Harvard Proposes Agentic Data Cracking, Cuts Unstructured QA Cost 53% on FanOutQA
Harvard · hf · 2026-09-03
Harvard researchers propose agentic data cracking, a method that lets LLM agents structure unstructured data (PDFs, reports, earnings calls) as a byproduct of reasoning itself, sharply cutting cost.
Problem: agents repeatedly reopen large documents to recover scattered evidence—up to a million tokens per question; over a pre-structured store, FanOutQA reasoning is 28X cheaper, with the gap growing to orders of magnitude as questions fan out.
Method: structuring is adaptive (observed queries decide when and what to extract) and speculative (serving future related questions). When an agent opens a document, a cracking sub-agent forks from the already-loaded context at marginal cost and extracts grounded structure for reuse.
Results: with just one related question added per FanOutQA test question, cracking cuts cost by 53% while preserving accuracy. The authors frame it as a first step toward next-generation data infrastructure for agentic reasoning.
More from coding & agent
- Stanford launches CS329Z, a new fall course on building AI agents from scratch — jyangballin · 2026-09-03
- Naveen Rao Says LLM Compiler Is Real: Turning Python Into Optimized Assembly, Guarantees Next — NaveenGRao · 2026-09-03
- A 9-step roadmap from general-purpose agents to composable specialist agents — MaryamMiradi · 2026-09-03
- Run Claude Code on Google Cloud in 60 seconds with gcloud auth and 3 env vars — fhinkel · 2026-09-03
- Full prompt released for single-file realistic three.js waterfall with WebGPU and TSL — majidmanzarpour · 2026-09-03
- Same-prompt test: fable 5.1 vs sol 5.6 generating a realistic three.js waterfall — majidmanzarpour · 2026-09-03