CORE retools the deep-research harness: only the best snippet enters context, and on a 553M-document corpus its answer accuracy leads by 12.28 points
Synopsis
CORE, a training-free harness from Qualcomm AI Research, splits retrieval into coarse and fine stages: a dense retriever returns whole passages, then token-level similarities from a late-interaction reranker are reused to hand the agent only the most relevant snippet from each passage. In controlled closed-search comparisons over fixed corpora, CORE reaches the highest answer accuracy and evidence recall in every model-benchmark group across three agent models and three benchmarks, with corpora ranging from 100K web documents to 553M ClimbMix documents.
Interpretation
The evidence a deep-research agent receives per search shrinks from whole passages to snippets, while retrieval still runs over passages long enough for reliable matching, so matching reliability and context cost stop trading against each other. Earlier open-source harnesses search at the document or passage level and append retrieved text largely as is, leaving the language model to find the few useful sentences inside redundant text. In controlled closed-search comparisons against the Tongyi DeepResearch and OpenResearcher harnesses, CORE has the highest answer accuracy and evidence recall in every one of the three agent models by three benchmarks groups.
The design moves the burden of locating evidence from the agent to the search itself: coarse retrieval picks candidate passages, and fine reranking picks the most relevant snippet inside each. Fine-grained reranking uses the ColBERT-based late-interaction model Reason-ModernColBERT, which compares query and passage token by token and sums each query token's strongest match into a passage score. The token-level similarity matrix computed during reranking is reused to score overlapping segments within each retained passage and select the top segment, with no additional encoding; context management lets the agent expand context around already retrieved sources and replaces repeated snippets with compact references.
Each component contributes a measurable gain, and the gain does not come from using more context. With Tongyi-DeepResearch-30B-A3B as the agent on BrowseComp-Plus, accuracy is 4.58% without tools and 62.61% with coarse retrieval only. Adding coarse-to-fine search on top of coarse retrieval raises accuracy to 70.84%, and adding context management reaches 80.12%; on the 553M-document corpus, CORE has the lowest median peak context for both GPT-5.4 and Qwen3.8-27B.
Perspective
The results rest on three benchmarks with fixed text collections and reference answers: BrowseComp-Plus with 830 questions over about 100K web documents, BrowseComp-Wikipedia with 232 questions over about 6.08M Wikipedia articles, and BrowseComp-PlusCM with 57 questions over 553M ClimbMix documents; the authors explicitly leave evaluation on live web search and open-ended research reports to future work.
Fixed corpora do not capture the dynamic and heterogeneous nature of the open web, and reference-answer-based evaluation may not fully reflect long-form synthesis quality in less constrained research tasks; on BrowseComp-PlusCM, evidence recall is much lower than answer accuracy for every harness because many documents state the same facts and the relevance sets include verified duplicates.