Taming Data Chaos: Agentic Knowledge Warehousing for Contextual Intelligence

ICLR 2026 Conference Submission257 Authors

Keywords: complex knowledge discovery, contextual intelligence, large language models

TL;DR: We present AWARE (Agentic Knowledge Warehouse), a retrieval paradigm that converts massive unstructured data into minimal, task-specific, LLM-ready context for complex knowledge discovery tasks such as BrowseComp.

Abstract:

Information seeking can be viewed as bridging the knowledge gap between a query and its answer. While large language models (LLMs) perform strongly across diverse tasks, their capacity to fill this gap is bounded by pretraining data and deteriorates on queries requiring specialized or up-to-date knowledge. A common solution is to augment LLMs with external knowledge, either by injecting retrieved evidence into the context or by interleaving retrieval with reasoning. The former restricts exploration of layered dependencies, whereas the latter is constrained by context length, limiting both efficiency and scalability. Yet complex tasks often involve intricate dependencies and may require processing large volumes of raw text, under which both strategies become inadequate.

To tackle this bottleneck, we present Agentic Knowledge Warehouse (AWARE), a retrieval paradigm that transforms vast unstructured data into minimal, task-specific knowledge consumable by LLMs. Rather than simply returning raw information, AWARE curates knowledge through an agentic process that plans, explores, and synthesizes evidence into coherent context. Specifically, it organizes raw corpora with document-level gist memory for global coverage, applies diffusion-based exploration with vertical exploitation to recover layered dependencies, and employs map–reduce inspired synthesis to integrate large-scale evidence into a compact, LLM-ready context. This design enables both in-depth exploration and scalable integration, reconstructing the knowledge space needed to address task-specific knowledge gaps. Experiments on GAIA, WebWalker, and BrowseComp show that AWARE outperforms baselines, validating its effectiveness and generality.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 257