Many factual errors in language model outputs are not invented from nothing. The model takes a real fact from the source and attaches it to the wrong entity, patient, drug, or year, or it shifts a number slightly. Detectors that retrieve "the most similar passage" and ask an entailment model whether it supports the claim handle these binding errors badly. The retrieved passage contains the right predicate and a plausible-looking number, and the wrong subject is a single token away. We propose a cheap retrieval-only test for this failure. Split each generated claim into its anchors, meaning the exact strings that must match (names, identifiers, numbers, dates), and its content, meaning what is asserted about them. Find the source sentence whose meaning best matches the content using a dense sentence embedding. Then check whether every anchor appears by exact lexical match within a few sentences of that content sentence. A correct claim has its anchors and its content in the same local stretch of text. A swapped claim has its content in one place and its anchor somewhere else, or nowhere.
The idea joins three threads that have so far been combined only in pairs. Neighboring-sentence expansion, which widens a retrieved sentence to include the sentences around it, is known to rescue evidence that crosses sentence boundaries. Hybrid retrieval pairs sparse keyword matching, which is exact on rare tokens, with dense semantic matching, which handles paraphrase. Hallucination detection is the application. Our contribution is to use the two retrieval signals differently from how hybrid systems use them. Hybrid systems add the two scores together to rank a single unit. We keep the signals apart and treat the distance, measured in sentences, between the sparse evidence and the dense evidence as the verdict itself.
Core Claim
Consider a claim $c$ with anchor set $A(c)$ and a source document split into sentences $s_1, \dots, s_n$. Let $s_{j^*}$ be the sentence with the highest dense similarity to the claim with its anchors masked, and let $W_r(j^*)$ be the sentences within distance $r$ of it. The co-location score is
$$ \mathrm{CoLoc}(c) = \min_{r \in \{0,1,2\}} \Big\{ r \;:\; \forall a \in A(c),\ a \text{ matches some sentence in } W_r(j^*) \Big\}, $$
and it is set to "unplaced" when no window of radius 2 or less contains all anchors. Matching is exact after normalizing number formats and units.
In practice we take the minimum over the top few dense candidates rather than only the best one. The radius limit follows the evidence that expansion beyond roughly three sentences adds noise (LongDocFACTScore) and that verifiers degrade past about 200 tokens of grounding text (TRUE).
We claim two things. First, "unplaced" is a high-precision signal for entity swaps, role swaps, and near-miss numbers. Second, combining it with a standard entailment verifier run on the minimal window improves detection specifically on these binding errors, at a small fraction of the cost of multi-model pipelines. The masking step matters: it stops the dense matcher from simply finding the anchor, so the two signals stay independent.
Relation to Prior Work
Two systems motivate the design by showing that neighbor context is necessary for verification. FIZZ checks atomic facts against a single source sentence and zooms out to adjacent sentences when support is missing, reaching 71.2% balanced accuracy on AGGREFACT. However, it runs an 11-billion-parameter coreference model plus a decomposition model and repeated entailment checks, and its authors call it too slow for real-time use. LongDocFACTScore retrieves the most similar sentences with dense embeddings, expands each to its neighbors, and scores the snippets, reaching Kendall's tau 0.61 with experts on LongSciVerify. Its seed retrieval is dense-only. Both systems treat the neighbor window as better input for an entailment scorer. Neither treats the position of exact-match evidence relative to semantic evidence as information in its own right. We reuse their window sizes and replace most of their heavy machinery with a lexical check.
Two further papers establish that dense-only matching misses exact entities, and they supply the rationale for the sparse channel. Mala et al. show that hybrid retrieval raises MAP@3 on HaluBench to 0.897, against 0.768 for dense and 0.724 for BM25, and lowers medical hallucination from 20% (dense) to 4%. Luan et al. show that single dense vectors lose their ability to separate relevant from irrelevant text as length grows, while sparse matching stays stable. Both work at passage level and fuse scores into one ranking. We work at sentence level and deliberately do not fuse.
FinGround is the closest work in spirit. It decomposes answers into typed atomic claims, uses hybrid retrieval and formula re-computation, and reports 91.4% F1. It also reports that recall on near-miss numbers within ±5% drops to 71.4%, and it uses static chunking for prose. That near-miss failure is the specific gap we target. DENSPI shows that keeping sparse and dense vectors side by side over small units is fast at scale, which constrains our design toward precomputed sentence indexes. ChunkRAG shows that filtering noisy chunk content helps generation, but it treats chunks in isolation once they are formed.
Research Program
We will build a sentence-level index that stores, for each sentence, a BM25 posting, a dense vector, and its document position. At verification time, a small language model extracts anchors and a masked content string from each claim. Numbers, dates, and percentages are normalized with a fixed rule set, and alias lists handle entity paraphrase where the corpus provides them.
We will then compute CoLoc and run an off-the-shelf entailment model only on the minimal window $W_{r}(j^*)$. The final detector is a two-feature logistic combination of the entailment score and CoLoc, fit on a small development split.
The analysis stage labels each benchmark error by type: binding error, fabricated anchor, unsupported predicate, or reasoning error. It then reports gains broken down by type. We expect little help on unsupported-predicate and reasoning errors, and we will say so. A secondary study uses CoLoc at generation time on HaluBench. There we drop retrieved sentences whose query anchors fall outside their window, which tests whether the same signal also reduces hallucination before any claim is generated.
The known risks are specific. Abstractive paraphrase of anchors, such as "halved" for "50%" or "the firm" for a company name, will cause false flags. The extractive bias noted for LongDocFACTScore may reappear, and we will measure this by running the method separately on extractive and abstractive outputs.
Baselines
The naive reference point is dense-only snippet verification in the style of LongDocFACTScore: dense top-3 sentences, fixed ±1 expansion, and an entailment score. This is what the field does by default. On LongSciVerify its reported Kendall's tau is 0.61 (LongDocFACTScore), and we will rerun it with the same entailment model we use, so that only retrieval and co-location differ.
The strongest detection baseline on news summarization is FIZZ, at 71.2% balanced accuracy on AGGREFACT with AlignScore at 70.1% (FIZZ). We will report FIZZ's published number and also rerun it where feasible, recording wall-clock time per claim on the same hardware. Beating it means matching or exceeding 71.2% balanced accuracy overall and clearly exceeding it on the binding-error subset, at far lower latency. Matching it at one-tenth the cost would already be a useful result.
On financial text the reference is FinGround's reported near-miss recall of 71.4% and overall detection of 91.4% F1 (FinGround). We will compare against it only if FinHalu is available. If it is not, we will reconstruct near-miss cases on FinanceBench items inside HaluBench and state that this comparison is not like-for-like.
The strongest cheap competitor, and the one we most need to beat, is hybrid-seeded windowed entailment. This means Mala et al.'s weighted reciprocal-rank fusion (Mala et al.) used to pick sentence seeds, with ±1 neighbors and the same entailment model, but no co-location feature. All systems share the same entailment model, index, candidate depth, and development-set tuning budget for any thresholds.
Three ablations isolate the key ingredient. The first uses unmasked dense retrieval, which merges the two signals. The second checks anchors anywhere in the document instead of within the window. The third uses fixed radius 1 instead of the adaptive minimum.
Why not just…?
Why not just index overlapping three-sentence windows with hybrid retrieval and run entailment on them? This is the most serious threat, and it is the hybrid-seeded baseline above run over sliding windows. It already puts entity and predicate in one unit. The weakness is that the window's single dense vector blends the neighboring sentences, which is the length dilution Luan et al. describe. A window mentioning Drug A and Drug B can then support a swapped claim under a lenient entailment model. We expect that approach to close most of the gap on fabricated anchors and to lose on swaps between entities that both appear nearby. If it matches our detector on the binding-error subset, the co-location feature adds nothing and we will report that.
Why not just check that every anchor string appears somewhere in the source? This catches fabricated numbers and names, and it is the document-wide ablation. It cannot catch swaps where both entities occur in the document, which is exactly the role-swapping pattern documented in clinical summaries (Milecki). The gap between the document-wide check and the windowed check on swap cases directly measures the contribution of neighbor locality.
Why not just ask a strong language model to judge the claim against the full document? Self-detection is weak. ChatGPT reached 58.53% on HaluEval summarization (HaluEval), and FinGround reports a 78% hallucination reduction relative to GPT-4o. Frontier models have improved since, so we include a long-context model judge under the same token budget as an upper-reference baseline rather than dismissing it.
Why not just use FIZZ's full pipeline? It is accurate but slow by its authors' own account. Our claim is about accuracy per unit of compute on binding errors, and the latency-matched comparison settles it.
Decisive Experiments
We will run detection on AGGREFACT, LongSciVerify, the TRUE collection (where the three-metric ensemble reports ROC AUC 86.0, TRUE), and the finance and medical parts of HaluBench. We report balanced accuracy, ROC AUC, and Kendall's tau as each benchmark prescribes, together with per-claim latency.
We will also build a controlled diagnostic set. For each benchmark we take claims labeled consistent and create swaps: an entity replaced by another entity from the same document, a number replaced by another number from the same document, and near-miss numbers within ±5%. These controlled cases are reported separately from natural errors, because synthetic perturbations can overstate gains.
The decisive check has two parts. On natural binding errors, the combined detector must beat hybrid-seeded windowed entailment by a margin that holds under bootstrap resampling. On natural non-binding errors, it must not hurt. If CoLoc helps only on synthetic swaps, or if the sliding-window hybrid baseline matches it, the idea fails as stated, and the negative result will be reported as such.
References
- LongDocFACTScore: Evaluating the Factuality of Long Document Abstractive Summarisation
- TRUE: Re-evaluating Factual Consistency Evaluation
- FIZZ: Factual Inconsistency Detection by Zoom-in Summary and Zoom-out Document
- Hybrid Retrieval for Hallucination Mitigation in Large Language Models: A Comparative Analysis
- Sparse, Dense, and Attentional Representations for Text Retrieval
- FinGround: Detecting and Grounding Financial Hallucinations via Atomic Claim Verification
- Real-Time Open-Domain Question Answering with Dense-Sparse Phrase Index
- ChunkRAG: Novel LLM-Chunk Filtering Method for RAG Systems
- MEDIMP: 3D Medical Images and clinical Prompts for renal transplant representation learning
- HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models