Tencent's InsightEmb: Training Agentic Experience Retrieval Using Only Math Data

_reachsumit · x · 2026-08-07

Self-improving agents accumulate reusable insights from prior trajectories, but existing retrieval methods rely primarily on semantic similarity, overlooking whether a retrieved insight actually resolves the agent's current decision bottleneck.

To address this, researchers at Tencent proposed InsightEmb, a contrastive embedding framework. Trained exclusively on mathematical reasoning data, InsightEmb learns a transferable progress-oriented retrieval geometry by aligning concrete situations with abstract heuristic rules and clustering reasoning trajectories with similar progress structures.

Without any environment-specific training, InsightEmb surpasses existing reasoning embedding models on dynamic agent tasks and static skill-retrieval benchmarks, demonstrating that effective training can be achieved using publicly available reasoning data without expensive environment-specific supervision.

Original post →

More from coding & agent

coding & agent channel →