Exploring “CORE: Coarse-to-Fine Search with Context Management for Deep Research” for SEO/GEO Insights: A Hamsterdam Research Post

By Ethan Lazuk

Welcome to another edition of Hamsterdam Research! 🐹

This is where we look at recent AI research papers to learn what they’re talking about and explore their hypothetical implications for the future of search and SEO/GEO strategies.

This time, we’ll look at research called, “CORE: Coarse-to-Fine Search with Context Management for Deep Research.”

It was submitted to arXiv on November 5, 2026, and its authors are Hoang Tran Vuong, Ngoc Quan Pham, Hung Bui, Fatih Porikli, and Dat Quoc Nguyen.

First off, why should SEO/GEO professionals care about this research?

This paper explores how AI research agents find, select, and manage information before generating an answer. In short, getting a source retrieved isn’t the same as getting its information used.

Here are the principal takeaways from the paper that may intrigue SEO/GEO professionals:

1. AI search is increasingly about retrieving passages.

CORE first retrieves broader passages to evaluate relevance, then identifies smaller snippets that directly address a research question.

2. Extractability matters alongside relevance.

The research demonstrates that an AI agent can perform better when it receives carefully selected, relevant snippets instead of large amounts of potentially distracting text. So always ask yourself, how easily can a retrieval system isolate the specific information that makes my page useful?

3. Context remains essential.

Although CORE prioritizes concise snippets, it also allows agents to retrieve additional surrounding information when necessary. This suggests that content should balance self-contained explanations with sufficient supporting context, particularly when discussing relationships between entities, complex concepts, or multi-step processes.

4. AI search is an iterative research process.

CORE’s agents formulate follow-up queries, discover additional evidence, revisit earlier sources, and avoid processing redundant snippets. It’s thus useful to consider how your content supports different stages of a research or purchase journey, rather than just matching an initial prompt.

5. The retrieval system can influence answer quality independently of the LLM.

The researchers kept agent models constant while changing the search harness. CORE achieved higher answer accuracy and evidence recall across all nine model-benchmark comparisons. This type of point reinforces why GEO can’t be fully understood just through prompt engineering or the language model itself.

In summary, for AI search, discoverability is only part of the challenge. What happens to content after retrieval, including which excerpts are selected and how they are incorporated into an agent’s context, can be just as consequential for the answer.

Let’s start by reviewing the paper’s abstract.

“Deep research agents answer complex questions through repeated search. They rely on a harness that runs their tool calls and returns evidence to their context. This creates a trade-off: Search works best over long passages, which give enough context for reliable semantic matching. But the returned evidence should be short, because every observation stays in the agent’s context for the rest of the trajectory. We introduce CORE, a training-free harness that combines coarse-to-fine search with context management. CORE retrieves passages with a dense retriever. It then reuses the token-level similarities from a late-interaction reranker to return only the most relevant snippet from each passage. Context management lets the agent expand the context around sources it has already retrieved and suppresses repeated snippets. In controlled closed-search comparisons with the Tongyi DeepResearch and OpenResearcher harnesses, CORE achieves state-of-the-art answer accuracy and evidence recall across three agent models and three benchmarks (BrowseComp-Plus, BrowseComp-Wikipedia and BrowseComp-PlusCM), on corpora containing up to 553M documents.”

Next, let’s break down the paper’s key vocabulary terms:

• Coarse-to-Fine Retrieval: A two-stage search approach that first identifies broadly relevant passages and then narrows the results to the most relevant pieces of information.
• Late-Interaction Reranking: A retrieval technique that compares individual query and passage token representations to refine relevance rankings after an initial search.
• Snippet Localization: Identifying the specific portion of a retrieved passage that most closely matches a search query.
• Context Expansion: A process that allows an AI agent to retrieve additional surrounding information from a previously discovered source without repeating a full search.
• Cross-Search Deduplication: A context-management technique that identifies previously retrieved snippets and replaces repeated instances with compact references.

Bravo. Let’s take a deeper look at the research paper’s contents now.

If you want to follow along, you can grab a PDF or the HTML version on arXiv.

The paper has five main sections.

We’ll focus on the most relevant ones below.

Introduction

Below I’ll paste in excerpts from the introduction that provide essential context for understanding the paper:

“… deep research is context-intensive. A single question may require dozens of searches, and every observation remains in the context for all subsequent turns (Chen et al., 2026a). Current harnesses search at the document or passage level and append the retrieved text largely as is (Zheng et al., 2025; Meng et al., 2026; Li et al., 2026a), although the information resolving a sub-query often spans only a few sentences. The rest consumes the context budget and dilutes the evidence on which later decisions depend.

Searching at a finer granularity is not a straightforward remedy, because retrievers need context for reliable matching. Prior work illustrates this difficulty: to make fine-grained retrieval effective, Chen et al. (2024) use a language model to rewrite the entire corpus into self-contained propositions.

This creates a conundrum: the text returned to the agent should be concise, but finding it requires searching over broad, contextual passages. Existing systems search at the document level and pass the retrieved text to a language model largely unfiltered. They rely on the model’s attention and reasoning capability to locate the few relevant sentences buried in redundant context. This model may be the agent itself browsing the documents (Li et al., 2026b), or a summarization model applied either to the visited documents (Tongyi DeepResearch Team et al., 2025) or, afterwards, to the accumulated history (Wu et al., 2025; Chen et al., 2026a). In all cases, a language model bears the cost of locating evidence within the largely irrelevant text that retrieval returns. On the other hand, locating relevant text within a passage is a matching problem, and a specialized retrieval model can solve it more cheaply.

The key idea of our CORE, a coarse-to-fine deep research harness, is to shift this burden from the agent to the search itself: relevant text is found by searching over broad passages, while the agent receives only concise snippets. We organize the search space into documents, passages, segments, and snippets, and treat evidence as the text admitted into the agent’s context to serve the local goal of a sub-query or the global goal of the question. CORE’s search tool first performs coarse retrieval of passages with a reasoning-aware dense retriever (Chen et al., 2026b), then fine-grained reranking with a late-interaction model (Khattab and Zaharia, 2020; Chaffin, 2025), whose token-level similarities also locate the most relevant snippet within each passage at no additional encoding cost. CORE’s context manager lets the agent expand the context around previously retrieved sources and deduplicates snippets across searches.”

Figure 1 below shows how CORE works:

Experiments

First the researchers tested three harness setups:

• Tongyi DeepResearch: A system that retrieves relevant passages and uses additional tools to examine or summarize documents.
• OpenResearcher: A system that enables agents to search and inspect retrieved documents.
• CORE: The proposed system, which combines coarse-to-fine retrieval, snippet localization, context expansion, and cross-search deduplication.

All systems searched the same underlying passage indexes, with the agent models and evaluation settings held constant within each comparison.

CORE achieved the highest answer accuracy and evidence recall in all nine model-benchmark comparisons. The improvements were particularly noticeable when searching larger document collections.

Among the findings that stood out were:

• Better retrieval improved answer accuracy: CORE consistently found more of the relevant evidence and produced more correct answers.
• The advantage generally increased with corpus size: With GPT-5.4, CORE’s lead over OpenResearcher grew from 3.49 percentage points on the smallest corpus to 12.28 points on the largest.
• More context was not necessarily better: On the largest benchmark, CORE achieved higher accuracy while using less peak context than OpenResearcher with GPT-5.4 and Qwen3.8-27B.
• Lower evidence recall did not always mean poor answers: The largest corpus contained many documents reporting the same information. Agents could answer correctly without retrieving every relevant document.

In summary, the findings suggest that improving how evidence is delivered to an AI agent can substantially improve its performance, even without changing the underlying language model.

For SEO/GEO professionals, an important takeaway is that the quality of information selected from a source can matter as much as the source’s initial retrieval. This research demonstrates that AI agents can generate more accurate answers when retrieval systems identify relevant passages, extract concise evidence, and preserve access to supporting context. This provides a technical foundation for thinking about content beyond traditional page-level relevance, raising questions about how clearly a document communicates individual facts, how easily those facts can be extracted, and whether sufficient context exists for an AI agent to interpret the facts accurately.

Conclusion

Below is the full conclusion:

“We presented CORE, a training-free deep research harness that combines coarse-to-fine search with context management. CORE retrieves whole passages to judge relevance, then reuses the reranker’s token-level similarities to return only the most relevant snippet from each one. Context management lets the agent expand the context around sources it has already retrieved and suppresses repeated snippets. Across three agent models and three benchmarks, CORE achieves higher answer accuracy and evidence recall than Tongyi DeepResearch and OpenResearcher in every model–benchmark comparison, while returning substantially fewer tokens per search than Tongyi DeepResearch. These results show that deep research agents can answer more accurately from shorter search responses when relevant text is selected before it enters their context.”

Knowing what we do now, why should SEO/GEO professionals care about “CORE: Coarse-to-Fine Search with Context Management for Deep Research”?

Perhaps the most interesting implication of this research is that AI visibility isn’t just about whether a page gets retrieved. It’s also about how effectively that page supports an agent’s subsequent research.

Here are practical takeaways implied by the research paper’s findings:

1. Optimize for follow-up questions, not just the initial prompt.

CORE’s agents conduct multiple searches, with each new query informed by previously retrieved evidence. Practically speaking, map the follow-up questions a user or research agent might ask after encountering your content, then address those questions in relevant sections or supporting pages.

2. Make your content useful for intermediate reasoning steps.

An AI agent may retrieve a source to resolve one small part of a larger research problem rather than answer the original question directly. Practically speaking, identify the supporting facts, comparisons, definitions, and relationships that would help an AI agent complete individual steps of a complex task.

3. Strengthen relationships between related information.

CORE’s context expansion lets agents revisit previously discovered documents to investigate entities and relationships mentioned in earlier snippets. Thus, when introducing an important entity, methodology, or concept, provide meaningful connections to related information within the document.

4. Prioritize information gain over repetition.

CORE suppresses snippets already encountered during the research process, reducing unnecessary repetition in the agent’s context. Audit related pages for repeated generic explanations and look for opportunities to contribute distinct evidence, original observations, or additional context.

5. Build content collections that support multi-hop research.

CORE’s experiments demonstrate the importance of gathering evidence across multiple sources when solving complicated questions. Practically speaking, develop complementary content that addresses different aspects of a broader topic.

Outro

The paper suggests a shift in how we think about content optimization. We’re going from “How do we make our content relevant enough to be retrieved?” toward “How do we make our content valuable throughout an AI agent’s research process?”

Making content relevant throughout an AI agent’s research process includes contributing evidence to intermediate reasoning steps, supporting follow-up investigations, providing distinctive information, and making the surrounding context useful when an agent needs to investigate further

I hope you’ve enjoyed this edition of Hamsterdam Research! 🐹

Feel free to comment below or contact me with your feedback.

Stay tuned for another new article, hopefully next week, or check out related research posts below.

Until next time, enjoy the vibes:

Thanks for reading. Happy optimizing!