Long-context QA stops re-searching every query: a trained builder agent lifts accuracy from 0.42 to 0.75

Preparation for long-document question answering moves ahead of the query: a builder agent first reorganizes raw documents into a multi-level filesystem, and an answering agent then looks for evidence inside that workspace instead of re-searching the raw corpus for each query. Retrieval-augmented generation searches raw documents only after a query arrives, while precomputed structures such as knowledge graphs commit to fixed abstractions like named-entity extraction and help less when queries do not match the index. On BrowseComp-Plus with a frozen Qwen3.5-4B answerer, accuracy is 0.75 against 0.42 for embedding retrieval and 0.44 for reranking; with Gemini 3.1 Flash Lite it is 0.77 against 0.30 and 0.42.

The builder's organizing actions are themselves learned: it creates, merges and deletes synthetic files so that downstream answering becomes easier. The same framework without reinforcement-learning fine-tuning reaches only 0.28 accuracy, indicating the base model does not already organize memory for another agent. Construction is cast as a finite-horizon Markov decision process optimized with GRPO, rewarded by answerability changes probed every eight turns plus terminal answer correctness, and the builder never sees evaluation questions during construction.

One constructed filesystem serves many queries over the same document set, so construction cost is spread across them. Per-query retrieval pays the cross-document reasoning cost repeatedly, whereas construction cost is incurred once during curation. Across 443 synthetic queries over 50 document sets, the same filesystems reach 0.64 accuracy against 0.24 for the base builder; at a hypothetical reuse of 20 queries per document set the reported construction cost projects to roughly 18K tokens per query, and normalized at ten questions per corpus the method uses 33.8% fewer recorded tokens per expected correct answer than the base builder.

Outside the training distribution, the builder still produces usable workspaces. Training uses only the 680 BrowseComp-Plus tasks, and the BRIGHT domains are unseen. Zero-shot on 150 BRIGHT questions across AOPS, biology and Stack Overflow, average pass@1 is 0.52 and pass@3 is 0.67, above the listed retrieval baselines; the trained builder averages 53.4 active synthetic files, 45.4 clusters and 8.0 merges, against 4.8, 4.2 and 0.6 for the base builder.

Evaluation is limited to question answering over static corpora: BrowseComp-Plus uses 680 training tasks and 50 held-out tasks, and BRIGHT supplies 50 questions in each of three domains for zero-shot transfer. The builder sees only documents during construction, never evaluation questions, and evaluation questions are not used to select the reported checkpoint. The builder operates over a large file context, and the source suggests pairing it with clustering or retrieval to extend the framework to larger domains.

The source limits evaluation to question answering over static corpora and leaves maintaining anticipatory memories as documents change an open problem. No finite set of synthesized artifacts can anticipate every query, and compression may discard details needed for needle-in-a-haystack tasks; the workspace keeps reference links to raw documents for fallback, but deciding when to bypass the synthetic layer remains a challenge. The builder's large file context may limit scalability.