teacher model, or on sub-question annotations produced by a stronger external system. In contrast, we propose SD-Search, which derives step-level supervision from the policy itself through on-policy hindsight self-distillation, requiring neither an external teacher nor additional annotations. In SD-Search, a single model plays two roles that differ only in conditioning: a student that sees only the context available at inference time, and a teacher that additionally conditions on a compact hindsight block summarizing the search queries and final outcomes of a group of rollouts sampled from the same question. Since the teacher knows how each rollout unfolded and which ones succeeded, its query distribution implicitly marks which decisions were worth making, and the student is trained to recover this behavior by minimizing the token-level Jensen--Shannon divergence to the teacher at search-query positions. This layers a dense, step-level signal on top of GRPO's coarse trajectory reward. Crucially, this signal is produced by the policy itself within the standard RL training loop, without external model inference, auxiliary annotation pipeline, or additional training stage.

Yufei Ma, Zihan Liang, Ben Chen, Zhipeng Qian, Huangyu Dai, Lingtao Mao, Xuxin Zhang, Chenyi Lei, Wenwu Ou• 2026

Related benchmarks

TaskDatasetResultRank
Single-hop Question AnsweringPopQA
EM49.4
186
Single-hop Question AnsweringTriviaQA
EM66.8
133
Question AnsweringNQ (test)
EM Accuracy47
133
Question AnsweringPopQA (test)
Accuracy46.7
111
Question AnsweringTriviaQA (test)
EM62.4
80
Question AnsweringMuSiQue (test)
EM18.8
76
Multi-hop Question AnsweringHotpotQA
Exact Match (EM)47.1
66
Single-hop Question AnsweringNQ
Exact Match (EM)50
60
Multi-hop Question AnsweringBamboogle
Exact Match (EM)54.4
55
Multi-hop Question AnsweringMuSiQue
Exact Match (EM)22.2
51
Showing 10 of 16 rows

Other info

Follow for update