IBM Research India replaces flat RAG's bag-of-chunks with a harness framework that synthesizes a corpus into a hierarchical, typed, navigable wiki — document overviews, cross-document topic pages, and page-level source records — and trains WikiResearcher-9B with on-policy RL (rewarding answer correctness, retrieval quality and trajectory efficiency) to navigate it through multi-turn tool use, progressively refining retrieval when the first lookup misses.

Evaluated on ViDoRe-V3 (8 domains), FinanceBench and memory benchmarks. A retrieval-layer sibling to WikiSkill's persistent-knowledge harness layer.

Paper

agentsagent-harnessretrievalrl

Related