InsightEmb: Learning Action-Intent Embeddings for Agentic Insight Retrieval

2026-08-05

AI Summary

Existing agent insight retrieval methods rely solely on semantic similarity, which often fails to align with the current decision-making bottleneck. This work proposes a contrastive embedding framework oriented toward decision progress, which learns a transferable geometric structure for matching states to insights from publicly available mathematical reasoning data. By jointly aligning concrete situations with abstract heuristic rules and clustering reasoning trajectories exhibiting similar progress patterns, the method constructs— for the first time—a decision-progress-oriented retrieval embedding space. Notably, it enables efficient cross-domain retrieval without task-specific training. Experiments demonstrate that the approach significantly outperforms existing reasoning embedding models on both dynamic agent tasks and static skill retrieval benchmarks, validating the generalization capability of the learned matching structure.

Abstract

Self-improving agents accumulate reusable insights from prior trajectories, making retrieval increasingly important for turning accumulated experience into actionable guidance. At each decision step, retrieving the right insight can help the agent progress toward its goal, a setting we refer to as agentic insight retrieval. However, existing retrieval methods primarily model semantic similarity, while overlooking whether a retrieved insight resolves the agent's current decision bottleneck. We propose InsightEmb, a contrastive embedding framework that learns transferable progress-oriented retrieval geometry using only mathematical reasoning data. InsightEmb jointly learns to align concrete situations with abstract heuristic rules and to cluster reasoning trajectories with similar progress structures. We evaluate InsightEmb on dynamic agent tasks and a static skill-retrieval benchmark. Without any environment-specific training, InsightEmb improves over all these evaluations, surpassing the performance of existing reasoning embedding models. These results suggest that the geometry of state-insight matching can transfer across domains, enabling effective training from publicly available reasoning data without expensive environment-specific supervision.

Problem

• agentic insight retrieval
• decision bottleneck
• reusable insights
• reasoning trajectories
• state-insight matching

Innovation

• agentic insight retrieval
• contrastive embedding
• progress-oriented retrieval
• transferable reasoning geometry
• heuristic alignment