TL;DR
DetouR is a zero-cost plug-in for search agents that improves multi-hop question answering by following "association routes"—hyperlinks and entity connections—rather than relying solely on the model's ability to guess the next search term. By rerouting redundant queries to linked pages and using Personalized PageRank to keep evidence chains intact, DetouR boosts accuracy across various agents without requiring additional LLM calls or search budget.
What Problem Does This Paper Address?
Current search agents (like ReAct or IRCoT) solve complex, multi-hop questions by interleaving reasoning with retrieval. They typically follow semantic routes: the model looks at what it knows, figures out what’s missing, and generates a new text query.
However, this approach has a massive blind spot: a model can only search for what it can already name. In many multi-hop scenarios, the "gold" evidence is just one hyperlink away from a page the agent has already found, but the agent fails to "click" it because it hasn't yet inferred the specific name of that next entity.
Furthermore, even if an agent accidentally retrieves the right evidence, standard context selection methods (like Reciprocal Rank Fusion) often discard it. These methods favor passages that appear in multiple search results; a "detour" page that only appears once is frequently dropped before the reader LLM ever sees it.
Methods
The authors introduce DetouR, a two-part plug-in designed to exploit the existing association structure of the web (or Wikipedia) without offline graph construction.
- Link Rerouting: The system monitors the agent's queries. If a query is a "near-repeat" (determined by Jaccard similarity) of a previous one, DetouR intercepts it. Instead of wasting a search call on redundant results, it identifies hyperlinks within already-retrieved pages. A lightweight LightGBM scorer then selects the most promising link to follow, replacing the redundant query with a targeted search for the linked entity.
- Chain-aware Context Selection: Once all searches are complete, the agent must pick which passages to show the "reader" LLM. Instead of using standard rank fusion, DetouR runs Personalized PageRank (PPR) over the links between all retrieved passages, seeded by entities named in the original question. This ensures that entire "chains" of evidence are kept together, even if individual links in the chain don't look highly relevant to the query in isolation.
Crucially, DetouR adds zero extra LLM or search calls, making it a "free" upgrade for existing agent architectures.
Datasets
The authors evaluated DetouR using four different search agents (IRCoT, ReAct, Search-R1, and a "Plan-then-search" baseline) on two primary multi-hop benchmarks:
- HotpotQA: Utilizing the 5.23M Wikipedia abstracts and their 19.5M internal hyperlinks.
- 2WikiMultiHopQA (2Wiki): Since this dataset lacks explicit hyperlinks, the authors constructed an association graph by linking paragraphs where one paragraph's text mentions another's title.
They also conducted a "live web pilot" using DeepSeek Search to test the approach in a noisier, real-world environment.
Key Insights and Ideas
- The "Association" Advantage: On HotpotQA, 53–73% of the gold pages that agents typically miss are actually linked from the agent's own top-three search results. The evidence is literally right there, just ignored.
- Synergy is Mandatory: The authors found that rerouting (finding the link) and chain-aware selection (keeping the link) must work together. Rerouting alone adds very little value because the "detour" pages get filtered out by standard rankers. When combined, the gain is greater than the sum of the parts.
- Lightweight Beats Heavyweight: Surprisingly, the PageRank-based selector (which uses no model calls) outperformed 8B-parameter rerankers and LLM-based judges on 2Wiki and matched them on HotpotQA. This suggests that for multi-hop tasks, the structure of information is more informative than raw semantic relevance scores.
- The Redundancy Tax: About a quarter of agent follow-up calls are "near-repeats" that return almost no new gold evidence. DetouR turns this wasted budget into productive exploration.
Key Results
- Accuracy Boost: DetouR improved Exact Match (EM) scores for all four agents tested. On average, it added 4.0 points on 2Wiki and 1.4 points on HotpotQA.
- Significant Gains: The improvements were statistically significant in 13 out of 16 tested scenarios (4 agents × 2 datasets × replay/live modes).
- Evidence Retention: DetouR nearly eliminated the "lost at the merge" problem. A retrieved gold page missed the reader's context in only 0.2% of wrong answers on 2Wiki, compared to much higher loss rates for standard fusion methods.
- Efficiency: The LightGBM scorer and PPR calculations run on a CPU with negligible overhead (occupying only a few megabytes of memory), making it highly practical for production systems.
Limitations and Future Work
- Graph Dependency: DetouR relies on an association graph. While Wikipedia hyperlinks are clean, the "live web" is much noisier. The pilot study showed smaller (though still positive) gains on the open web, where links are often broken or replaced by ads.
- Entity Matching: The system currently struggles when multiple entities share the same name. Better entity disambiguation could further improve link selection.
- Reading vs. Retrieval: The authors note that while DetouR fixes many retrieval and retention issues, over 50% of remaining errors are "reading errors"—where the LLM has all the right evidence but still gives the wrong answer. Future work could involve teaching the reader LLM to better utilize the link structure provided in the context.
Summary updated