IGSD Uses Environment-Verified Hindsight Self-Distillation to Train Search Agents

_reachsumit · x · 2026-09-29

A new paper proposes IGSD (Information-Gain-Gated Self-Distillation), a teacher-free on-policy self-distillation method for training search agents.
Key idea: a policy conditioned on privileged hindsight guides its own unprivileged rollouts, but prior approaches distill such preferences directly or filter them with model-internal scores—neither verifies the query's actual retrieval outcome.
IGSD treats each query token as a micro-action: it completes both the teacher's proposed query and the student's sampled query, executes both from the same failed state with the same retriever, and uses shared counterfactual controls to isolate the executed paired information gain as a relative utility signal for the retrieved documents.
This contrast serves as a positive-only soft weight for candidate-pair distillation, leaving the GRPO objective unchanged; verification happens only during training.
Results: across 7 single-hop and multi-hop QA benchmarks, IGSD reaches macro-average exact-match accuracies of 42.8% (3B) and 47.0% (7B).

More from coding & agent