Knowledge Compilation Cuts Agentic Search Query Tokens by 91%, Paper Finds
_reachsumit · x · 2026-11-02
An arXiv paper proposes knowledge compilation as an infrastructure layer for agentic retrieval: LLMs enrich documents at ingestion so agents receive compact, navigable context.
- Key observation: retrieval is going agentic — a harness with just BM25/grep and MiniMax M3 answers 85% of BrowseComp-Plus and 91% of multi-hop FRAMES without compiled or vector indexes
- The real bottleneck is context: undisciplined loops repeat queries, reread documents, and fill the window with unused passages
- The ingestion pipeline generates navigation hierarchies with claim-evidence leaves, entity graphs, timelines, mind maps, ontologies, wikis, and memory
- Counterintuitive finding: under a competent harness, structural artifacts gave no detectable accuracy gain, and a PageIndex-style outline degraded at scale — compilation is a supply layer, not a better retriever
- Measured mechanism: a claim-evidence navigation tree cuts query tokens by 91% when evidence suffices and adds 1.9% when it doesn't (38% average reduction), implying a 2% break-even sufficiency rate vs. a measured 43%
More from coding & agent
- ArtCraft rebuilds Adobe's Creative Cloud in pure Rust: 15 free open-source apps on GitHub — anthara_ai · 2026-11-02
- Measure launches as fully open-source Crashlytics alternative with MCP-accessible debugging agent — Arindam_1729 · 2026-11-02
- Nous-Girl: an open-source StackChan robot powered by Nous Research's Hermes Agent — alexcovo_eth · 2026-11-02
- A scorecard for agent pilots: tasks, interruptions, missed decisions — krishnan · 2026-11-02
- Managing agents is babysitting: every ping arrives equally loud — krishnan · 2026-11-02
- A task you can't verify yet isn't done — agents need a 'deferred' task state — Key-Register-1969 · 2026-11-02