Neighborhood Relays for Lexically Divergent Multi-Hop Retrieval

Many questions need two documents: one that names an intermediate "bridge" entity, and a second that states the fact about it. Retrieval systems usually find the second document by searching with the question plus the first document's text. That works when both documents describe the bridge in the same words. It fails when they don't. The first passage may say "Nolan" while the answer passage says only "the director of Memento." It may say "was acquired by" while the other says "following the amalgamation." In that case neither the question nor the first passage shares vocabulary with the target.

Our observation is that the corpus usually contains a third passage that uses both vocabularies. We propose to build, offline, a small graph around every passage that links it to passages sharing an explicit mention with it. At query time, a lightweight attention module reads that neighborhood, guided by the question, and adds the neighbors' vocabulary to the next-hop search vector. The system therefore uses local neighborhood aggregation, the graph-learning practice of representing a node by pooling its immediate neighbors. Here it serves as a vocabulary relay between hops, with no language-model text generation at test time.

Core Claim

A bounded, question-conditioned aggregation over the hop-1 passage's mention neighborhood recovers hop-2 evidence that has near-zero word overlap with both the question and the hop-1 passage. It does this more cheaply than having a language model write a new query, and more reliably than enlarging the beam of an existing dense multi-hop retriever. The claim is specific to the lexically divergent part of multi-hop retrieval. We expect small or no gains where the two hops already share words, and the evaluation is built to show that split explicitly.

Written out, the next-hop query vector is

$$q_2 = f(q, p_1) + \sum_{j=1}^{k} \alpha_j W n_j, \qquad \alpha_j = \mathrm{softmax}_j\!\left(f(q,p_1)^\top U n_j\right).$$

The symbols are as follows. $f(q, p_1)$ is the base retriever's encoding of the question $q$ joined with the retrieved first passage $p_1$. The $n_j$ are precomputed embeddings of the $k$ neighbors of $p_1$. $U$ and $W$ are small learned matrices. The weight $\alpha_j$ says how much neighbor $j$'s vocabulary should shift the search, given what the question needs.

The residual form means the module can learn to contribute nothing when the base query already works. We cap the neighborhood size $k$ because the one-shot knowledge-graph work in One-Shot Relational Learning for Knowledge Graphs found that encoding more than about 30 neighbors injects noise. That paper encoded one-hop graph neighborhoods to represent rare entities. It is the source of both our neighbor encoder design and our budget constraint.

Position Relative to Prior Work

Three graph-based retrieval systems already mix neighborhoods into multi-hop retrieval, and they motivate this proposal more than they pre-empt it.

Hierarchical Lexical Graph for Enhanced Multi-Hop Retrieval breaks text into atomic propositions and connects documents through shared, symbolically extracted entity nodes. When two documents name an entity or relation differently and extraction does not merge them, the path simply does not exist. Our relay passes through a third passage in embedding space instead of requiring the two ends to be symbolically merged.

SiReRAG builds separate similarity and entity-relatedness summary trees. It showed that similarity alone covers under 20% of the connections multi-hop questions need. It also reported that letting the two trees interact during construction hurt, so it fused them only at the end. We read that negative result as a warning against mixing neighborhoods blindly at index time. Our mixing is conditioned on the question at retrieval time and can switch itself off.

Query-Centric Graph Retrieval Augmented Generation links synthetic Doc2Query questions by nearest-neighbor similarity. Its entry point therefore fails whenever the user's wording was not anticipated by the query generator. None of these three systems learns what to take from the neighborhood.

The closest mechanism in the knowledge-graph line is JAKET. JAKET runs graph attention over entity neighborhoods initialized from text descriptions and caches the results in a momentum memory. It assumes entities are already aligned to a curated knowledge graph. FIRE plays the same role for few-shot multi-hop reasoning, encoding each node's neighborhood inside a reinforcement-learning path search over closed knowledge graphs.

We reuse JAKET's idea of an offline neighbor cache so that the test-time cost is one attention over $k$ vectors. We drop the curated graph: our graph comes from raw passages, hyperlinks where they exist, and exact-string mention overlap.

Two lines of work supply the training signal and the problem statement. For training, RaFe trains small query rewriters from cross-encoder reranker feedback instead of language-model feedback. We apply the same supervision to the aggregator on domain corpora without gold hop labels. For the problem, WikiQA documents that 20.3% of correct answers share no content words with their question, which motivates stratifying our evaluation by lexical overlap.

Research Program

Index construction. For each passage we collect up to $K$ neighbors. These come from hyperlink targets and passages sharing an exact-match noun-phrase mention, found with an off-the-shelf tagger and no language-model calls. We rank them by the base encoder's similarity and store each neighbor's embedding.

Test-time retrieval. The base retriever returns hop-1 candidates, as in two-step dense multi-hop retrieval. For each hop-1 candidate, the aggregator forms $q_2$ from that candidate's cached neighbors and performs the hop-2 search.

Base encoders. We use two. The first is Contriever, the dense retriever that Optimizing Multi-Hop Document Retrieval Through Intermediate Representations maps language-model hidden states into. The second is the decomposition-free encoder of GRITHopper. Both are kept frozen, so any gain is attributable to the aggregator.

Training. The aggregator is trained with a contrastive loss on gold hop-2 passages from the MuSiQue, 2WikiMultiHopQA, and HotpotQA training sets. For domain corpora without gold chains, it is trained with reranker-derived preferences in the style of RaFe.

Analysis. We will inspect which neighbors receive high attention weight. This tests whether the mechanism actually works as a relay: high weight should go to neighbors whose words overlap the gold hop-2 passage, not to generic topical passages.

Evaluation slices. We build the evaluation around a lexical-divergence split. The first slice stratifies hop-2 gold passages in MuSiQue and 2WikiMultiHopQA by content-word overlap with the question and the hop-1 passage. MuSiQue was built to remove shortcut questions answerable without true composition. The second slice is a controlled perturbation: the bridge-entity mention inside each hop-2 gold passage is replaced by an alias or a definite description drawn from elsewhere in the corpus, which forces a true lexical gap. The third slice is out-of-distribution: the SEC-10Q and NTSB corpora from the Hierarchical Lexical Graph paper, where no Wikipedia alias tables exist.

Baselines

Naive reference. Single-pass dense retrieval of the question alone. It shows how much of the multi-hop gap exists on each overlap stratum. For scale, MultiHop-RAG reports a best Hits@10 of 0.7467 even with reranking on its news corpus.

Strongest dense competitor. GRITHopper, which embeds the accumulated context directly into the next-hop query. It is the strongest competitor because it is also generation-free and already strong out of distribution: on MoreHopQA it reports Hits@1 of 95.44 against 82.16 for MDR. We found no reported GRITHopper retrieval Hits number on MuSiQue on Lacuna; only its runtime of 6 minutes 42 seconds is reported. We will therefore rerun it ourselves on identical corpora and candidate budgets. Beating it means a statistically significant Recall@k gain on the low-overlap strata and the perturbation set, with no more than a one-point loss on high-overlap strata.

Hidden-state competitor. L-RAG, which reports 0.352 accuracy on MuSiQue at about 2.8 seconds per query. It is run from its released code with the same reader model as all other systems.

Generation-based competitor. IRCoT, which interleaves chain-of-thought sentences with retrieval. It reports 11–21 point recall gains over one-step retrieval, but we found no absolute per-dataset figure on Lacuna. It is run with the same retriever, the same reader, and a matched number of retrieval calls. Here the target is parity or better at a fraction of the latency, not necessarily outright wins.

Graph competitors. QCG-RAG, which reports 79.60% accuracy on MultiHop-RAG, and HLG's StatementGraphRAG, which reports 73.6% average correctness on its domain sets. Both are run through their public pipelines, and we report their index-construction language-model calls alongside ours.

Ablations. Each isolates one ingredient. Replacing $p_1$'s neighbors with random passages of equal count tests whether the relay content matters. Replacing attention with mean pooling tests whether question conditioning prevents drift. Sweeping $K$ from 4 to 64 tests the noise threshold reported in GMatching.

Why not just…?

Why not just let an LLM write the next-hop query, as IRCoT does? A fluent model can often paraphrase its way across a lexical gap, so this is a real competitor on accuracy. It costs repeated generation calls: iterative methods exceed 20 seconds per query in the L-RAG measurements, and GPT-4o decomposition took over an hour on MuSiQue in GRITHopper's timing. It also cascades when hop 1 is wrong, as BeamAggR documents. More fundamentally, the model can only paraphrase into vocabulary it knows. On the SEC-10Q and NTSB corpora, the relay terms are corpus-specific. If IRCoT matches us there at equal retrieval calls, our contribution reduces to a latency gain, and we will report it that way.

Why not just concatenate the hop-1 passage and do dense retrieval, or expand the query once? Concatenating the hop-1 passage is exactly what $f(q, p_1)$ already does. The neighborhood term is added only on top of it, so this alternative is the $\alpha = 0$ baseline in every comparison. One-shot expansion such as RaFe or InteR cannot name a bridge entity before hop 1 is retrieved. InteR also degrades beyond two expansion iterations. The perturbation set is designed so that the words shared by the question and $p_1$ are absent from the hop-2 target.

Why not just widen the dense beam or add a third hop, so the relay passage gets retrieved on its own? This is the most serious threat, because a relay passage is, in effect, an unlabeled intermediate hop. We will give MDR-style and GRITHopper chains a third hop and a beam large enough to match our candidate budget, measured as passages scored per question. If they recover the same hop-2 targets, the neighborhood adds only efficiency. Our bet is that relay passages sit low in the unconditioned ranking because they are not about the question. Pre-linking them to $p_1$ by explicit mention sharing is what surfaces them.

Why not just resolve aliases with an entity linker? On Wikipedia this may close much of the gap, so we report an alias-dictionary baseline on the Wikipedia benchmarks. It does not address relational paraphrase ("acquired" versus "amalgamation"). It is also unavailable for financial and aviation corpora, which is why those corpora are in the evaluation.

Proposed Experiments and Decisive Checks

The main experiment reports Recall@2, Recall@5, and Recall@10 of hop-2 gold passages, together with downstream answer F1 using a fixed reader. Results are broken out by lexical-overlap stratum on MuSiQue, 2WikiMultiHopQA, and HotpotQA, on the alias-perturbation set, and on SEC-10Q and NTSB. Per-query latency and offline indexing cost are reported for every system.

The decisive check is the interaction pattern. The proposal is supported only if the aggregator's gain over GRITHopper and over the widened-beam baseline grows as overlap falls, and exceeds the random-neighbor ablation on the perturbation set.

Two secondary checks follow. First, attention weights should concentrate on neighbors whose vocabulary overlaps the gold hop-2 passage; if weight goes mainly to topical neighbors instead, the relay explanation fails. Second, performance should peak at a moderate $K$ and fall beyond roughly 30 neighbors, reproducing GMatching's noise threshold in the text setting.

Falsification is concrete. If the widened third-hop beam matches the aggregator on the low-overlap strata at equal budget, the explicit neighborhood carries no information beyond what dense search already finds. We would then report the idea as an efficiency trick rather than a retrieval gain.