Research map / Large language models / Retrieval and knowledge

Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: Retrieval-augmented reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes direct corpus interaction (DCI), letting LLM agents search raw corpora with terminal tools (grep, read, bash) instead of a top-k retriever, and shows gains on BrowseComp-Plus, multi-hop QA, and BRIGHT/BEIR ranking.

Problem

Fixed top-k similarity retrievers (sparse or dense) compress corpus access into one step, which bottlenecks agentic search requiring exact lexical constraints, weak-clue conjunctions, local context checks, and multi-step refinement.

Approach

Replace the retriever interface with direct grep/read/bash access to the raw corpus, so the agent performs semantic interpretation itself and can pipeline fine-grained search primitives; no embeddings, index, or retrieval API. Explained via "retrieval interface resolution" and coverage/localization metrics.

Key results

• On BrowseComp-Plus with Claude Sonnet 4.6, DCI raises accuracy from 69.0% (Qwen3-Embedding-8B retrieval) to 80.0% while cutting cost from $1,440 to $1,016.
• DCI-Agent-CC averages 83.0% on multi-hop QA, 30.7 points above ASearcher-Local-14B (52.3%).
• DCI-Agent-CC reaches 68.5 average NDCG@10 on BRIGHT/BEIR subsets, 21.5 points above ReasonRank-32B (47.0).

Surprising findings

• DCI-Agent-Lite has lower mean gold-document coverage than the embedding retriever (28.0 vs 56.7) yet achieves higher accuracy (73 vs 45 on the 100-question subset).
• Scaling the corpus from 100K to 400K documents degrades DCI-Agent-CC accuracy to 37.5%, showing a cost limit on breadth search.

Limitations

• Cost and tool calls grow sharply as corpus size increases (100K to 400K docs), limiting breadth-heavy search.
• Gains depend on the strength of the agent backbone's strategic search ability.
• Multi-hop QA and IR evaluations use 50-example random samples and an LLM judge (GPT-4.1).

Experimental setup

Datasets: BrowseComp-Plus NQ TriviaQA Bamboogle HotpotQA 2WikiMultiHopQA MuSiQue 2018 Wikipedia dump BRIGHT BEIR ArguAna BEIR SciFact FineWeb

Benchmarks: BrowseComp-Plus NQ TriviaQA Bamboogle HotpotQA 2WikiMultiHopQA MuSiQue BRIGHT BEIR

Metrics: accuracy NDCG@10 cost coverage localization average tool calls latency

Baselines: Qwen3-Embedding-8B retrieval agent BM25 retrieval agent ASearcher-Local-14B Search-R1-32B R1-Searcher-7B ZeroSearch-7B Verl-Tool-Search-7B-DAPO ReasonRank-32B Rank1-32B Rank-R1-14B GTE-Qwen2-7B-Instruct OpenAI-text-embedding-3-large

Builds on: Pi Claude Code ASearcher BrowseComp-Plus BRIGHT

Closest papers by meaning

• WideSearch: Benchmarking Agentic Broad Info-Seeking (ICLR 2026 · similarity 0.89)
• InteractComp: Evaluating Search Agents With Ambiguous Queries (ICML 2026 · similarity 0.89)
• Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation (NeurIPS 2026 · similarity 0.89)
• MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains (ICLR 2026 · similarity 0.88)
• Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge (NeurIPS 2025 · similarity 0.88)
• AgentHop: A Diagnostic Benchmark for Agentic Multi-Hop Scientific Question Answering (NeurIPS 2026 · similarity 0.88)