Upload folder using huggingface_hub

paper/main.tex

@@ -81,9 +81,11 @@
 \section{Conclusion}
 \input{sections/conclusion}
 
+\twocolumn[
 \appendix
 \section{Reproducibility Details}
 \input{sections/appendix}
+]
 
 \clearpage
 \bibliographystyle{plain}

paper/references.bib

@@ -6,7 +6,7 @@
 }
 
 @article{santhanam2022colbertv2,
-  title={ColBERTv2: Efficient and Effective Retrieval via Lightweight Late Interaction},
+  title={ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction},
   author={Santhanam, Keshav and Khattab, Omar and Saad-Falcon, Jon and Potts, Christopher and Zaharia, Matei},
   journal={Proceedings of NAACL},
   year={2022}

paper/sections/abstract.tex

@@ -1 +1 @@
-Late-interaction retrievers compare query and document token representations and have become a strong foundation for neural retrieval. Yet practical retrieval systems often continue to include lexical search, typically BM25, even when a semantic retriever is available. This paper asks a focused question: when does lexical search help a late-interaction retrieval pipeline? We evaluate BM25, semantic retrieval, lexical+semantic reciprocal-rank fusion (RRF), and ModernColBERT reranking over first-stage candidates on seven heterogeneous BEIR-style datasets. Our main finding is that lexical search helps most as part of hybrid candidate generation rather than as a standalone top-rank signal. On the seven complete datasets, ModernColBERT reranking over lexical+semantic RRF top-100 candidates improves average nDCG@10 from 0.3826 to 0.4211 over reranking BM25 top-100 candidates, winning on every dataset. The mechanism is candidate recall: average Recall@100 rises from 0.5262 for BM25 to 0.6743 for the hybrid first stage. These results suggest a practical design rule for late-interaction retrieval systems: use lexical search as complementary candidate-pool insurance, then apply token-level semantic interaction as the precision layer.
+Late-interaction retrievers compare query and document token representations, so it is natural to ask whether BM25 still adds anything once such a reranker is available. We study this question in a two-stage retrieval pipeline. The first stage is BM25, semantic retrieval, or lexical+semantic reciprocal-rank fusion (RRF); the second stage is ModernColBERT reranking over the resulting top-100 candidates. Across seven BEIR-style datasets, reranking lexical+semantic RRF candidates improves average nDCG@10 from 0.3826 to 0.4211 relative to reranking BM25 candidates, with gains on every dataset. The explanation is candidate recall. BM25 reaches 0.5262 Recall@100, while the hybrid first stage reaches 0.6743. In this setting, lexical search helps less by being a better final ranker and more by giving the late-interaction model relevant documents it would otherwise never see.

paper/sections/analysis.tex

@@ -1,12 +1,14 @@
 \paragraph{Candidate recall is the bottleneck.}
-The central result is not that BM25 is a stronger ranker than semantic retrieval. In fact, semantic retrieval beats BM25 as a first-stage ranker on six of seven complete datasets. Instead, the result is that BM25 remains useful when paired with semantic retrieval to construct candidates. The reranker is powerful but bounded: if a relevant document is absent from the top-100 candidates, no reranking method can move it into the final top 10.
+The central result is not that BM25 is a stronger ranker than semantic retrieval. Semantic retrieval beats BM25 as a first-stage ranker on six of the seven complete datasets. The point is narrower and, for system design, more useful: BM25 still helps when it is paired with semantic retrieval to build candidates. The reranker is powerful but bounded. If a relevant document is missing from the top-100 candidates, no reranking method can move it into the final top 10.
 
 \paragraph{Lexical overlap alone does not explain the effect.}
-We computed simple lexical features, including average query length, rare-token ratio, entity-like token ratio, and lexical Jaccard overlap between queries and relevant documents. A naive expectation would be that datasets with higher lexical overlap should favor BM25. The observed pattern is more complicated. ArguAna and CQADupStack Android have relatively high query--relevant-document lexical overlap, yet semantic retrieval strongly outperforms BM25 as a first-stage ranker. Meanwhile, SciFact is the only complete dataset where BM25 slightly beats semantic retrieval at nDCG@10.
+We computed simple lexical features: average query length, rare-token ratio, entity-like token ratio, and lexical Jaccard overlap between queries and relevant documents. A tempting rule would be to expect high lexical overlap to favor BM25. The data do not support such a simple cutoff. ArguAna and CQADupStack Android have relatively high query--relevant-document lexical overlap, yet semantic retrieval strongly outperforms BM25 as a first-stage ranker. SciFact is the only complete dataset where BM25 slightly beats semantic retrieval at nDCG@10.
 
-\begin{table*}[t]
+\begin{table}[t]
 \centering
-\small
+\scriptsize
+\setlength{\tabcolsep}{2pt}
+\resizebox{\columnwidth}{!}{%
 \begin{tabular}{lccc}
 \toprule
 Dataset & Lexical Jaccard & BM25 $-$ Sem. & RRF-rerank gain \\
@@ -20,16 +22,17 @@ CQADupStack Android & 0.1146 & -0.1243 & +0.0440 \\
 CQADupStack Mathematica & 0.0569 & -0.0892 & +0.0269 \\
 \bottomrule
 \end{tabular}
+}
 \caption{Simple lexical-overlap features do not directly predict BM25 superiority. The practical benefit of BM25 is more robust when used as one channel in a hybrid candidate generator. BM25 $-$ Sem. is first-stage nDCG@10 difference.}
 \label{tab:features}
-\end{table*}
+\end{table}
 
-Table~\ref{tab:features} supports a more operational interpretation. Lexical overlap is not a sufficient decision rule for choosing BM25 over semantic retrieval. But BM25 can still contribute complementary candidates even when semantic retrieval is stronger on average. This is exactly the setting where RRF is attractive: it does not require deciding which retriever is globally better; it preserves high-ranked evidence from both.
+Table~\ref{tab:features} points to a more operational interpretation. Lexical overlap is not enough to choose BM25 over semantic retrieval. BM25 can still contribute complementary candidates even when semantic retrieval is stronger on average. This is the setting where RRF is useful: it does not ask us to decide which retriever is globally better, and it preserves high-ranked evidence from both.
 
 \paragraph{Why RRF before late interaction?}
-One could fuse BM25 scores after late-interaction reranking, but our results suggest that the more important intervention is before reranking. Late interaction already provides a strong token-level semantic score. The main failure mode is missing relevant documents in the first-stage candidate set. RRF addresses that failure by broadening the candidate pool while keeping the budget fixed at 100 documents. This makes hybrid retrieval a recall-improving stage, not merely a score-mixing trick.
+One could also fuse BM25 scores after late-interaction reranking. The results here suggest that the more important intervention happens before reranking. Late interaction already gives a strong token-level semantic score. The failure mode we observe is more basic: relevant documents are missing from the first-stage candidate set. RRF addresses that failure by broadening the pool while keeping the budget fixed at 100 documents. In this setup, hybrid retrieval is a recall-improving stage, not just a score-mixing trick.
 
-Formally, post-hoc score interpolation can only reorder $\C_f^{100}(q)$ for a fixed $f$. Hybrid candidate generation changes the feasible set from $\C_{\text{BM25}}^{100}(q)$ to $\C_{\text{RRF}}^{100}(q)$. The observed gain is therefore consistent with a feasible-set expansion:
+Formally, post-hoc score interpolation can only reorder $\C_f^{100}(q)$ for a fixed $f$. Hybrid candidate generation changes the feasible set from $\C_{\text{BM25}}^{100}(q)$ to $\C_{\text{RRF}}^{100}(q)$. The gain is consistent with a feasible-set expansion:
 \begin{equation}
   \G_q\cap\C_{\text{RRF}}^{100}(q)
   \supsetneq
@@ -38,4 +41,4 @@ Formally, post-hoc score interpolation can only reorder $\C_f^{100}(q)$ for a fi
 \end{equation}
 
 \paragraph{Practical recommendation.}
-For late-interaction retrieval systems with a fixed reranking budget, we recommend treating BM25 as a complementary candidate generator. A simple lexical+semantic RRF first stage is easy to implement, requires no score calibration, and improved every dataset in this study after ModernColBERT reranking. If compute is constrained, the most valuable measurement is Recall@100 of the first-stage pool: low BM25 Recall@100 indicates that BM25-only reranking is likely leaving quality on the table.
+For late-interaction retrieval systems with a fixed reranking budget, BM25 is best treated as a complementary candidate generator. A lexical+semantic RRF first stage is simple, requires no score calibration, and improved every dataset in this study after ModernColBERT reranking. If compute is constrained, the most useful diagnostic is Recall@100 of the first-stage pool. Low BM25 Recall@100 is a warning sign that BM25-only reranking is leaving quality on the table.

paper/sections/appendix.tex

@@ -2,7 +2,7 @@
 
 \subsection{Artifact Layout}
 
-The reproducibility artifact is organized around a canonical configuration and runner:
+The reproducibility artifact is organized around one canonical configuration and runner:
 \begin{itemize}[leftmargin=*]
   \item \texttt{experiments/configs/paper\_main.yaml}: dataset list, model names, retrieval parameters, and output directories.
   \item \texttt{experiments/scripts/prepare\_paper\_data.py}: downloads standard BEIR datasets and exports CQADupStack domains into BEIR format.
@@ -13,11 +13,11 @@ The reproducibility artifact is organized around a canonical configuration and r
 
 \subsection{Full Main Table}
 
-\begin{table*}[t]
+\begin{center}
 \centering
 \scriptsize
-\setlength{\tabcolsep}{3.5pt}
-\begin{tabular}{lrrrrr}
+\setlength{\tabcolsep}{0pt}
+\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}lrrrrr}
 \toprule
 Dataset & BM25 & Dense & RRF & BM25 rerank & RRF rerank \\
 \midrule
@@ -31,13 +31,13 @@ CQADupStack Mathematica & 0.1793 & 0.2685 & 0.2443 & 0.2694 & 0.2963 \\
 \midrule
 Average & 0.2956 & 0.3831 & 0.3606 & 0.3826 & 0.4211 \\
 \bottomrule
-\end{tabular}
-\caption{Full nDCG@10 table for the seven complete datasets.}
+\end{tabular*}
+\captionof{table}{Full nDCG@10 table for the seven complete datasets.}
 \label{tab:full-ndcg}
-\end{table*}
+\end{center}
 
 \subsection{Main Commands}
 
-The core experiment can be reproduced with \texttt{uv sync}, followed by \texttt{prepare\_paper\_data.py} and \texttt{run\_paper\_main.py} using \texttt{experiments/configs/paper\_main.yaml}. The exact shell commands are included in the artifact command sheet.
+The core experiment can be reproduced with \texttt{uv sync}, followed by \texttt{prepare\_paper\_data.py} and \texttt{run\_paper\_main.py} using \texttt{experiments/configs/paper\_main.yaml}. The artifact command sheet contains the exact shell commands.
 
 The runner wraps BM25 retrieval, semantic retrieval, first-stage RRF fusion, ModernColBERT reranking over BM25 candidates, ModernColBERT reranking over RRF candidates, feature extraction, and table compilation.

paper/sections/conclusion.tex

@@ -1 +1 @@
-We studied when lexical search helps late-interaction retrieval. Across seven complete datasets, the answer is clearest in a two-stage pipeline: BM25 helps as part of hybrid candidate generation. Lexical+semantic RRF increases Recall@100 over BM25 alone, and ModernColBERT reranking over that hybrid candidate pool improves nDCG@10 on every dataset. The result suggests a simple design rule. Use semantic retrieval for semantic coverage, BM25 for exact-match coverage, fuse them with a robust rank-based method such as RRF, and let a late-interaction model perform the final precision ranking.
+We studied when lexical search helps late-interaction retrieval. Across seven complete datasets, the clearest answer appears in the two-stage pipeline: BM25 helps as part of hybrid candidate generation. Lexical+semantic RRF increases Recall@100 over BM25 alone, and ModernColBERT reranking over that hybrid pool improves nDCG@10 on every dataset. The design rule is simple. Use semantic retrieval for semantic coverage, use BM25 for exact-match coverage, fuse the candidate lists with a robust rank-based method such as RRF, and let the late-interaction model do the final precision ranking.

paper/sections/experiments.tex

@@ -1,11 +1,11 @@
 \paragraph{Datasets.}
-We evaluate on seven complete datasets: SciFact, NFCorpus, ArguAna, SciDocs, FiQA, CQADupStack Android, and CQADupStack Mathematica. These cover scientific fact verification, biomedical/scientific retrieval, argument retrieval, scientific document retrieval, financial question answering, and technical duplicate-question retrieval. Partial BM25-rerank runs on CQADupStack English and Gaming are excluded from the main average because semantic and hybrid first-stage runs were not completed.
+We evaluate on seven complete datasets: SciFact, NFCorpus, ArguAna, SciDocs, FiQA, CQADupStack Android, and CQADupStack Mathematica. The collection spans scientific fact verification, biomedical and scientific retrieval, argument retrieval, financial question answering, and technical duplicate-question retrieval. Partial BM25-rerank runs on CQADupStack English and Gaming are left out of the main average because the matching semantic and hybrid first-stage runs were not completed.
 
 \paragraph{Systems.}
-We compare five systems: BM25; semantic retrieval; lexical+semantic RRF; ModernColBERT reranking of BM25 top-100; and ModernColBERT reranking of lexical+semantic RRF top-100. The last system is the proposed pipeline. All first-stage systems use $k=100$ for fair reranking.
+We compare five systems: BM25; semantic retrieval; lexical+semantic RRF; ModernColBERT reranking of BM25 top-100; and ModernColBERT reranking of lexical+semantic RRF top-100. The last system is the proposed pipeline. All reranking runs use the same first-stage budget, $k=100$.
 
 \paragraph{Implementation.}
-BM25 uses $k_1=0.9,b=0.4$. The semantic retriever is BGE-M3 with normalized embeddings and maximum sequence length 512. The late-interaction reranker is \texttt{lightonai/GTE-ModernColBERT-v1}. Documents are split into 180-word chunks with stride 120; each chunk is encoded with document length 300 and queries with length 48. We aggregate chunk scores by max pooling. Experiments ran on a MacBook Air with Apple MPS acceleration when available; non-finite MPS embeddings were repaired by CPU re-encoding before scoring.
+BM25 uses $k_1=0.9,b=0.4$. The semantic retriever uses normalized BGE-M3 embeddings with maximum sequence length 512. The late-interaction reranker is \texttt{lightonai/GTE-ModernColBERT-v1}. We split documents into 180-word chunks with stride 120, encode chunks with document length 300, and encode queries with length 48. Document scores are the maximum over chunk scores. The experiments ran on a MacBook Air with Apple MPS acceleration when available; when MPS produced non-finite embeddings, those batches were re-encoded on CPU before scoring.
 
 \paragraph{Metrics.}
 For query $q$, ranked list $R_q=(d_1,\ldots,d_k)$, and gain $\rel(q,d_i)$, we compute
@@ -24,7 +24,7 @@ Candidate recall is
   \sum_{q\in\Q}
   \frac{|\C_f^k(q)\cap \G_q|}{|\G_q|}.
 \end{equation}
-We report nDCG@10 as the primary quality metric and Recall@100 as the candidate-pool diagnostic. Reranking can improve nDCG@10 but cannot increase Recall@100 beyond its first-stage candidate set.
+We report nDCG@10 as the primary quality metric and Recall@100 as the candidate-pool diagnostic. Reranking can improve nDCG@10, but it cannot increase Recall@100 beyond the first-stage candidate set it receives.
 
 \paragraph{Reproducibility.}
-The artifact contains a canonical configuration, dataset preparation script, end-to-end runner, table compiler, and generated TREC run files. The main reproduction entry point is \texttt{run\_paper\_main.py} with \texttt{experiments/configs/paper\_main.yaml}; Appendix~\ref{app:repro} gives details.
+The artifact contains the canonical configuration, dataset preparation script, end-to-end runner, table compiler, and generated TREC run files. The main reproduction entry point is \texttt{run\_paper\_main.py} with \texttt{experiments/configs/paper\_main.yaml}; Appendix~\ref{app:repro} gives the command-level details.

paper/sections/introduction.tex

@@ -1,10 +1,10 @@
-Late-interaction retrieval models, introduced by ColBERT-style architectures, represent queries and documents as sets of contextual token vectors and score a document by aggregating maximum token similarities \cite{khattab2020colbert,santhanam2022colbertv2}. This design occupies a useful middle ground: it is more expressive than a single-vector dense retriever, but substantially cheaper at retrieval time than a cross-encoder reranker. As a result, late-interaction models are increasingly attractive for retrieval-augmented generation, search, and domain-specific question answering.
+Late-interaction retrieval models, introduced by ColBERT-style architectures, represent queries and documents as sets of contextual token vectors and score a document by aggregating maximum token similarities \cite{khattab2020colbert,santhanam2022colbertv2}. This gives them a useful tradeoff: they keep token-level matching, but they are much cheaper at retrieval time than cross-encoder rerankers. That makes late-interaction models a good fit for retrieval-augmented generation, search, and domain-specific question answering.
 
-At the same time, lexical retrieval remains stubbornly useful. BM25 is simple, fast, interpretable, and especially strong when relevance depends on exact terminology, identifiers, abbreviations, entities, or rare domain words \cite{robertson2009bm25}. Many production systems therefore combine lexical and neural retrieval, often with reciprocal rank fusion (RRF) or a learned reranking stage \cite{cormack2009rrf}. Most empirical work on hybrid retrieval, however, studies BM25 combined with single-vector dense embeddings. The late-interaction setting is less clear. Because late-interaction models already perform token-level matching, one might expect BM25 to add little beyond implementation complexity. Alternatively, lexical retrieval may still recover relevant documents that semantic models miss, improving the candidate pool available to a reranker.
+Lexical retrieval remains hard to discard. BM25 is fast, transparent, and strong when relevance depends on exact terminology, identifiers, abbreviations, entities, or rare domain words \cite{robertson2009bm25}. Many deployed systems therefore combine lexical and neural retrieval, often with reciprocal rank fusion (RRF) or a reranking stage \cite{cormack2009rrf}. Most hybrid-retrieval studies pair BM25 with single-vector semantic embeddings. The late-interaction case is less obvious. Since the final scorer already performs token-level matching, BM25 might be redundant. On the other hand, BM25 may still recover relevant documents that semantic retrieval misses, giving the reranker a better candidate set.
 
-This paper studies that question in a controlled two-stage retrieval setting. We compare BM25, semantic retrieval, lexical+semantic RRF, and ModernColBERT reranking over first-stage candidates. In our implementation, the semantic retriever is BGE-M3 \cite{chen2024bge}. Rather than asking only which standalone retriever has the highest nDCG@10, we ask which first-stage candidate pool produces the best final ranking after late-interaction reranking. This distinction matters: a reranker cannot recover documents that never appear in its candidate set, so first-stage Recall@100 can be as important as top-10 first-stage ranking quality.
+We study that question in a controlled two-stage setup. The first stage is BM25, semantic retrieval, or lexical+semantic RRF. The second stage is ModernColBERT reranking over the first-stage candidates. In our implementation, the semantic retriever is BGE-M3 \cite{chen2024bge}. We are not only asking which standalone retriever has the best nDCG@10. We are asking which candidate pool gives the late-interaction reranker the best chance to produce a good final ranking. This matters because a reranker cannot recover a document that never enters its candidate set.
 
-Our experiments on seven complete BEIR-style datasets show a consistent pattern. Hybrid lexical+semantic candidate generation increases Recall@100 over BM25 alone, and ModernColBERT reranking over the hybrid candidate pool improves nDCG@10 on every dataset. The average gain is substantial for a retrieval-system ablation: 0.4211 nDCG@10 for hybrid-candidate reranking versus 0.3826 for BM25-candidate reranking. The result reframes the role of lexical search. BM25 is not always the best standalone ranker, but it is a useful complement to semantic retrieval when constructing candidates for a token-level reranker.
+Across seven BEIR-style datasets, the pattern is consistent. Hybrid lexical+semantic retrieval has higher Recall@100 than BM25 alone, and ModernColBERT reranking over the hybrid pool improves nDCG@10 on every dataset. The average nDCG@10 is 0.4211 for hybrid-candidate reranking and 0.3826 for BM25-candidate reranking. The role of BM25 is therefore narrower than ``best ranker'' but still important: it supplies candidate-pool coverage that the semantic retriever misses.
 
 The paper makes three contributions:
 \begin{enumerate}[leftmargin=*]

paper/sections/limitations.tex

@@ -1 +1 @@
-This study is intentionally narrow. We evaluate one semantic retriever, one late-interaction reranker, and a fixed top-100 candidate budget. Stronger semantic retrievers, larger candidate pools, or learned fusion methods may change the magnitude of the gains. We also report quality metrics rather than general latency claims, because the experiments were run on a MacBook Air with Apple MPS acceleration and CPU fallback for non-finite embeddings. Finally, the analysis uses public benchmark datasets; production traffic may have different query distributions, document lengths, freshness requirements, and latency constraints.
+This study is intentionally narrow. It uses one semantic retriever, one late-interaction reranker, and a fixed top-100 candidate budget. Stronger semantic retrievers, larger candidate pools, or learned fusion methods may change the size of the gains. We also avoid broad latency claims, because the experiments ran on a MacBook Air with Apple MPS acceleration and CPU fallback for non-finite embeddings. The datasets are public benchmarks; production traffic may have different query distributions, document lengths, freshness requirements, and latency constraints.

paper/sections/method.tex

@@ -1,8 +1,8 @@
-Let $\C=\{d_1,\ldots,d_N\}$ be a corpus, $\Q$ a query set, and $\G_q=\{d:\rel(q,d)>0\}$ the relevant set for query $q$. A first-stage retriever $f$ induces a scoring function $s_f(q,d)$, a ranked list $R_f(q)$, and a candidate set
+Let $\C=\{d_1,\ldots,d_N\}$ be a corpus, $\Q$ a query set, and $\G_q=\{d:\rel(q,d)>0\}$ the relevant set for query $q$. A first-stage retriever $f$ defines a score $s_f(q,d)$, a ranked list $R_f(q)$, and a candidate set
 \begin{equation}
   \C_f^k(q)=\Top_k\{s_f(q,d):d\in\C\}.
 \end{equation}
-The reranker receives only $\C_f^k(q)$; therefore $\C_f^k(q)$ defines the maximum achievable recall of any downstream model.
+The reranker only sees $\C_f^k(q)$. This set therefore fixes the recall ceiling for every downstream model, no matter how strong the final scorer is.
 
 \paragraph{Lexical and semantic first stages.}
 The lexical retriever is BM25:
@@ -13,12 +13,12 @@ s_{\text{BM25}}(q,d)=
 \frac{tf_{t,d}(k_1+1)}
 {tf_{t,d}+k_1(1-b+b|d|/\overline{|d|})}.
 \end{equation}
-The semantic retriever embeds $q$ and $d$ into normalized vectors $e_q,e_d$ and scores by inner product:
+The semantic retriever embeds $q$ and $d$ as normalized vectors $e_q,e_d$ and scores their inner product:
 \begin{equation}
   s_{\text{sem}}(q,d)= e_q^\top e_d,\qquad
   \|e_q\|_2=\|e_d\|_2=1 .
 \end{equation}
-In our implementation the semantic model is BGE-M3, but the experimental role is model-agnostic: it supplies candidates not constrained by exact term overlap.
+Our implementation uses BGE-M3 for this stage. In the argument of the paper, however, the semantic retriever mainly plays a structural role: it supplies candidates that are not limited to exact term overlap.
 
 \paragraph{Hybrid candidate generation.}
 We fuse lexical and semantic rankings with reciprocal rank fusion:
@@ -27,15 +27,15 @@ We fuse lexical and semantic rankings with reciprocal rank fusion:
   \sum_{r\in\{\text{lex},\text{sem}\}}
   \frac{\mathbf{1}[d\in R_r(q)]}{c+\rank_r(q,d)} .
 \end{equation}
-The main hybrid candidate set is $\C_{\text{RRF}}^{100}(q)$. RRF is attractive here because BM25 and semantic scores are not calibrated, while ranks are directly comparable.
+The hybrid candidate set is $\C_{\text{RRF}}^{100}(q)$. RRF is useful here for a mundane reason: BM25 scores and semantic inner products are not calibrated, but their ranks can still be combined without fitting a calibration model.
 
 \paragraph{Late-interaction reranking.}
-ModernColBERT maps a query to token vectors $Q=\{u_i\}_{i=1}^{m}$ and a document chunk to token vectors $D=\{v_j\}_{j=1}^{n}$. The exact MaxSim score is
+ModernColBERT maps a query to token vectors $Q=\{u_i\}_{i=1}^{m}$ and a document chunk to token vectors $D=\{v_j\}_{j=1}^{n}$. The MaxSim score is
 \begin{equation}
   s_{\text{LI}}(q,d)=
   \sum_{i=1}^{m} \max_{1\le j\le n} u_i^\top v_j .
 \end{equation}
-For long documents, we score chunks independently and aggregate at document level:
+For long documents, we score chunks independently and aggregate at the document level:
 \begin{equation}
   s_{\text{doc}}(q,d)=
   \max_{z\in \operatorname{chunks}(d)} s_{\text{LI}}(q,z).
@@ -49,7 +49,7 @@ The final two-stage ranking for first-stage retriever $f$ is
 \end{equation}
 
 \paragraph{Hypothesis.}
-The candidate-recall bottleneck predicts that lexical search helps if it increases the hybrid candidate set seen by the late-interaction scorer:
+The candidate-recall bottleneck gives a concrete prediction. Lexical search should help the late-interaction reranker when it increases the relevant-document mass inside the top-100 candidate pool:
 \begin{align}
   \CR_{100}(\text{RRF}) &> \CR_{100}(\text{BM25})
   \nonumber\\[-1mm]
@@ -59,4 +59,4 @@ The candidate-recall bottleneck predicts that lexical search helps if it increas
   &\hspace{1.7em}>
   \operatorname{nDCG}_{10}(R_{\text{LI}\circ\text{BM25}}).
 \end{align}
-Thus the paper tests a candidate-generation claim, not merely a score-fusion claim.
+The experiment therefore tests candidate generation. It is not only a comparison of score-fusion recipes.

paper/sections/related_work.tex

@@ -1,17 +1,17 @@
 \paragraph{Lexical retrieval.}
-BM25 remains the standard lexical baseline for ad hoc retrieval \cite{robertson2009bm25}. Its strength comes from exact term matching with document-length normalization and term-frequency saturation. This makes it particularly competitive for entity-heavy and terminology-heavy queries. Its weakness is equally clear: it cannot retrieve documents that express the same information with different vocabulary.
+BM25 remains the standard lexical baseline for ad hoc retrieval \cite{robertson2009bm25}. It works well when the query and relevant documents share terminology, and it is especially useful for entities, identifiers, abbreviations, and rare domain words. Its limitation is direct: if the relevant document uses different wording, BM25 has little signal.
 
 \paragraph{Dense and late-interaction retrieval.}
-Semantic retrievers encode queries and documents into continuous representations, enabling matching beyond exact token overlap. DPR demonstrated that learned dense representations can outperform strong BM25 baselines for open-domain QA candidate retrieval \cite{karpukhin2020dpr}. BGE-M3 is a recent embedding model designed for multiple retrieval modes, including dense, sparse, and multi-vector use cases \cite{chen2024bge}. Late-interaction models such as ColBERT retain multiple token vectors per query and document, scoring with MaxSim-style token matching \cite{khattab2020colbert}. ColBERTv2 improves the efficiency and quality of this family through denoised supervision and compression \cite{santhanam2022colbertv2}. In this paper, we use \texttt{lightonai/GTE-ModernColBERT-v1}, an open ModernBERT-based ColBERT-style model trained for PyLate-style multi-vector retrieval \cite{lighton2025gte}.
+Semantic retrievers encode queries and documents into continuous representations, which lets them match beyond exact token overlap. DPR showed that learned dense representations can outperform strong BM25 baselines for open-domain QA candidate retrieval \cite{karpukhin2020dpr}. BGE-M3 is a recent embedding model with dense, sparse, and multi-vector retrieval modes \cite{chen2024bge}. Late-interaction models such as ColBERT keep multiple token vectors per query and document and score with MaxSim-style matching \cite{khattab2020colbert}. ColBERTv2 improves this family with denoised supervision and compression \cite{santhanam2022colbertv2}. We use \texttt{lightonai/GTE-ModernColBERT-v1}, an open ModernBERT-based ColBERT-style model trained for PyLate-style multi-vector retrieval \cite{lighton2025gte}.
 
 \paragraph{Hybrid retrieval and fusion.}
-Hybrid retrieval combines lexical and neural evidence. A common approach is reciprocal rank fusion, which combines ranked lists using only ranks rather than calibrated scores \cite{cormack2009rrf}. This is attractive when BM25 and semantic scores live on different scales. Recent work studies complementarity objectives that explicitly encourage sparse and dense retrievers to retrieve different relevant evidence \cite{lee2023complementarity}, and hybrid hierarchical retrieval shows that sparse+dense combinations can improve robustness in open-domain QA pipelines \cite{arivazhagan2023hhr}. These works establish the value of hybrid retrieval as a first-stage retrieval strategy.
+Hybrid retrieval combines lexical and neural evidence. Reciprocal rank fusion is a common choice because it uses ranks rather than calibrated scores \cite{cormack2009rrf}. That is useful when BM25 and semantic scores are not comparable. Recent work studies complementarity objectives that encourage sparse and dense retrievers to retrieve different relevant evidence \cite{lee2023complementarity}. Hybrid hierarchical retrieval also shows that sparse+dense combinations can improve open-domain QA pipelines \cite{arivazhagan2023hhr}. These papers establish hybrid retrieval as a strong first-stage strategy.
 
 \paragraph{Hybrid retrieval with reranking.}
-Several studies combine sparse/dense retrieval with reranking. HYRR selects reranker training data from hybrid retrievers and shows improved robustness across first-stage retrievers on MS MARCO and BEIR \cite{lu2024hyrr}. HybRank incorporates lexical and semantic properties from upstream retrievers into a collaborative reranker \cite{zhang2023hybrank}. Domain-specific systems similarly find that hybrid retrieval followed by cross-encoder reranking performs well in scientific, biomedical, conversational, and text-and-table settings \cite{fathallah2025climatecheck,ajith2024litsearch,akarsu2026texttable}. Our study is closest in system shape to these retrieve-then-rerank pipelines, but differs in the final scorer and diagnosis: we use a late-interaction reranker rather than a cross-encoder or LLM reranker, hold the reranker fixed, and isolate how the first-stage candidate set changes the final ranking.
+Several studies combine sparse/dense retrieval with reranking. HYRR selects reranker training data from hybrid retrievers and improves robustness across first-stage retrievers on MS MARCO and BEIR \cite{lu2024hyrr}. HybRank incorporates lexical and semantic signals from upstream retrievers into a collaborative reranker \cite{zhang2023hybrank}. Domain systems also find that hybrid retrieval followed by cross-encoder reranking works well in scientific, biomedical, conversational, and text-and-table settings \cite{fathallah2025climatecheck,ajith2024litsearch,akarsu2026texttable}. Our system has the same retrieve-then-rerank shape, but the question is different. We use a late-interaction reranker, hold it fixed, and vary only the candidate generator.
 
 \paragraph{Benchmarks.}
-BEIR provides a heterogeneous benchmark suite for zero-shot retrieval evaluation across domains and query styles \cite{thakur2021beir}. We use BEIR-style datasets because the central question is comparative: lexical search may help differently in scientific fact verification, argument retrieval, financial question answering, and technical duplicate-question retrieval.
+BEIR provides heterogeneous retrieval benchmarks across domains and query styles \cite{thakur2021beir}. We use BEIR-style datasets because the value of lexical search should vary with the task: scientific fact verification, argument retrieval, financial question answering, and technical duplicate-question retrieval do not stress the same retrieval behavior.
 
 \paragraph{Same and different from prior hybrid pipelines.}
-The present paper is similar to prior hybrid+reranking work in three respects: it uses BM25 as the lexical channel, uses a semantic retriever as a complementary first stage, and evaluates a retrieve-then-rerank architecture. It differs in the object of study. HYRR and HybRank modify or train rerankers using hybrid evidence; domain RAG systems often evaluate end-to-end generation or cross-encoder rerankers. We instead hold the late-interaction reranker fixed and vary only the candidate generator. The contribution is therefore diagnostic rather than architectural: we show that lexical search still matters even when the final scorer is token-level semantic interaction, and that the measurable pathway is $\CR_{100}$ expansion followed by higher nDCG@10 after reranking.
+This paper overlaps with prior hybrid+reranking work in its ingredients: BM25, a semantic retriever, RRF, and a retrieve-then-rerank pipeline. It differs in what it tries to isolate. HYRR and HybRank modify or train rerankers using hybrid evidence; domain RAG systems often evaluate answer quality or cross-encoder reranking. We hold the late-interaction reranker fixed and change only the candidate generator. The contribution is diagnostic: lexical search still matters when the final scorer is already token-level semantic interaction, and the measurable pathway is $\CR_{100}$ expansion followed by higher nDCG@10.

paper/sections/results.tex

@@ -1,4 +1,4 @@
-Table~\ref{tab:main} reports the complete-seven-dataset average. The first three rows compare first-stage candidate generators directly. The last two rows evaluate the late-interaction reranker over BM25 candidates and over hybrid RRF candidates.
+Table~\ref{tab:main} reports the average over the seven complete datasets. The first three rows compare first-stage candidate generators. The last two rows keep the reranker fixed and change only the candidate pool: BM25 top-100 versus hybrid RRF top-100.
 
 \begin{table*}[t]
 \centering
@@ -18,9 +18,9 @@ ModernColBERT rerank of RRF top-100 & \textbf{0.4211} & \textbf{0.6743} \\
 \label{tab:main}
 \end{table*}
 
-The main comparison is between the two reranked systems. Reranking BM25 candidates reaches 0.3826 average nDCG@10. Reranking hybrid RRF candidates reaches 0.4211, an absolute gain of +0.0385. This improvement occurs on all seven complete datasets. Because both systems use the same late-interaction reranker and the same candidate budget, the difference is attributable to candidate-pool construction rather than to a stronger final scorer.
+The main comparison is between the two reranked systems. Reranking BM25 candidates reaches 0.3826 average nDCG@10. Reranking hybrid RRF candidates reaches 0.4211, an absolute gain of +0.0385. The gain appears on all seven complete datasets. Since both systems use the same late-interaction reranker and the same candidate budget, the difference comes from the candidates supplied to the reranker, not from a different final scorer.
 
-The first-stage results explain the mechanism. BM25 has average Recall@100 of 0.5262. Semantic retrieval increases this to 0.6738, and lexical+semantic RRF reaches 0.6743. Thus, the hybrid first stage gives the reranker many more opportunities to place relevant documents in the top 10. Even though semantic retrieval has higher average first-stage nDCG@10 than RRF, RRF slightly improves Recall@100 and includes lexical matches that semantic retrieval may miss. For a reranking pipeline, that recall-oriented property is valuable.
+The first-stage results explain the mechanism. BM25 has average Recall@100 of 0.5262. Semantic retrieval raises this to 0.6738, and lexical+semantic RRF reaches 0.6743. The hybrid first stage therefore gives the reranker more chances to place a relevant document in the final top 10. Semantic retrieval has higher average first-stage nDCG@10 than RRF, but RRF slightly improves Recall@100 and preserves lexical matches that the semantic retriever may miss. In a reranking pipeline, that recall-oriented behavior matters more than the first-stage nDCG ordering by itself.
 
 \begin{table*}[t]
 \centering
@@ -42,4 +42,4 @@ CQADupStack Mathematica & 0.2694 & 0.2963 & +0.0269 \\
 \label{tab:rerank}
 \end{table*}
 
-Table~\ref{tab:rerank} shows the per-dataset reranking gains. The largest improvement is on FiQA (+0.0820), followed by ArguAna (+0.0526) and CQADupStack Android (+0.0440). The smallest gain is on SciDocs (+0.0088), where all systems have relatively low recall and the hybrid pool offers less additional useful evidence. The consistency of the sign, however, is notable: hybrid candidate generation never hurts the late-interaction reranker in this study.
+Table~\ref{tab:rerank} shows the per-dataset reranking gains. The largest improvement is on FiQA (+0.0820), followed by ArguAna (+0.0526) and CQADupStack Android (+0.0440). The smallest gain is on SciDocs (+0.0088), where all systems have relatively low recall and the hybrid pool adds less useful evidence. The sign is still consistent: hybrid candidate generation does not hurt the late-interaction reranker on any complete dataset in this study.