Bridge Documents: Why Static Retrieval Utility Fails to Predict Causal Utility in Multi-Step Agentic Search
> *"Knowledge is a double-edged sword: it can help you, and it can blind you."* — adapted from Richard Feynman
*(Full English translation of a Chinese-language paper explainer by Xiaokai, originally published on zhichai.net.)*
The Detective's Dilemma
Imagine a detective investigating a complex case, with an assistant hired to sift through files. The assistant's criterion is simple: pick up a document, read it, and ask, "Does this directly answer the detective's question?" If yes, mark it useful; if not, discard it.
But discarded files turn out to matter: a bill reveals an unfamiliar phone number leading to a key witness; a restaurant receipt's timestamp breaks a suspect's alibi; a shopping list hints at a rare chemical recipe—the method of the crime. These documents don't directly answer "who did it," but they provide connections—bridges from known information to critical leads.
Debayan Mukhopadhyay, Utshab Kumar Ghosh, and Shubham Chatterjee found that LLM search agents systematically make this mistake: they discard documents that appear useless but are causally critical. The authors call these "Bridge Documents."
The Illusion of Retrieval: Do We Really Know What's Useful?
In traditional RAG systems, a document's "usefulness" is judged by pairing it with a reader model and checking whether the answer improves. This works for single-step reading. But when the language model becomes an active search agent—issuing multiple queries, reasoning across turns, adjusting strategy—this evaluation breaks down.
Why? In multi-step search, a document's value lies not in what it says, but in what it enables the agent to do next. A procurement record may not answer "who is the culprit," but a supplier name inside it may become the keyword of the next query, unlocking a chain of evidence.
This is the gap between Static RAG Utility (SRU) and Counterfactual Trajectory Utility (CTU).
The Counterfactual Experiment: Delete Every Document and Watch the World Collapse
The experimental setup:
- A ReAct-style search agent (alternating reasoning and action) answers 1,000 development questions from HotpotQA.
- For every document the agent reads, researchers remove it and re-run the entire trajectory from that step—asking: "What if the agent had never read this document?"
- Comparing original vs. counterfactual runs yields a CTU score based on three deltas: 1. Change in final answer quality 2. Change in next-query quality 3. Change in number of turns needed
- High SRU + High CTU: "Star documents"—good in both evaluations; usually contain the answer directly.
- High SRU + Low CTU: "Vase documents"—look useful, but dispensable in actual search.
- Low SRU + Low CTU: "Junk documents"—truly useless.
- Low SRU + High CTU: "Bridge documents"—rated useless statically, yet causally essential.
- Entities from relevant documents appear in the next query 6.1% of the time
- Entities from non-relevant documents: 1.5%
- Sample size: 227,139 entity observations
- Entity references: links to other documents or external resources
- Categorical information: labels that help the agent narrow the search
- Relational statements: descriptions of how things relate
- Exclusion information: explicit guidance on what not to consider
- Mukhopadhyay, D., Ghosh, U. K., & Chatterjee, S. (2026). *Bridge Evidence: Static Retrieval Utility Does Not Predict Causal Utility in Multi-Step Agentic Search*. arXiv:2607.15253.
- arXiv: https://arxiv.org/abs/2607.15253
- ReAct: Yao, S., et al. (2023). *ReAct: Synergizing Reasoning and Acting in Language Models*. ICLR 2023.
- OER (Observable Entity Relevance): Prior work on entity-based relevance measurement.
They then compare CTU against traditional SRU (does the document improve answer quality when given directly to a reader model?).
The result was startling.
Statistical Independence: A Rupture Between Two Worlds
Across 23,322 document observations, the Spearman correlation between CTU and SRU was -0.026—the two scores are nearly statistically independent.
Partitioning documents into four quadrants:
The last quadrant is the study's focus: roughly one-third of documents are bridge documents. The finding is robust across BM25 and cross-encoder agents (27.2% bridge units), indicating a structural feature of multi-step search agents, not a quirk of a particular model.
The Mechanism of Bridges: The Entity That Redirects the Search
Using Observable Entity Relevance (OER)—a measure of whether a document's entities distinguish relevant from non-relevant candidates—the experiments revealed a clear mechanism:
Discriminative entities appearing in the agent's next query are 4.02x more likely to come from relevant documents than non-relevant ones.
Bridge documents provide discriminative entities—information that helps the agent "rule out wrong options," "narrow the search," and "redirect queries." In the detective analogy: a bridge document is not testimony about the culprit, but a note saying "go interview this witness."
Why This Matters: Optimizing the Wrong Objective
Most retrieval systems—academic and industrial—optimize static retrieval utility: training models to judge whether a document directly helps answer a question. But this research proves that optimizing SRU does not improve causal retrieval utility—the two objectives are statistically independent.
This implies the need for a new evaluation framework for systems supporting multi-step search agents:
1. Introduce multi-step trajectories into evaluation: measure how deleting a document changes the whole trajectory, not just its direct help. 2. Identify structural features of bridge documents: what kinds of documents tend to be bridges? 3. Design retrieval models that recognize bridges: train models to predict "what will the agent do next because of this document" rather than "what answer does this document contain."
The Bigger Picture: The Value of Information Lies Not in the Information Itself
In classical information theory, value depends on content. In multi-step search, value depends on position—which node of the information network a document occupies, what it connects, what possibilities it opens. Like a small road serving little traffic but serving as the only link between two major communities: close it, and the city's connectivity collapses.
Bridge documents are those critical small roads. Their "surface value" is low, but their "structural value" is immense. They are not destinations, but paths to destinations.
The Future: Information Systems Built for Bridges
Bridge documents likely contain:
Future retrieval systems may have two modes: a "direct answer mode" (for simple questions) and a "bridge discovery mode" (for complex questions), the latter deliberately seeking documents that expand the search space, redirect queries, and provide connective clues—even when they contain no direct answers.
References
*Explainer: Xiaokai | Feynman-style deep dive | 2026-07-18*