📄 AWARE: Agentic Knowledge Warehousing for Contextual Intelligence

Summary

LLMs often fail on questions that need specialized or recent knowledge, and current fixes involve trade-offs. Classic RAG retrieves once and cannot follow chains of dependent facts. Tool-integrated reasoning agents keep all gathered evidence inside their context window, which limits how widely they can search. The authors describe real-world sources such as web pages, PDFs and local files as noisy and low in useful content per token.

They propose AWARE, which treats retrieval as agentic knowledge construction. Offline, each document receives a short header describing its scope and structure, which allows cheap screening. Online, the agent plans a sequence of information intents and splits each into atomic sub-queries. It explores many threads in parallel and refines wherever dependencies remain unresolved. Evidence is then extracted and merged in map–reduce fashion outside the main reasoning loop. The result is a compact, step-ordered context that any LLM can use to answer.

AWARE outperforms all baselines on GAIA, WebWalker and BrowseComp-Plus. It separates costly reasoning from bulk text processing, so standalone LLMs can handle long multi-hop search within fixed context budgets.

Key points

• The evaluation benchmarks are GAIA, WebWalker and BrowseComp-Plus, and the paper reports gains over every baseline. Exact scores are not in the excerpt provided.
• The pipeline has four modules:
  • offline document structuring with headers;
  • multi-threaded exploration with targeted refinement;
  • map–reduce evidence integration;
  • task-aware contextualization into LLM-ready input.
• Header-level screening discards close to 90% of irrelevant retrieved pages before detailed processing.
• Raising exploration depth from 1 to 5 improves results, after which performance fluctuates as evidence saturates. Deeper search also reduces the number of intents needed.
• In a GAIA-style case study (answer: "Mice"), one intent used 4 atomic queries, retrieved 36 pages and browsed 13. Reasoning used about 8.9K tokens versus about 227.2K for bulk processing. This shows that a strong central model can be paired with lightweight auxiliary models.