DetouR plug‑in boosts multi‑hop QA agents by leveraging hyperlink routes
The DetouR plug‑in lets search agents follow hyperlink‑based association routes, improving multi‑hop question answering accuracy by up to 4 EM points without.
Source: Grafy Research LabWhat is DetouR and why it matters
Search agents that interleave reasoning with retrieval have become the standard way to tackle multi‑hop question answering (QA). The new plug‑in named DetouR (short for Detour Routing) shows that these agents can be made more effective simply by exploiting the hyperlinks that already appear on retrieved pages. According to the arXiv paper by Chen et al. (2026), DetouR improves exact‑match (EM) scores by points on the 2Wiki benchmark and by points on HotpotQA, without adding any extra search queries or large‑language‑model (LLM) calls.
How association routes differ from semantic routes
Current agents usually generate a semantic route: they infer what information is missing, turn that inference into a textual query, and rely on a retriever to match the query to passages. This approach only reaches pages the model can explicitly name. The authors point out that the decisive evidence for many multi‑hop questions is often just one hyperlink away from a page already retrieved, a path they call an association route. Empirical analysis on HotpotQA shows that – of gold pages missed by four baseline agents are linked from a page in the agents' top‑three results. When the same search budget is allocated to association routes, the agents reach of gold pages versus for semantic routes.
Key components of DetouR
DetouR is not a new search agent; it is a lightweight plug‑in that intervenes at two decision points:
- Link Rerouting – When a follow‑up query repeats an earlier one, the plug‑in swaps the query for a direct hyperlink to a page linked from the previously retrieved evidence. A small learned scorer (implemented with LightGBM) picks the most promising link. The paper notes that even upgrading this scorer with cross‑encoders up to parameters changes EM by at most points, confirming the lightweight model is sufficient.
- Chain‑aware Context Selection – After retrieval, the reader component must decide which passages to feed to the LLM. DetouR runs personalized PageRank over the hyperlink graph formed by the retrieved pages, seeded with the question‑named pages, and merges the ranked list with a round‑robin strategy. This keeps entire evidence chains together, preventing the loss of crucial linked pages that traditional relevance‑based or rank‑fusion methods (e.g., RRF) would discard.
Both steps add zero extra search calls and no additional LLM invocations, preserving the original agent’s computational budget.
Experimental results and impact
The authors attached DetouR to four distinct agents – IRCoT, ReAct, Search‑R1, and a plan‑then‑search baseline – and evaluated on two standard multi‑hop QA datasets:
- Overall gain – Average EM improvement of on 2Wiki and on HotpotQA.
- Component contribution – Link Rerouting alone adds EM on 2Wiki and on HotpotQA, while Chain‑aware Context Selection contributes and respectively. The combined effect exceeds the sum of the parts, indicating synergy.
- Model size effect – Heavier LLMs struggled to make use of the non‑semantic evidence; in some cases they performed slightly worse than the lightweight selectors.
The study also includes a live web pilot confirming that DetouR’s gains hold in real‑time settings, not just offline replay.
How to explore DetouR with Grafy
While DetouR itself is not yet a built‑in Grafy feature, creators and researchers can experiment with the same ideas using existing Grafy apps:
- Grafy Reader (/reader) lets you upload PDFs, such as the DetouR paper, and ask the model to extract and summarize specific sections. You can query the tool for the definition of association route or for the quantitative results, reproducing the paper’s findings on‑the‑fly.
- Grafy Chat (/chat) supports multi‑model conversations. You can prompt a model to simulate a search‑agent loop, then manually apply a hyperlink‑based detour by feeding the linked URL as a new query. This mimics DetouR’s Link Rerouting without any code changes.
By combining these two apps, you can build a lightweight prototype: retrieve a passage with Reader, inspect its outbound links, and feed the most promising link back into Chat as the next query. The workflow mirrors DetouR’s plug‑in logic and helps you gauge whether association routes improve your own retrieval‑augmented generation tasks.
Frequently asked questions
Does DetouR require a pre‑built hyperlink graph?
No. DetouR leverages the hyperlinks that arrive together with each retrieved page, so it works with on‑the‑fly web retrieval and does not need an offline graph construction step.
Can DetouR be used with any existing search agent?
Yes. The authors tested it on four diverse agents and reported consistent gains, indicating that the plug‑in is model‑agnostic and can be attached to any agent that performs query formulation and context selection.
Will DetouR increase the computational cost of a QA system?
According to the paper, DetouR adds no extra search calls and only a lightweight scoring step, so the overall computational overhead is negligible compared with the baseline agents.


