CLAP: Coreference-Linked Passage Retrieval

• The paper introduces a lightweight LLM-based framework that segments passages, resolves coreference, and generates localized pseudo-queries to enhance first-stage retrieval.
• It employs a five-stage pipeline combining semantic chunking, dual encoding, and local-global score fusion to mitigate semantic drift and improve query alignment.
• Empirical results show significant nDCG@10 gains over baselines, validating CLAP’s logical augmentation strategy across both in-domain and out-of-domain settings.

Searching arXiv for the CLAP paper and closely related coreference-aware retrieval work. Coreference-Linked Augmentation for Passage Retrieval (CLAP) is a lightweight LLM-based expansion framework for first-stage passage retrieval that segments passages into coherent chunks, resolves coreference chains, and generates localized pseudo-queries aligned with dense retriever representations. It is motivated by two linked retrieval failures: LLM-based passage expansion often underperforms with dense retrievers because of semantic drift and misalignment with the retriever’s pretrained semantic space, and chunking can break coreference continuity so that only part of a passage is query-relevant while the remainder introduces noise. CLAP addresses these issues through a logic-centric, domain-agnostic pipeline that combines global topical signals with fine-grained subtopic signals, yielding consistent gains even as retriever strength increases and, in reported settings, enabling dense retrievers to match or surpass BM25 + MonoT5-3B (Xu et al., 9 Aug 2025).

1. Problem setting and conceptual basis

CLAP is designed for first-stage passage retrieval rather than for second-stage reranking or downstream answer generation. Its starting point is the observation that conventional passage expansion can be counterproductive for dense retrievers: added text may drift away from the original semantic content or fail to align with the retriever’s embedding space. The framework also targets a structural problem in passage indexing. Only a portion of a passage is typically relevant to a query, while the remainder contributes noise, and this issue is compounded when chunking breaks the referential continuity needed to interpret pronouns and other anaphoric expressions (Xu et al., 9 Aug 2025).

The framework is therefore “logic-centric” rather than knowledge-centric. In the paper’s formulation, CLAP grounds augmentation in the logical surface structure of passages instead of relying on domain knowledge. This matters particularly under domain shift. Prior passage and query expansion methods such as query2doc, docT5Query, HyDE, and LameR are reported to often harm retriever performance with strong dense backbones, especially in out-of-domain or zero-shot settings, because of semantic mismatch and brittle reliance on world knowledge. CLAP is presented as domain-agnostic precisely because it operates through segmentation, referential clarification, and local query synthesis rather than external factual elaboration (Xu et al., 9 Aug 2025).

A common misconception is that passage expansion helps dense retrieval simply by adding more lexical material. CLAP’s design rejects that premise. Its central assumption is that augmentation should improve representational alignment and reduce ambiguity, not merely enlarge the indexed text. A plausible implication is that the method’s gains come from controlling semantic granularity and referential clarity simultaneously.

2. Five-stage pipeline

CLAP consists of a five-stage modular pipeline. The stages are sequential, but the first two are implemented jointly in a single LLM call for simplicity.

Stage	Function	Output
1	Semantic chunking	Subtopic chunks
2	Coreference resolution	Disambiguated chunks
3	Pseudo-query generation	Multiple query-like texts
4	Dual encoding	Embeddings in same space
5	Local-global score fusion	Final passage scores

In Stage 1, an LLM agent $\mathcal{G}_c$ segments each passage $p_j \in \mathcal{P}$ into semantically coherent chunks with descriptive titles:

$$\mathcal{G}_c(p_j) \rightarrow \{(c_k, \text{title}_k)\}_{k=1}^{K}$$

This stage is intended to isolate subtopic units and prevent expansion from being dominated by unrelated content in multi-topic passages. The chunk titles serve as short semantic descriptors for later pseudo-query generation (Xu et al., 9 Aug 2025).

In Stage 2, an agent $\mathcal{G}_r$ resolves coreference within each chunk:

$\mathcal{G}_r(c_k) \rightarrow \hat{c}_k$

Here $\hat{c}_k$ is the coreference-resolved chunk in which pronouns and vague expressions are replaced by explicit antecedents. The purpose is to make each chunk self-contained after segmentation, since chunk boundaries can otherwise sever the context needed to interpret references (Xu et al., 9 Aug 2025).

In Stage 3, CLAP generates multiple localized pseudo-queries for each resolved chunk-title pair:

$$\mathcal{G}_q(\hat{c}_k, \text{title}_k) \rightarrow \{\psi(\hat{c}_k)_l\}_{l=1}^{L}$$

These pseudo-queries are intended to simulate diverse, user-like queries that are answerable from the chunk. The paper states the objective as minimizing the representational gap between the generated pseudo-query set $\mathcal{PQ}$ and the genuine query space $\mathcal{Q}$:

$$\min_{\mathcal{G}_q} \quad \mathbb{E}\left[\operatorname{dist}(\mathcal{Q}, \mathcal{PQ})\right]$$

The reported rationale is that chunk-bounded, coreference-resolved pseudo-queries better match the dense retriever’s learned query geometry than broader, knowledge-heavy expansions (Xu et al., 9 Aug 2025).

Stages 4 and 5 connect augmentation to retrieval. Queries, passages, and pseudo-queries are encoded by the same frozen dense retriever, and the resulting signals are fused during ranking rather than replacing the original passage representation.

3. Retrieval model and score fusion

CLAP uses a dual-view retrieval formulation. The first view is global passage relevance, computed directly between the query and the full passage:

$p_j \in \mathcal{P}$0

The second view is local relevance, defined as the maximum similarity between the query and any pseudo-query generated from any chunk of the passage:

$p_j \in \mathcal{P}$1

The final passage score is a convex interpolation of these two views:

$p_j \in \mathcal{P}$2

The local component is effectively a max-pooled passage-side expansion signal. It retains the strongest match among a passage’s pseudo-queries rather than averaging across them, which preserves fine-grained subtopic evidence when only one chunk is truly relevant. The global component preserves broader topical compatibility at the passage level. The paper reports that this fusion is more robust than relying on either view alone (Xu et al., 9 Aug 2025).

The same frozen dense retriever is used to encode all object types. This is a central architectural choice: pseudo-queries are not treated as free-form text to be matched lexically, but as additional points placed within the retriever’s semantic space. The paper argues that because these pseudo-queries are grounded in chunk-level logical structure rather than world knowledge, they are less susceptible to the semantic mismatch that affects other LLM-based expansion strategies (Xu et al., 9 Aug 2025).

The paper also reports empirical patterns for the fusion weight $p_j \in \mathcal{P}$3. Higher $p_j \in \mathcal{P}$4 values, meaning greater reliance on the global signal, are optimal for datasets with high passage/query multiplicity or longer passages. Lower $p_j \in \mathcal{P}$5 values, meaning stronger emphasis on the local signal, are better for short queries and tightly focused datasets. These are dataset-level observations rather than universal prescriptions.

4. Empirical results and ablation findings

CLAP is evaluated both in-domain and out-of-domain. On MS MARCO, the reported nDCG@10 results include a BM25 baseline of 38.01 and BM25 + CLAP at 47.41, an ANCE baseline of 56.52 and ANCE + CLAP at 59.16, and a SimLM baseline of 59.00 with SimLM + CLAP at 59.77 (Xu et al., 9 Aug 2025).

On BEIR datasets, the reported out-of-domain gains are larger. Examples include ArguAna with SimLM improving from 36.05 to 53.88, FiQA with ANCE improving from 29.49 to 55.78, FiQA with BM25 improving from 23.60 to 52.38, and e5-large-v2 on ArguAna improving from 51.77 to 63.08. The abstract characterizes the framework as delivering “up to 20.68% absolute nDCG@10 improvement,” while the detailed results list gains as large as +28.78 on FiQA with BM25. The paper also reports average gains of roughly +10 to +14 nDCG@10 across BEIR datasets on ANCE, SimLM, MiniLM, and e5-large-v2 (Xu et al., 9 Aug 2025).

Setting	Baseline nDCG@10	With CLAP
BM25 on MS MARCO	38.01	47.41
ANCE on MS MARCO	56.52	59.16
SimLM on ArguAna	36.05	53.88
ANCE on FiQA	29.49	55.78
BM25 on FiQA	23.60	52.38

The ablation evidence is especially important for understanding the role of coreference. On ArguAna, the paper reports that chunk-based pseudo-queries without coreference reduce nDCG@10 from a 50.15 baseline to 35.77, while adding coreference increases performance to 54.61; fusing local and global scores with $p_j \in \mathcal{P}$6 yields 58.11. This is a direct indication that localized pseudo-query generation alone is not sufficient and can be harmful when referential ambiguity is left unresolved (Xu et al., 9 Aug 2025).

Additional analysis shows that more than 90% of queries in FiQA and NFCorpus achieve positive similarity gain at chunk-level matching. The reported median gains are +0.19 to +0.22 in FiQA and +0.12 in NFCorpus. The paper also states that lower LLM temperature, specifically deterministic generation at 0.0, produces better retrieval performance, whereas higher temperature increases irrelevant drift. These observations are consistent with the framework’s emphasis on logical precision rather than creative paraphrastic variation (Xu et al., 9 Aug 2025).

5. Relation to coreference-aware retrieval research

CLAP belongs to a broader research trajectory in which coreference is treated as a retrieval variable rather than solely as a preprocessing step for information extraction or reading comprehension. Several adjacent lines of work illuminate the technical space that CLAP occupies.

A closely related study on Retrieval-Augmented Generation systematically investigates how explicit coreference resolution affects both retrieval and generation. It applies an LLM-based function to generate resolved documents, $p_j \in \mathcal{P}$7, and reports consistent gains in nDCG@k across retrievers as well as QA improvements across instruction-tuned LLMs. That study further reports that mean pooling demonstrates superior context capturing ability after coreference resolution and that smaller LLMs benefit more from disambiguation, likely because they have less inherent capacity to handle referential ambiguity (Jang et al., 10 Jul 2025). CLAP shares the same underlying intuition—that explicit antecedent recovery changes the semantic accessibility of passage content—but operationalizes it for first-stage retrieval through chunk-local augmentation and local-global fusion rather than end-to-end RAG evaluation.

A second related direction focuses on reranking rather than first-stage retrieval. The Embedding-Based Context-Aware Reranker (EBCAR) reranks candidate passages directly in embedding space using document ID embeddings, passage position encodings, and a hybrid attention mechanism that combines shared full attention with dedicated masked attention. The reported purpose is to capture both global cross-document interactions and local intra-document structure, including coreference and entity disambiguation. On ConTEB, EBCAR is reported to outperform rerankers that process passages in isolation, particularly on tasks such as Football, Geography, and Insurance where non-local mention linking and structural cues are essential (Yuan et al., 15 Oct 2025). Compared with EBCAR, CLAP intervenes earlier in the pipeline and modifies the indexed passage representation itself rather than jointly reranking a retrieved candidate set.

A third line of work reframes cross-document event coreference as a search problem. Cross-document Event Coreference Search introduces CoreSearch, a large-scale dataset of approximately 2.77 million passages derived from WEC-Eng, and adapts DPR using SpanBERT and mention boundary tokens. It then adds an integrated coreference scorer for mention-level matching, improving test performance from MRR@10 = 88.18 and F1 = 82.92 to MRR@10 = 90.06 and F1 = 84.16 (Eirew et al., 2022). This work demonstrates that coreference links can be built directly into retrieval objectives and ranking functions, especially when the task itself is inherently mention-centric.

An earlier neural modeling approach incorporates coreference annotations into recurrent computation through the Coref-GRU layer. Instead of relying only on sequential recency, the layer propagates hidden state along coreference chains and dynamically weights sequential and coreferent signals. Reported gains on Wikihop, LAMBADA, and the bAbi tasks are especially strong in low-data settings, and simply appending one-hot coreference features is shown to be much less effective than structural propagation along coreference arcs (Dhingra et al., 2018). This provides an architectural precedent for the broader claim that coreference is most useful when it shapes the representation pathway itself. CLAP is consistent with that principle, but instantiates it through passage segmentation, resolution, and pseudo-query synthesis rather than recurrent state design.

6. Efficiency, limitations, and interpretive issues

CLAP is described as plug-and-play: it requires no retriever finetuning and changes only the passage-side representation. The main computational cost is offline. The paper reports an LLM preprocessing cost of approximately $p_j \in \mathcal{P}$81.6k for MS MARCO dev. It also reports no increase in online retrieval latency because the augmentation is performed before indexing (Xu et al., 9 Aug 2025).

The primary efficiency trade-off is index growth. Each passage reportedly expands the index by 20–32$p_j \in \mathcal{P}$9 because multiple pseudo-queries are generated for multiple chunks. This makes CLAP lightweight at inference time but not free in storage or preprocessing. The framework therefore shifts computation from online reranking to offline indexing (Xu et al., 9 Aug 2025).

Two interpretive cautions follow from the reported ablations. First, CLAP should not be reduced to “coreference resolution for retrieval.” The ArguAna ablation indicates that chunking plus pseudo-query generation without coreference can harm performance, but the best result comes from fusing local and global signals rather than relying on local expansion alone. Second, CLAP should not be understood as a generic endorsement of LLM-based passage expansion. The paper explicitly states that conventional LLM-based expansion methods often underperform with dense retrievers because of semantic drift and misalignment with the pretrained semantic space. A plausible implication is that CLAP’s contribution is not the use of an LLM per se, but the use of an LLM under strict logical constraints: semantic chunking, referential disambiguation, localized pseudo-query generation, and shared-space encoding (Xu et al., 9 Aug 2025).

In that sense, CLAP occupies a specific position in retrieval system design. It is neither a purely lexical expansion method nor a heavy reranking architecture. It is a passage-side augmentation framework that treats referential coherence as part of retrieval representation, and it does so with a domain-agnostic pipeline that is reported to remain effective even as the underlying dense retriever becomes stronger.