Pith. sign in

REVIEW 3 major objections 4 minor 31 references

Beyond Retrieval: Ensembling Cross-Encoders and GPT Rerankers with LLMs for Biomedical QA

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A retrieval-ensemble RAG system achieves rank-1 factoid answers and rank-10 document retrieval on the BioASQ 2025 biomedical QA benchmark.

desk verdict A genuine and useful BioASQ system-description with verifiable leaderboard results, but its internal design conclusions rest on small, unrepresentative validation splits and should not be read as robust evidence. read the letter →

arxiv 2507.05577 v1 pith:7CIBHCWM submitted 2025-07-08 cs.IR cs.CLcs.LG

Shashank Verma , Fengyi Jiang , Xiangning Xue This is my paper
classification cs.IRcs.CLcs.LG
keywords biomedicalquestionansweringretrieval-augmentedgenerationdenseretrievalcross-encoderre-rankinglistwisefew-shotpromptingBioASQTask13bensemble
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 tries to establish that a tiered retrieval-augmented generation (RAG) stack can handle all four biomedical question types of the BioASQ 2025 Task13b challenge at a competitive level. The stack combines dense vector retrieval over roughly 34 million PubMed abstracts, an ensemble of a finetuned MiniLM-L12 cross-encoder and a GPT-4o listwise reranker for document selection, and few-shot prompts to instruction-tuned LLMs for answer generation. On the official leaderboard the system placed 10th for document retrieval and, notably, 1st for factoid exact answers by MRR. A sympathetic reading is that the paper's contribution is the full-pipeline recipe plus evidence that weighted ensembling of rerankers and prompt-style choices matter more than model scale alone.

What carries the argument

The central object is the weighted ensemble re-ranker. It is built from three ranked sources: bge-large-en dense embeddings stored in an approximate-nearest-neighbor index for recall; a finetuned ms-marco-MiniLM-L12 cross-encoder (33.4M parameters) that jointly scores question-document pairs; and GPT-4o applied as a listwise re-ranker to the top-30 candidates, re-ordering them by prompt. The two re-ranker outputs are fused by a weighted sum of relevance scores, with weights 1 and 7 found by grid search on the internal validation split. This object carries the paper's Phase A argument: the ensemble outperforms each member and creates the document context on which all Phase B answers depend.

What would settle it

Run a full ablation on a held-out portion of the BioASQ test questions with confidence intervals: if replacing the ensemble with the finetuned cross-encoder alone does not significantly lower MAP@10, the paper's central retrieval claim is not supported.

Watch

Extended reading notes

Core claim

The authors argue that the decisive ingredient for retrieval quality is the two-stage re-ranking ensemble: retrieving the top 1,000 candidates with a bge-large-en bi-encoder, passing them through a finetuned ms-marco-MiniLM-L12 cross-encoder to obtain a top-30, and then having GPT-4o perform listwise re-ranking of those 30 to produce the final top-10, combined by weighted score summation with weights 1 and 7. This ensemble reached an internal MAP@10 of 0.4551 against 0.4337 for the finetuned cross-encoder alone. For answer generation, the paper claims that GPT-4o-turbo with its 'style 3' answer-formatting prompt, which instructs the model to extract verbatim phrases from passages, yields the best list and ideal answers, and that few-shot prompting alone was sufficient to reach the top of the factoid leaderboard with an MRR of 0.64.

Load-bearing premise

The system's design choices rest on small internal test sets that do not represent the official evaluation, so the reported benefits of finetuning, data-drift filtering, and the specific ensemble weights may not hold on the actual leaderboard.

Editorial extensions

If this is right

  • If the weighted ensemble claim holds, small domain-finetuned cross-encoders are the cost-effective workhorse, and an LLM listwise reranker adds the last precision gain.
  • If the rank-1 factoid result generalizes, few-shot instruction-tuned prompting is a viable replacement for span-extraction heads in biomedical factoid QA.
  • The prompt-style effect implies that answer formatting instructions are a primary control variable for list and ideal answers under n-gram overlap metrics.
  • The tiered bi-encoder-to-cross-encoder-to-LLM-reranker design is a reusable recipe for large-corpus domain retrieval in other specialised literatures.

Reading between the lines

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

  • The authors leave implicit that the gap between internal MAP@10 (0.4551) and leaderboard MAP@10 (0.1581) indicates their internal proxy is systematically optimistic; a uniform sample of questions across all five batches would give a more honest ablation estimate.
  • An editorial extension: the style-3 verbatim-extraction instruction suggests ROUGE-based evaluation rewards lexical copying; evaluating ideal answers with a semantic-equivalence measure would likely change the ranking of GPT-4o versus the smaller Mistral model.
  • A testable follow-up would be to replace GPT-4o with an open-weight listwise reranker to measure how much of the ensemble gain depends on raw model capability versus the listwise prompting protocol itself.
  • If factoid MRR of 0.64 is stable across repeated answer generations, it would indicate that instruction-tuned LLMs with a handful of snippets can outperform dedicated medical reading-comprehension models when the question is a standard factoid.
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.

Referee Report

3 major / 4 minor

Summary. The paper describes a retrieval-augmented generation (RAG) system submitted to the BioASQ 2025 Task 13b challenge. For Phase A (document retrieval), the authors build a dense index over roughly 34 million PubMed abstracts using bge-large-en embeddings, retrieve the top 1,000 candidates, and then rerank them with a finetuned ms-marco-MiniLM-L12 cross-encoder and a listwise GPT-4o reranker, combined in a weighted ensemble (weights 1 and 7). The system reportedly achieved a leaderboard MAP@10 of 0.1581 (rank 10) for batch 4. For Phase B (answer generation), the authors use few-shot prompting of GPT-4o-turbo and Mistral-7B-Instruct-v0.3 with three prompt styles, achieving leaderboard scores of 0.95 macro-F1 for yes/no questions (rank 12), 0.64 MRR for factoid questions (rank 1), 0.63 mean-F1 for list questions (rank 5), and 0.29 ROUGE-SU4 F1 for ideal answers (rank 11). The paper also reports internal experiments on cross-encoder finetuning, ensemble design, and prompt styles on a 100-question test split and an 80-question sampled set from the training data. The central leaderboard results are externally verifiable, but the internal design conclusions rely on small, non-representative test sets with no uncertainty quantification.

Significance. If taken at face value, the leaderboard results demonstrate that a relatively small finetuned cross-encoder plus an LLM reranker can be competitive on a public biomedical QA benchmark, and that few-shot prompting of instruction-tuned LLMs can achieve rank 1 on factoid answers. The paper's strengths include the use of a large-scale real PubMed index, the clear system architecture, and the fact that the headline scores are externally verified on the official leaderboard, which avoids circularity in the main empirical claim. The principal weakness is that all internal comparisons used to motivate the design—choice of cross-encoder, finetuning benefits, ensemble weights, retrieval count, and prompt styles—are evaluated on a single 100-question split (for retrieval) and an 80-question sample (for generation), without error bars or significance testing. The paper's own data show a large distribution gap between internal and leaderboard performance (e.g., internal MAP@10 0.4551 vs. leaderboard 0.1581). This does not invalidate the reported challenge scores, but it substantially weakens any claim of optimality or generalizable superiority for the specific design choices.

major comments (3)
  1. [§4.1, Table 2; §4.2.2; §5] The internal retrieval evaluation used to select and validate the reranking configuration is based on a single test split of the last 100 questions in the dataset, with no error bars or significance tests. The authors themselves note in §5 that the internal perfect yes/no scores did not carry over to the leaderboard, and the large drop from internal MAP@10 of 0.4551 to the leaderboard MAP@10 of 0.1581 indicates that this split is not representative of the actual batch distribution. Consequently, the claims that the weighted ensemble with weights 1 and 7 outperforms the cross-encoder alone (0.4551 vs. 0.4337), and that retrieving 1k before reranking is optimal, are not robustly established. I recommend reporting bootstrap confidence intervals or paired significance tests across several random splits of the dataset, and tempering the optimality claims if the differences are not statistically significant.
  2. [§4.4, Tables 4 and 5; §5] The prompt-style comparison for answer generation is based on 80 randomly sampled training questions (20 per type) using gold snippets. The internal differences between prompt styles (e.g., GPT-4o-turbo factoid MRR of 0.46 for style 2 vs. 0.35 for style 3 in Table 4) are presented as meaningful, but with only 20 questions per type, these differences could easily be noise; no variance or significance information is given. The paper itself acknowledges in §5 that "sampled evaluation questions may not comprehensively represent the complexity and variability encountered in real-world questions," which undercuts the strong design conclusions drawn from Tables 4 and 5. Additionally, the leaderboard results in Tables 6 and 7 show only small differences between styles (e.g., 0.01–0.03 in MRR/F1), further suggesting that the internal gains are not stable. Please add uncertainty quantification and discuss the reliability of the observed differences, or explicitly present the prompt-style experiments as anecdotal rather than as validated conclusions.
  3. [§3.2.2] The statement that "Discarding older questions from training ... turned out to be true in our case" is a design-relevant claim but is never supported with any quantitative comparison, table, or figure. Since this finetuning decision is used to justify the final model, the absence of any reported metric makes the claim unverifiable. Provide at least a summary of the MAP@10 values with and without older questions, or remove the claim if the experiment is not reproducible from the paper.
minor comments (4)
  1. [Author footnote] The author footnote contains a corrupted string "envel⌢pe-⌢pensverma342@gatech.edu" that appears to be an encoding artifact; please fix the contact-email formatting.
  2. [References] Reference [23] has an unusual and likely corrupted author list ("D. S. Chaplot, Albert q. jiang, alexandre sablayrolles, ..."); the entry should be corrected to the proper citation for the Mistral-7B technical report.
  3. [Tables 3, 6, and 7] The paper uses the terms "dense rank" and "non-dense rank" without defining them; add a sentence explaining how dense rank differs from the standard competition rank when ties are present.
  4. [§5] The sentence "It is interesting to note that each next stage is as much or more accurate as well as more expensive on the IR task" is vague; please specify which stages and which metric are meant.

Circularity Check

1 steps flagged · score 2.0 of 10

Central leaderboard results are externally validated; the only construction artifact is the acknowledged ROUGE-oriented Style 3 prompt, a secondary design claim.

  1. fitted input called prediction [Section 4.4 (Phase B: Answer Generation), ideal-answer discussion after Tables 5 and 7]
    "The improvement observed with style 3 highlights a limitation of ROUGE metrics: they focus on surface-level n-gram overlap rather than semantic equivalence. Style 3 explicitly instructed the model to“only combining direct phrases or sentences from the Passage” and to “avoid paraphrasing the original wording unless necessary, ” which increased overlap with the gold snippets and thus boosted ROUGE scores."

    ROUGE-2 and ROUGE-SU4 measure n-gram overlap between the generated text and the gold answer. The internal simulation constructed queries using gold-standard snippets, so the Passage in the prompt is the gold snippet text. Style 3's formatting template instructs the model to copy direct phrases from that Passage and avoid paraphrasing, which by construction increases n-gram overlap with gold answers derived from the same snippets. The paper explicitly admits this, yet still presents Style 3 as giving 'measurable improvements in summary F1 scores.' The measured improvement is therefore engineered into the metric definition, not an independent finding. This is a secondary, self-acknowledged design observation; the headline leaderboard results remain externally evaluated.

full rationale

The paper's headline claims are evaluated against the external BioASQ 2025 Task13b leaderboard: MAP@10 of 0.1581, MRR of 0.64, list mF1 of 0.63, yes/no maF1 of 0.95, and ROUGE-SU4 F1 of 0.29. Those numbers are not produced by the paper's own equations or fitted values, so the central empirical claim is self-contained and not circular. The retrieval pipeline (bge-large-en index, finetuned MiniLM-L12 cross-encoder, GPT-4o listwise reranker, weighted ensemble) is selected using standard train/validation/test splits and grid search; selection on validation data followed by a separate leaderboard evaluation is normal model selection, not a fitted input renamed as a prediction. The one construct-level issue is the Style 3 prompt for ideal answers: the paper explicitly instructs the model to reuse passage phrasing, which mechanically raises ROUGE n-gram overlap, and the paper itself acknowledges this as a limitation of ROUGE metrics. This makes the 'Style 3 improves ideal answers' result partially by construction, but it is a secondary, transparently disclosed design observation rather than a load-bearing part of the externally verified leaderboard claims. The small internal evaluation sizes and the mismatch between internal and leaderboard scores are robustness and generalizability concerns, not circularity. There are no load-bearing self-citations and no imported uniqueness theorems, so the appropriate circularity score is low.

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

The paper introduces no new entities. Its burden on the reader comes from fitted hyperparameters (ensemble weights, retrieval depth, epochs, data cutoff, few-shot count) that are tuned on small validation sets, plus standard domain assumptions about embedding quality, ANN search, LLM reliability, and label correctness. The most fragile assumption, that the last 100 training questions approximate the leaderboard distribution, is contradicted by the paper's own numbers and undermines the internal comparisons.

free parameters (5)
  • ensemble weights for re-ranker combination = 1 for finetuned MiniLM-L12 cross-encoder, 7 for GPT-4o reranker
    Selected by grid search on the validation set in Section 3.2.4; the paper reports these weights yield MAP@10 of 0.4551 on its internal test set but gives no sensitivity analysis.
  • number of documents retrieved from index before reranking = 1000
    Chosen by comparing MAP@10 for Top 10 through Top 1k in Table 2; the paper notes 2k performed better on the leaderboard than 1k for a single reranker, so this value is not robust.
  • number of finetuning epochs = 3 for the final cross-encoder (1 epoch for screening)
    Section 4.2.1 reports the best model was finetuned for 1 epoch then for 3 epochs, reaching MAP@10 of 0.4337; no learning curve or early stopping details are given.
  • training data time cutoff for discarding older questions = not specified (last few years)
    Section 3.2.2 claims this dataset variation improved results but provides no numbers or exact cutoff.
  • number of few-shot examples (1 vs 10) = 1-shot for GPT-4o submissions, 10-shot for Mistral submissions
    Section 3.3 and Table 6; the paper reports no significant difference between 1-shot and 10-shots for style 1, but submits different counts per model.
assumptions (5)
  • domain assumption BioASQ gold documents and answers are correct and complete for training and evaluation.
    All finetuning labels and evaluation rely on the BioASQ-QA corpus labels being accurate; stated in Section 1.
  • domain assumption Dense embeddings from bge-large-en capture semantic relevance for biomedical query-document pairs.
    Section 3.1.2 adopts bge-large-en as the index encoder without task-specific evaluation; the recall improvement over PubMed API in Table 1 is the only evidence.
  • domain assumption The HNSW approximate nearest neighbor search in Qdrant returns sufficiently accurate top-k results.
    Section 3.1.3 relies on Qdrant's ANN search to retrieve the candidate pool; no recall analysis for ANN accuracy is provided.
  • domain assumption Instruction-tuned LLMs (GPT-4o-turbo, Mistral-7B-Instruct) provide reliable few-shot reasoning for biomedical answers.
    Section 3.3 uses these models as black-box generators without probing their failure modes; the paper notes quality issues (verbosity, synonyms) that prompt styles partially fix.
  • ad hoc to paper The last 100 questions of the BioASQ training set are a valid proxy for the leaderboard question distribution.
    Table 2 is evaluated on this split; the paper's own leaderboard results show this proxy is poor (internal MAP@10 0.4551 vs leaderboard 0.1581), so the assumption is contradicted by the paper's data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Retrieval: Ensembling Cross-Encoders and GPT Rerankers with LLMs for Biomedical QA." pith.science (2026). https://pith.science/paper/7CIBHCWM

@misc{pith2026250705577,
  author       = {Pith},
  title        = {Pith review of: Beyond Retrieval: Ensembling Cross-Encoders and GPT Rerankers with LLMs for Biomedical QA},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7CIBHCWM}},
  note         = {Machine review of arXiv:2507.05577}
}
read the original abstract

Biomedical semantic question answering rooted in information retrieval can play a crucial role in keeping up to date with vast, rapidly evolving and ever-growing biomedical literature. A robust system can help researchers, healthcare professionals and even layman users access relevant knowledge grounded in evidence. The BioASQ 2025 Task13b Challenge serves as an important benchmark, offering a competitive platform for advancement of this space. This paper presents the methodologies and results from our participation in this challenge where we built a Retrieval-Augmented Generation (RAG) system that can answer biomedical questions by retrieving relevant PubMed documents and snippets to generate answers. For the retrieval task, we generated dense embeddings from biomedical articles for initial retrieval, and applied an ensemble of finetuned cross-encoders and large language models (LLMs) for re-ranking to identify top relevant documents. Our solution achieved an MAP@10 of 0.1581, placing 10th on the leaderboard for the retrieval task. For answer generation, we employed few-shot prompting of instruction-tuned LLMs. Our system achieved macro-F1 score of 0.95 for yes/no questions (rank 12), Mean Reciprocal Rank (MRR) of 0.64 for factoid questions (rank 1), mean-F1 score of 0.63 for list questions (rank 5), and ROUGE-SU4 F1 score of 0.29 for ideal answers (rank 11).

Figures

Figures reproduced from arXiv: 2507.05577 by the authors.

Figure 1
Figure 1. Distribution of question types in BioASQ training data (left) and number of golden documents & snippets per question (right). a complete set of items without ranking. In addition to the exact answers, each of the question types also contain an ideal answer which is a natural language response to the question. In this paper, we present our automated Retrieval-Augmented Generation (RAG) system, which aims to solve thi… view at source ↗
Figure 2
Figure 2. Architecture of the best performing system for Phase A. We use bge-large-en for creating our index and perform ensembled re-ranking to get the top-10 most relevant documents for a query [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overall Prompt Template Structure. Template texts wrapped in {} are subject to change according to styles and questions. Query Template 1 Passage: {All snippets} Question: {Question body} {Answer Formatting Template, e.g. Answer this question in “yes” or “no”. } Query Template 2 (changed from Template 1 only for ideal answer prompts:) Passage: {All snippets} Question: {Question body} (Hint: short answer is {System g… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Query Templates for exact and ideal answers. We experimented with two primary dimensions in our prompting strategy: 1. Number of Few-Shot Examples: We tested using either a single example (1-shot) or ten examples (10-shots) as contextual demonstrations, corresponding t…
Figure 5
Figure 5. Figure 5: Answer Formatting Templates. Template 1 was presented in Ateia & Kruschwitz’s published work [26] and thus omitted. 2. Prompts Templates Design: Each prompt varied along two core components: the Query Template, which defined how few-shot examples and questions were pre…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 15 canonical work pages

  1. [1]

    Nentidis, G

    A. Nentidis, G. Katsimpras, A. Krithara, G. Paliouras, Overview of BioASQ Tasks 13b and Synergy13 in CLEF2025, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025

    work page 2025

  2. [2]

    Nentidis, G

    A. Nentidis, G. Katsimpras, A. Krithara, M. Krallinger, M. Rodríguez-Ortega, E. Rodriguez-López, N. Loukachevitch, A. Sakhovskiy, E. Tutubalina, D. Dimitriadis, G. Tsoumakas, G. Giannakoulas, A. Bekiaridou, A. Samaras, G. M. Di Nunzio, N. Ferro, S. Marchesin, M. Martinelli, G. Silvello, G. Paliouras, Overview of bioasq 2025: The thirteenth bioasq challeng...

    work page 2025

  3. [3]

    Krithara, A

    A. Krithara, A. Nentidis, K. Bougiatiotis, G. Paliouras, BioASQ-QA: A manually curated corpus for Biomedical Question Answering, Scientific Data 10 (2023) 170

    work page 2023

  4. [4]

    S. E. Robertson, S. Walker, S. Jones, M. M. Hancock-Beaulieu, M. Gatford, et al., Okapi at trec-3, Nist Special Publication Sp 109 (1995) 109

    work page 1995

  5. [5]

    Lewis, E

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, et al., Retrieval-augmented generation for knowledge-intensive nlp tasks, arXiv preprint arXiv:2005.11401 (2020) 1–22

    arXiv 2020

  6. [6]

    Izacard, E

    G. Izacard, E. Grave, Leveraging passage retrieval with generative models for open domain question answering, Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume (2021) 1545–1556

    work page 2021

  7. [7]

    Y. A. Malkov, D. A. Yashunin, Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs, arXiv preprint arXiv:1603.09320 (2016) 1–14

    arXiv 2016

  8. [8]

    Johnson, M

    J. Johnson, M. Douze, H. Jégou, Billion-scale similarity search with gpus, arXiv preprint arXiv:1702.08734 (2017) 1–12

    arXiv 2017

Show all 31 references
  1. [9]

    Bernhardsson, Annoy: Approximate nearest neighbors in c++/python optimized for memory usage and loading speed, GitHub Repository (2018)

    E. Bernhardsson, Annoy: Approximate nearest neighbors in c++/python optimized for memory usage and loading speed, GitHub Repository (2018)

    2018

  2. [10]

    Y. Gu, R. Tinn, H. Cheng, M. Lucas, N. Usuyama, X. Liu, T. Naumann, J. Gao, H. Poon, Domain- specific language model pretraining for biomedical natural language processing, ACM Transactions on Computing for Healthcare (HEALTH) 3 (2021) 1–23

    2021

  3. [11]

    W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, M. Zhou, Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers, Advances in neural information processing systems 33 (2020) 5776–5788

    2020

  4. [12]

    Reimers, I

    N. Reimers, I. Gurevych, Sentence-bert: Sentence embeddings using siamese bert-networks, arXiv preprint arXiv:1908.10084 (2019) 1–16

    2019 arXiv

  5. [13]

    G. M. Rosa, R. C. Rodrigues, R. Lotufo, R. Nogueira, In defense of cross-encoders for zero-shot retrieval, arXiv preprint arXiv:2212.06121 (2022) 1–12

    2022 arXiv

  6. [14]

    Nogueira, K

    R. Nogueira, K. Cho, Passage re-ranking with bert, arXiv preprint arXiv:1901.04085 (2019) 1–6

    2019 arXiv

  7. [15]

    Thakur, N

    N. Thakur, N. Reimers, A. Rücklé, A. Srivastava, I. Gurevych, Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models, arXiv preprint arXiv:2104.08663 (2021) 1–20

    2021 arXiv

  8. [16]

    X. Ma, X. Zhang, R. Pradeep, J. Lin, Zero-shot listwise document reranking with a large language model, 2023. URL: https://arxiv.org/abs/2305.02156. arXiv:2305.02156

    2023 arXiv

  9. [17]

    Pradeep, S

    R. Pradeep, S. Sharifymoghaddam, J. Lin, Rankvicuna: Zero-shot listwise document rerank- ing with open-source large language models, 2023. URL: https://arxiv.org/abs/2309.15088. arXiv:2309.15088

    2023 arXiv

  10. [18]

    R. Luo, L. Sun, Y. Xia, T. Qin, S. Zhang, H. Poon, T.-Y. Liu, Biogpt: generative pre-trained transformer for biomedical text generation and mining, Briefings in bioinformatics 23 (2022) bbac409

    2022

  11. [19]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, W. Chen, Lora: Low-rank adaptation of large language models, arXiv preprint arXiv:2106.09685 (2021) 1–25

    2021 arXiv

  12. [20]

    Pfeiffer, A

    J. Pfeiffer, A. Rücklé, C. Poth, A. Kamath, I. Vulić, S. Ruder, K. Cho, I. Gurevych, Adapterfusion: Non-destructive task composition for transfer learning, Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume ...

    2020

  13. [21]

    J. Lee, W. Yoon, S. Kim, D. Kim, S. Kim, C. H. So, J. Kang, Biobert: a pre-trained biomedical language representation model for biomedical text mining, Bioinformatics 36 (2020) 1234–1240

    2020

  14. [22]

    Achiam, S

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al., Gpt-4 technical report, arXiv preprint arXiv:2303.08774 (2023)

    2023 arXiv

  15. [23]

    D. S. Chaplot, Albert q. jiang, alexandre sablayrolles, arthur mensch, chris bamford, devendra singh chaplot, diego de las casas, florian bressand, gianna lengyel, guillaume lample, lucile saulnier, lélio renard lavaud, marie-anne lachaux, pierre stock, teven le scao, thibaut ...

    2023 arXiv

  16. [24]

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learners, arXiv preprint arXiv:2005.14165 (2020) 1–75

    2020 arXiv

  17. [25]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, D. Zhou, Chain-of- thought prompting elicits reasoning in large language models, arXiv preprint arXiv:2201.11903 (2022) 1–62

    2022 arXiv

  18. [26]

    Ateia, U

    S. Ateia, U. Kruschwitz, Can open-source llms compete with commercial models? exploring the few-shot performance of current gpt models in biomedical tasks, arXiv preprint arXiv:2407.13511 (2024)

    2024 arXiv

  19. [27]

    URL: https: //pubmed.ncbi.nlm.nih.gov/download/

    National Library of Medicine, Download pubmed data, PubMed User Guide (2025). URL: https: //pubmed.ncbi.nlm.nih.gov/download/

    2025

  20. [28]

    URL: https://www.pinecone.io/ pricing/

    Pinecone Systems, Inc., Pricing, Pinecone Documentation (2025). URL: https://www.pinecone.io/ pricing/

    2025

  21. [29]

    URL: https://qdrant.tech/ documentation/overview/

    Qdrant Technologies, Qdrant collection, Qdrant Documentation (2025). URL: https://qdrant.tech/ documentation/overview/

    2025

  22. [30]

    Zamani, M

    H. Zamani, M. Dehghani, W. B. Croft, E. Learned-Miller, J. Kamps, From neural re-ranking to neural ranking: Learning a sparse representation for inverted indexing, in: Proceedings of the 27th ACM International Conference on Information and Knowledge Management, CIKM ’18, Assoc...

    2018

  23. [31]

    I. A. Prodromos Malakasiotis, Ioannis Pavlopoulos, A. Nentidis, http://participants-area.bioasq. org/Tasks/b/eval_meas_2022/, 2020

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.