The Point: RISE achieves similar accuracy to unbounded shell interaction within a limited interaction space, but reduces request costs to about one quarter and scales significantly better to large corpora.
Researchers propose a retrieval system designed for agent-based search rather than direct shell commands across the entire corpus, constructing a limited searchable interaction field. The RISE method uses BM25 ranking and document-specific indexing to dramatically reduce latency and costs.
Current agent-based search systems still adopt the paradigm of classical information retrieval: a retriever ranks the document collection and the agent reads a small amount of returned documents. However, newer approaches for direct corpus interaction (DCI) demonstrate that agents can instead interact directly with the raw collection through shell tools like grep and file readers. This model has a critical scaling problem: every broad shell command triggers a scan across the entire corpus, and latencies grow rapidly with corpus size.
The researchers argue that retrieval for agent-based search should not only select documents that fit into the LLM context window, but must construct a limited explorable interaction space—a subset of the corpus with associated exploration tools. Under this concept, the boundary of this space is drawn by retrieval, and the objects within it are optimized for navigation. The proof-of-concept system RISE uses BM25 to construct the interaction space; the documents it contains are prepared during indexing for shell-like navigation.
On the BrowseComp-Plus benchmark, RISE with gpt-4-mini achieves 78 percent accuracy at approximately one quarter of the request costs of the pure shell-DCI approach. On a corpus of one million documents, RISE with BM25 and gpt-4-mini reaches 81 percent accuracy, while the DCI baseline with gpt-4-nano drops to 60 percent and fails 33 of 100 requests within the time budget.
Source: arxiv.org · Published June 4, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.6.5.