Listwise Cross-Encoder Fine-Tuning vs. Agentic Instruction Tuning for LLM Rerankers: A Systematic Study in Medical Procedure Reranking

Matan Fainzilber; Shlomit Plavner

REVIEW 3 major objections 4 minor 11 references

Listwise Cross-Encoder Fine-Tuning vs. Agentic Instruction Tuning for LLM Rerankers: A Systematic Study in Medical Procedure Reranking

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Fine-tuning a 109M-parameter biomedical cross-encoder with the listwise ListNet loss yields better medical-procedure reranking than a 4B-parameter instruction model whose prompt was optimized by an agentic loop.

desk verdict A well-run empirical study whose headline margin is real on its own benchmark but teacher-circular: the 109M model's edge over the 4B reranker measures agreement with GPT-4o until an independent test set says otherwise. read the letter →

arxiv 2608.09650 v1 pith:W3R3NVMJ submitted 2026-08-10 cs.IR cs.CL

Matan Fainzilber , Shlomit Plavner This is my paper · ORCID

classification cs.IRcs.CL

keywords listwiselearningtorankcross-encoderrerankingmedicalprocedureretrievalagenticpromptoptimizationLLMdatasetsynthesisNDCGhealthinsurancesearchinstruction-tunedreranker

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to determine whether a small, domain-specialized reranker can beat a much larger instruction-tuned language model at matching patient queries to billable medical procedures. On a new dataset of 2,647 query-procedure lists built with an LLM synthesis pipeline, the best configuration, a 109M-parameter MedCPT cross-encoder fine-tuned with the listwise ListNet loss, reaches $0.961$ test $\mathrm{NDCG@3}$, ahead of a prompt-optimized 4B-parameter Qwen3 reranker by $2.6$ points on $\mathrm{NDCG@3}$ and $13.3$ points on Spearman correlation, at $37\times$ fewer parameters. The paper also introduces a two-phase LLM-based dataset construction pipeline and an agentic prompt-optimization loop that improves an instruction-tuned reranker's prompt from validation feedback. If the central claim is right, production reranking decisions under latency and cost constraints can favor small fine-tuned encoders over much larger generalist models.

What carries the argument

The mechanism that carries the argument is listwise supervision over full candidate lists, applied to a cross-encoder that already carries biomedical domain knowledge. The concrete identity is ListNet, which minimizes the cross-entropy between the softmax of ground-truth relevance labels and the softmax of the model's logits over the list; unlike pointwise or pairwise losses, it encodes the whole ordering into the training signal. The experiment grid contrasts this with two other listwise objectives (LambdaLoss and PListMLE), with layer-freezing regimes, and with an agentic prompt-optimization loop that rewrites the instruction of Qwen3-Reranker-4B from validation-set $\mathrm{NDCG@3}$ feedback. What the comparison isolates is the contribution of learning from domain-specific full-list signal versus adapting a large model through prompt text alone.

What would settle it

Take an independently labeled sample of roughly 300 query-procedure lists with human expert relevance judgments, evaluate the best MedCPT+ListNet model and the prompt-optimized Qwen3 reranker on those labels, and check whether the 2.6-point $\mathrm{NDCG@3}$ and 13.3-point Spearman gaps persist; if they vanish or reverse, the reported advantage is an artifact of training and evaluating on the same teacher's rankings.

Watch

Extended reading notes

Core claim

The paper's central claim is that listwise fine-tuning of a small domain-pretrained cross-encoder produces a reranker that beats an agentically prompt-optimized 4B-parameter instruction reranker on this medical procedure task. The strongest result is MedCPT plus ListNet with no frozen layers: $0.961 \pm 0.002$ test $\mathrm{NDCG@3}$, a $+4.6$ point gain over the unfine-tuned MedCPT baseline and a $+19.4$ point gain on Spearman. Against Qwen3-Reranker-4B with its optimized prompt, this configuration leads by $+2.6$ points on $\mathrm{NDCG@3}$, $+1.6$ points on $\mathrm{NDCG@1}$, and $+13.3$ points on Spearman. The authors conclude that prompt-only adaptation of a large generalist reranker cannot substitute for listwise fine-tuning on domain data, and that full-list ordering is what small listwise-trained models improve most.

Load-bearing premise

The load-bearing premise is that the GPT-4o-produced relevance rankings are valid ground truth for clinical relevance; the paper's own limitation section notes that any systematic bias in the teacher is inherited by the training labels, and only 100 of 2,647 ranked lists were checked by human experts.

Editorial extensions

If this is right

Deploying a 109M-parameter reranker on commodity CPU hardware can match or beat a 4B-parameter GPU-hosted model on this medical insurance task, changing the cost profile of real-time reranking.
Listwise training drives the win: the best fine-tuned model gains $+4.6$ points over its unfine-tuned base on $\mathrm{NDCG@3}$ and $+19.4$ points on Spearman, so teams with domain data should rank lists rather than pairs.
The agentic prompt-optimization loop improves Qwen3 from $0.936$ to $0.948$ validation $\mathrm{NDCG@3}$ in three iterations, making it a usable baseline-setting method even though it does not close the gap to fine-tuning.
The two-phase LLM dataset pipeline (query generation, teacher ranking, top-3 quality gate) is offered as a scalable route to domain-specific reranking datasets when expert annotation is scarce.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A fair stress test would re-score a random held-out sample with independent human judgments rather than the teacher model; if the 2.6-point and 13.3-point gaps persist, the ranking advantage is robust, and if they shrink, the teacher's top-3 filter is doing unseen work.
Because the paper's task has the same lexical structure as other lay-to-formal matching problems, the same listwise fine-tuning recipe is a plausible default for legal-benefit search, government form lookup, and product-category routing.
The unexplored reverse freezing strategy, freezing upper layers while training embeddings, is a natural next experiment; the paper notes it may matter for lexical-gap tasks, since adapting token representations rather than reasoning layers could be the dominant effect.
The $37\times$ parameter gap suggests an operational extension the paper does not calculate: the saved inference budget could be spent on retrieving and reranking longer candidate lists, which tends to improve recall and might further widen the gap.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things up front. The empirical work is more careful than most: an 18-cell grid (2 backbones × 3 listwise losses × 3 freeze strategies), Optuna tuning, five seeds with tight std, one held-out test evaluation. And the headline comparison is benchmark-circular exactly as the stress-test note says: GPT-4o wrote the queries, produced the relevance rankings, and the quality gate kept only rows where the teacher put the target in the top 3. The cross-encoders train on that teacher's softmax labels and are evaluated against them; the Qwen baseline is only prompted. So the 2.6pp NDCG@3 and 13.3pp Spearman margins measure agreement with GPT-4o on teacher-selected easy rows, not clinical relevance. The limitations section admits teacher-bias inheritance, but the admission doesn't unhitch the headline claim.

What the paper does well: the two-phase dataset pipeline is clearly specified with pseudocode and a code/sample release; the findings that all 18 fine-tuned configurations beat their base models, that freezing six layers costs about 1.6pp while freezing embeddings costs little, and that listwise training gives double-digit Spearman lifts are credible, internally consistent results that would likely survive in direction on independent labels even if the absolute numbers shift. The agentic prompt-optimization loop is clearly described.

Soft spots, in proportion. The teacher-circularity is load-bearing and real. The 100-row human validation covers 3.8% of the data and checks only top-1/top-3 acceptance — a sanity anchor for the top, but silent on the full-list ordering that drives the Spearman gap. The metric band is saturated too: with the target always in the teacher's top 3, NDCG@3 sits between 0.92 and 0.96 for everyone, so a 2.6pp margin is compressed. One correction to the reader's write-up: the initial-prompt test score is missing from the prose, but Figure 5 appears to show it — Test: 0.936, the same as the optimized prompt's test score. The agentic loop was thus a validation-only gain (0.936→0.948) with zero test transfer, a buried negative result the authors should report; it weakens their 'strong LLM baseline' framing but supports their small-model argument.

Who this is for: production IR teams deciding between distilling a teacher into a small cross-encoder and prompting a large reranker, plus anyone building LLM-synthetic benchmarks. The narrow claim — a 109M model fine-tuned on these labels beats a 4B model prompted on these labels — holds on the paper's own terms. The production-grade conclusion needs an independently labeled test set, even a modest one, before it counts as evidence about clinical relevance.

Recommendation: send it to peer review. It is a serious, competently executed study; the referee should push hard for external test labels and for the unoptimized baseline's test score. With those, it becomes a genuinely useful paper.

Referee Report

3 major / 4 minor

Summary. The paper presents a systematic comparison of two reranking paradigms for medical procedure reranking: small cross-encoders (MedCPT, MiniLM-L12) fine-tuned with listwise losses (LambdaLoss, ListNet, PListMLE) under three layer-freezing strategies, and a 4B-parameter Qwen3-Reranker whose instruction is refined by an agentic GPT-4.1 optimization loop. Using a purpose-built synthetic dataset of 2,647 query-procedure lists generated and ranked by GPT-4o with quality filtering, the authors report that a 109M-parameter MedCPT fine-tuned with ListNet and no freezing achieves 0.961 test NDCG@3, outperforming the prompt-optimized Qwen3 model by 2.6pp on NDCG@3 and 13.3pp on Spearman, at 37x fewer parameters. The paper also contributes a dataset-construction pipeline, an agentic prompt optimization framework, and practical deployment takeaways.

Significance. If the reported result holds as a claim about clinical relevance, it is practically important: it would show that a small, cheap, CPU-servable cross-encoder can beat a 4B-parameter instruction-tuned reranker on a domain-specific medical query-to-procedure alignment task. The study design has genuine strengths: an 18-cell experimental grid, 5-seed training with tight standard deviations, a validation-based model selection protocol, a credible agentic prompt optimization baseline, and a commitment to release code and a sample dataset. These are exactly the ingredients for reproducible applied IR research. However, the evaluation is built entirely on GPT-4o-generated relevance labels, and the fine-tuned models are trained to imitate that same teacher. Consequently, the current evidence establishes agreement with a teacher model on teacher-selected easy cases, not superiority for true clinical relevance. The significance of the paper's central claim is therefore contingent on the validity of the synthetic labels as ground truth; the manuscript's own limitations and the small human validation sample do not currently resolve this dependency.

major comments (3)

[§3.2, §4.2, Table 2] The central comparison is circular. In Phase 2 (§3.2), GPT-4o produces the relevance ranking π_i for each query, and the quality filter retains only rows where GPT-4o placed the target procedure in the top 3. The ListNet loss in Eq. (4) trains the small cross-encoder to match that teacher's softmax distribution, and the test set is drawn from the same teacher-filtered pool. The NDCG@k and Spearman scores in Table 2 therefore measure agreement with GPT-4o's ordering on teacher-selected confident cases. The Qwen3 baseline is not trained on these teacher labels, so the reported 2.6pp NDCG@3 and 13.3pp Spearman margins could reflect how faithfully each model mimics GPT-4o on GPT-4o's own outputs rather than clinical utility. This undermines the headline production claim in §6.5 and the Conclusion that the 109M model 'outperforms' the 4B model for the medical procedure reranking task.
[§3.2 Human Validation] The human validation does not provide an independent graded ground truth for the metrics reported. Only 100 of 2,647 rows (3.8%) were evaluated, and the experts judged only the acceptability of the teacher's top-1 and top-3 placements, not the full-list ordering that underlies the reported Spearman correlation and NDCG@5. Cohen's κ=0.52 is moderate. This validation cannot support the claim that the synthetic labels are reliable enough to compare fine-tuned models against an untrained LLM baseline on full-list ranking metrics. An independent human-annotated test set with graded relevance labels for all candidate positions, or at least with pairwise preferences, is needed to support the clinical-relevance conclusions.
[§3.2, Quality Filter] The top-3 quality filter introduces a selection bias that makes the test set easier for the teacher and for models trained to imitate it. Discarding all rows where the target is not ranked in the top 3 removes exactly the difficult, lexically distant cases that motivated the study. The distribution shift between the filtered evaluation set and the production query stream is not analyzed, so the reported metrics may not reflect real-world reranking difficulty. The authors should either evaluate on unfiltered queries, report performance as a function of teacher confidence, or provide a separate human-labeled set that includes hard cases.

minor comments (4)

[§5, §6.3] The training-details description is inconsistent: §5 states early stopping with patience 5 and a minimum of 15 epochs, while Appendix B reports patience 15 and a 100-epoch ceiling for the final recipe. The paper should present a single consistent training protocol.
[§7] The claim that 109M-parameter cross-encoders 'can be served on commodity CPU hardware' and the discussion of cost and latency advantages are not backed by any measurements in the paper. A benchmark table with inference time or cost would make the production takeaway concrete.
[Table 2] The abstract and conclusion report '37x fewer parameters,' but Table 2 lists 4B for Qwen3 and 109M for MedCPT; the exact ratio is about 36.7x. Please align the exact wording or use an approximation symbol.
[Figure 2] The colorbar in Figure 2 is not labeled with its units or meaning; adding a caption clarifying whether the color encodes validation NDCG@3 or the improvement over the base model would improve readability.

Circularity Check

1 steps flagged · score 6.0 of 10

Teacher-derived labels make the benchmark circular: the fine-tuned cross-encoder is trained and evaluated on GPT-4o's own rankings, so the reported advantage over Qwen measures teacher-imitation fidelity, not clinical relevance.

fitted input called prediction [§3.2 Dataset Construction (Phase 2 and Quality Filtering); §4.2 Eq. (4); §5 Evaluation protocol; Table 2]
"Phase 2: Relevance Ranking. A separate GPT-4o call (temperature 0.0) ranks all candidate procedures for each generated query. ... We retain only rows where the teacher model places the target procedure within the top 3 positions. ... LListNet = −Σ Py(pi) log Ps(pi) ... Relevance labels follow a linear scheme: for n candidates, scores are assigned as [n, n−1, . . . ,1]."

The 'ground-truth' relevance labels y_i in Eq. (4) are the linear score transform of the GPT-4o-produced ranking π_i from Phase 2, and the same π_i is used to compute NDCG and Spearman on the held-out test set. The quality gate further removes any row where GPT-4o did not rank the target procedure in the top 3, so the test distribution is teacher-selected. Training MedCPT to minimize cross-entropy against P_y therefore trains it to imitate GPT-4o's ordering, and the reported metrics (0.961 NDCG@3; +13.3pp Spearman vs. Qwen) measure agreement with that same teacher on those easy rows. Human validation checked only top-1/top-3 acceptance on 100 rows, not full-list order, so the Spearman comparison has no independent gold standard.

full rationale

The paper's central comparison is internally consistent as a supervised learning evaluation, but its target variable is not independent of the model being trained. GPT-4o generates the queries, produces the relevance rankings in Phase 2, and the quality filter retains only rows where GPT-4o placed the target procedure in the top 3. The ListNet objective (Eq. 4) trains MedCPT to match the softmax distribution induced by those same GPT-4o labels, and the held-out test metrics are computed against the same label source. Consequently, the headline result that a 109M-parameter cross-encoder outperforms Qwen3-Reranker-4B by 2.6pp NDCG@3 and 13.3pp Spearman is a statement about how faithfully each model reproduces GPT-4o's ordering on teacher-selected easy cases, not about clinical relevance per se. The human validation of 100 rows checks only top-1 and top-3 acceptance, not the full-list ranking that drives the Spearman comparison, and covers only 3.8% of the data. The Limitations section candidly acknowledges that 'any systematic biases of this teacher are inherited by the training labels,' but that acknowledgment does not remove the circularity: the reported gap could reflect differential teacher-imitation fidelity rather than clinical utility. No self-citation or uniqueness-import circularity is present; the circularity is specifically that the fitted model's prediction target is defined by the same teacher that generated its training labels and its test labels.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim has no theoretical derivation; it is an empirical comparison. The main load-bearing input is the synthetic training and evaluation set, whose generation relies on a set of hand-chosen pipeline parameters and assumptions about the teacher model's reliability. The Optuna tuning and prompt-optimization settings are additional fitted choices that shape the comparison.

free parameters (6)

GPT-4o query generation temperature = 0.7
Chosen by hand for Phase 1 diversity in §3.2; the central dataset depends on this sampling behavior.
Quality filter top-k threshold = 3
Rows retained only when teacher ranks target in top 3 (§3.2). This selects the dataset and directly shapes the benchmark difficulty.
Candidate subsample size = 20
For services with over 20 procedures, candidates are subsampled to target plus 19 random distractors (§3.2), affecting list composition.
Linear relevance label scheme = [n, n-1, ..., 1]
Ground-truth labels are assigned linearly from teacher rank (§5), turning ranking positions into numeric relevance.
Hyperparameters (LR, batch, warmup, patience, weight decay) = per-config, see Table 4
Selected via Optuna on validation NDCG@3; standard model selection but still tuned to the synthetic validation set.
Agentic prompt optimization loop parameters = k=5, patience=5, 200-query subsample, seed 42
Settings chosen in §4.4 for the prompt optimizer, which affects the Qwen baseline strength.

assumptions (5)

domain assumption GPT-4o relevance rankings reflect clinical relevance
The whole training and test label set is built from GPT-4o rankings (§3.2), and human validation covers only 100 rows.
ad hoc to paper Quality-filtered rows generalize to the full distribution
Keeping only rows where the target is top-3 may remove hard queries and inflate NDCG; see §3.2.
domain assumption Query-level splits prevent information leakage
Authors assert no query appears in more than one split, but candidate procedures may overlap across splits (§3.3), which is not fully analyzed.
domain assumption Domain-expert service-to-procedure mappings are correct
Expert mapping is the upstream source data; incorrect mappings would propagate into generated queries and labels.
ad hoc to paper Human validation sample is representative
Only 100 of 2,647 lists were expert-checked (§3.2); acceptance rates are high but the sample is small and drawn from the synthetic distribution.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Listwise Cross-Encoder Fine-Tuning vs. Agentic Instruction Tuning for LLM Rerankers: A Systematic Study in Medical Procedure Reranking." pith.science (2026). https://pith.science/paper/W3R3NVMJ

@misc{pith2026260809650,
  author       = {Pith},
  title        = {Pith review of: Listwise Cross-Encoder Fine-Tuning vs. Agentic Instruction Tuning for LLM Rerankers: A Systematic Study in Medical Procedure Reranking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W3R3NVMJ}},
  note         = {Machine review of arXiv:2608.09650}
}

read the original abstract

Reranking medical procedures against patient queries is a critical component of health insurance information retrieval, complicated by a substantial lexical gap between patient language and clinical nomenclature. We present a systematic comparison of two reranking paradigms for this production task: (1) small cross-encoders (MedCPT, MiniLM-L12) fine-tuned with listwise learning-to-rank objectives across layer freezing configurations, and (2) Qwen3-Reranker-4B, a 4B-parameter instruction reranker whose prompt is iteratively refined via an agentic optimization loop driven by GPT-4.1. On a purpose-built dataset of 2,647 queries across 708 insurance services, we find that a 109M-parameter cross-encoder fine-tuned with ListNet outperforms the 4B-parameter model by 2.6 percentage points on NDCG@3 and 13.3 points on Spearman correlation - at 37x fewer parameters. We report practical findings, a scalable LLM based dataset construction pipeline, and deployment trade-offs relevant to production reranking systems. We release our code and a sample dataset to support reproducibility and adaptation to other domains.

Figures

Figures reproduced from arXiv: 2608.09650 by the authors.

**Figure 1.** Convergence of agentic prompt optimization for Qwen3-Reranker-4B. Blue dots: all GPT-4.1- proposed candidate instructions per iteration; red line: best-so-far NDCG@3 on the 200-query validation subsample. The optimizer converged within 3 iterations and terminated at iteration 9 via early stopping (patience 5). See Appendix D for a schematic of the greedy search procedure. from the iterative signal, starting from a… view at source ↗

**Figure 3.** Best fine-tuned cross-encoder (MedCPT + ListNet, 109M) vs. prompt-optimized Qwen3-Reranker (4B) vs. unfine-tuned MedCPT baseline on the held-out test set. CE error bars show ±1 std across 5 seeds. +10-19pp for MiniLM), confirming that listwise training substantially improves full-list rank correlation regardless of the specific objective. 6.5 Cross-Encoder vs. LLM Reranker [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Parallel coordinates plot of all 18 fine-tuning configurations on the validation set. Each line traces a [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: Schematic of the agentic prompt optimization procedure as a greedy tree search. Blue nodes: selected best [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Dataset construction pipeline pseudocode. Each entry [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 9 canonical work pages

[1]

and Yeganova, Lana and Wilbur, W

Jin, Qiao and Kim, Won and Chen, Qingyu and Comeau, Donald C. and Yeganova, Lana and Wilbur, W. John and Lu, Zhiyong , title =. Bioinformatics , volume =. 2023 , doi =

work page 2023
[2]

Proceedings of the 2nd Clinical Natural Language Processing Workshop , pages =

Alsentzer, Emily and Murphy, John and Boag, William and Weng, Wei-Hung and Jin, Di and Naumann, Tristan and McDermott, Matthew , title =. Proceedings of the 2nd Clinical Natural Language Processing Workshop , pages =. 2019 , publisher =

work page 2019
[3]

Proceedings of the 2019 Conference of the North

Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , title =. Proceedings of the 2019 Conference of the North. 2019 , publisher =

work page 2019
[4]

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (

Sun, Weiwei and Yan, Lingyong and Ma, Xinyu and Wang, Shuaiqiang and Ren, Pengjie and Chen, Zhumin and Yin, Dawei and Ren, Zhaochun , title =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (. 2023 , publisher =

work page 2023
[5]

arXiv preprint arXiv:2305.02156 , year =

Ma, Xueguang and Zhang, Xinyu and Pradeep, Ronak and Lin, Jimmy , title =. arXiv preprint arXiv:2305.02156 , year =

arXiv
[6]

Proceedings of the 47th European Conference on Information Retrieval (

Pezzuti, Francesca and MacAvaney, Sean and Tonellotto, Nicola , title =. Proceedings of the 47th European Conference on Information Retrieval (. 2025 , doi =

work page 2025
[7]

arXiv preprint arXiv:2506.05176 , year =

Zhang, Yanzhao and Li, Mingxin and Long, Dingkun and Zhang, Xin and Lin, Huan and Yang, Baosong and Xie, Pengjun and Yang, An and Liu, Dayiheng and Lin, Junyang and Huang, Fei and Zhou, Jingren , title =. arXiv preprint arXiv:2506.05176 , year =

arXiv
[8]

Proceedings of the 27th

Wang, Xuanhui and Li, Cheng and Golbandi, Nadav and Bendersky, Michael and Najork, Marc , title =. Proceedings of the 27th. 2018 , publisher =

work page 2018

Show all 11 references

[9]

Proceedings of the 24th International Conference on Machine Learning (

Cao, Zhe and Qin, Tao and Liu, Tie-Yan and Tsai, Ming-Feng and Li, Hang , title =. Proceedings of the 24th International Conference on Machine Learning (. 2007 , publisher =

2007
[10]

Proceedings of the 30th Conference on Uncertainty in Artificial Intelligence (

Lan, Yanyan and Zhu, Yadong and Guo, Jiafeng and Niu, Shuzi and Cheng, Xueqi , title =. Proceedings of the 30th Conference on Uncertainty in Artificial Intelligence (
[11]

Advances in Neural Information Processing Systems 22 (

Chen, Wei and Liu, Tie-Yan and Lan, Yanyan and Ma, Zhi-Ming and Li, Hang , title =. Advances in Neural Information Processing Systems 22 (

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

and Yeganova, Lana and Wilbur, W

Jin, Qiao and Kim, Won and Chen, Qingyu and Comeau, Donald C. and Yeganova, Lana and Wilbur, W. John and Lu, Zhiyong , title =. Bioinformatics , volume =. 2023 , doi =

work page 2023

[2] [2]

Proceedings of the 2nd Clinical Natural Language Processing Workshop , pages =

Alsentzer, Emily and Murphy, John and Boag, William and Weng, Wei-Hung and Jin, Di and Naumann, Tristan and McDermott, Matthew , title =. Proceedings of the 2nd Clinical Natural Language Processing Workshop , pages =. 2019 , publisher =

work page 2019

[3] [3]

Proceedings of the 2019 Conference of the North

Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , title =. Proceedings of the 2019 Conference of the North. 2019 , publisher =

work page 2019

[4] [4]

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (

Sun, Weiwei and Yan, Lingyong and Ma, Xinyu and Wang, Shuaiqiang and Ren, Pengjie and Chen, Zhumin and Yin, Dawei and Ren, Zhaochun , title =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (. 2023 , publisher =

work page 2023

[5] [5]

arXiv preprint arXiv:2305.02156 , year =

Ma, Xueguang and Zhang, Xinyu and Pradeep, Ronak and Lin, Jimmy , title =. arXiv preprint arXiv:2305.02156 , year =

arXiv

[6] [6]

Proceedings of the 47th European Conference on Information Retrieval (

Pezzuti, Francesca and MacAvaney, Sean and Tonellotto, Nicola , title =. Proceedings of the 47th European Conference on Information Retrieval (. 2025 , doi =

work page 2025

[7] [7]

arXiv preprint arXiv:2506.05176 , year =

Zhang, Yanzhao and Li, Mingxin and Long, Dingkun and Zhang, Xin and Lin, Huan and Yang, Baosong and Xie, Pengjun and Yang, An and Liu, Dayiheng and Lin, Junyang and Huang, Fei and Zhou, Jingren , title =. arXiv preprint arXiv:2506.05176 , year =

arXiv

[8] [8]

Proceedings of the 27th

Wang, Xuanhui and Li, Cheng and Golbandi, Nadav and Bendersky, Michael and Najork, Marc , title =. Proceedings of the 27th. 2018 , publisher =

work page 2018