Retrieval Augmented Biomedical Question Answering with Weak Question Recovery and Neural Reranking for BioASQ Task 14b

Balaji Anandganesh; Lee Mai; Xueying Zhao

REVIEW 4 major objections 5 minor 20 references

Retrieval Augmented Biomedical Question Answering with Weak Question Recovery and Neural Reranking for BioASQ Task 14b

T0 review · 4 major / 5 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 pipeline that detects weak questions and re-expands them improved MAP@10 on BioASQ Batch 4 by about 28% over its own baseline, and the paper argues the gain comes from targeting hard queries rather than from larger rerankers.

desk verdict A cleanly written BioASQ system paper whose headline 28% MAP gain is not actually isolated to the claimed component, plus a concrete table inconsistency. read the letter →

arxiv 2608.01468 v1 pith:Y5IJMCA4 submitted 2026-08-02 cs.CL cs.AI

Xueying Zhao , Lee Mai , Balaji Anandganesh This is my paper

classification cs.CLcs.AI

keywords BioASQbiomedicalquestionansweringretrieval-augmentedgenerationneuralrerankingweak-questionrecoveryqueryexpansionreciprocalrankfusionOpenLLM

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

Biomedical question answering is only as good as the documents it retrieves, and this paper targets the cases where retrieval fails. The authors build a pipeline around PubMed search, MiniLM-based semantic reranking, reciprocal rank fusion, and snippet pruning, and add a conditional step that identifies 'weak' questions from evidence quality signals and re-expands them. On BioASQ Task 14b Batch 4, the final pipeline raises $\mathrm{MAP@10}$ from 0.0745 to 0.0955, a relative gain of about 28%, while the set of retrieved documents changes only modestly. Their central argument is that selectively repairing hard queries can help more than scaling up the reranker: several larger reranking variants scored lower in their comparison. If this holds, retrieval robustness for biomedical QA can be improved without a large computational footprint.

What carries the argument

The load-bearing object is the weak-question recovery gate: a composite lexical relevance score with thresholds (weak below 0.35, strong above 0.55) that decides which questions get additional semantic and relationship-aware expansion before results are selectively merged back into the ranked list. The same gate, together with retrieval signals like unique candidate count, MiniLM relevance, snippet redundancy, and lexical overlap, is what targets the hard questions. The thresholds are set empirically on the training collection.

What would settle it

Run the fixed recovery pipeline with the 0.35/0.55 thresholds on a held-out BioASQ batch that was not used to set thresholds, and compare MAP@10 against the pipeline without recovery; if the relative gain is much smaller than 28% or is negative, the calibration does not transfer.

Watch

Extended reading notes

Core claim

The paper's central claim is that a conditional weak-question recovery mechanism makes retrieval more robust on difficult biomedical questions. The system scores retrieved snippets with a composite lexical relevance score combining keyword recall, entity overlap, phrase matching, and topic-drift penalties. Thresholds below 0.35 are treated as weak evidence and above 0.55 as strong; questions dominated by weak evidence or with limited retrieval coverage are selected for recovery, then re-expanded with semantic expansion, relationship-aware augmentation, and synonym enrichment. Improved results are merged selectively, and a pruning stage drops redundant or low-relevance snippets. On Batch 4 th

Load-bearing premise

The load-bearing assumption is that the weak/strong evidence thresholds (0.35 and 0.55) and the retrieval signals that trigger recovery, which were calibrated empirically on the BioASQ training set, transfer to the same batch where the 28% improvement is measured; the paper gives no held-out validation of this calibration.

Editorial extensions

If this is right

If the recovery gains hold, retrieval pipelines for biomedical QA should treat hard questions separately rather than applying a single global retrieval configuration.
The improvement pattern (MAP up, F1 roughly flat) implies recovery mainly reorders relevant evidence to the top of the list, which is what exact-answer generation depends on.
Larger rerankers are not an automatic win; with targeted recovery, a fine-tuned lightweight MiniLM system can outperform generic or larger semantic rerankers on this benchmark.
Controlled snippet pruning to roughly two high-confidence snippets per question preserves answer coverage while reducing noise for generation.
Post-processing and format normalization materially affect BioASQ submission outcomes, independent of reasoning quality.

Reading between the lines

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

Inference: because the 0.35/0.55 thresholds were tuned on the training set and the reported gain is on the same Batch 4 used to develop them, the 28% figure is likely an upper bound for what would transfer to a genuinely new batch; a held-out evaluation would be needed to confirm.
Inference: the composite lexical score could be replaced by a learned failure predictor trained on retrieval outcomes, which would remove the manual threshold and possibly improve recall of weak questions.
Inference: the recovery loop's selective merging is a simple form of adaptive retrieval; it could be combined with agentic or self-reflective retrieval strategies to decide not just whether to re-expand a query but how many retrieval rounds to spend.
Inference: the snippet-coverage analysis (diminishing returns after about two snippets) suggests a general evidence-budget rule for RAG systems: prompt length is better spent on a few high-confidence snippets than on many candidates, a hypothesis testable outside BioASQ.

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

Dear colleague,

What you should know: this is a straightforward BioASQ 14b system paper, clearly written, with one genuinely new twist—a conditional weak-question recovery module that selectively re-queries when initial retrieval looks weak. The rest is standard: query expansion, MiniLM reranking, RRF, RAG with OpenBioLLM. The novelty is at the engineering-variant level, not a new principle.

The paper does several things well. It is honest about the biggest weakness of its own headline claim: Section 5.2 explicitly says the reported 28% MAP improvement is from the combined pipeline, not the recovery component alone. It also provides a useful analysis of snippet pruning versus answer coverage (Figure 2), and the local proxy MAP (0.0955) matches the official MAP (0.0956), which is a good sanity check.

Now the soft spots, in roughly increasing order of seriousness.

First, the abstract and discussion attribute the gains specifically to the "recovery and cleanup strategies," which goes beyond what the experiments show. Table 2 compares the full final pipeline to the base pipeline; no run disables only the recovery module. So the central causal claim is untested, and the authors themselves defer the ablation to future work.

Second, there is a concrete data inconsistency. Table 2 reports Precision@10 = 0.0367 for the final run, while Table 3 reports Precision@10 = 0.0733 for what appears to be the same configuration. Both have Recall@10 = 0.1539 and MAP@10 ≈ 0.0955/0.0956. These cannot both be right. One table has a typo, and a referee needs to catch it.

Third, the weak-evidence thresholds (0.35 and 0.55) were "empirically" set on the training set and then applied to Batch 4, the same batch on which the improvement is measured. There is no held-out validation that the calibration transfers, and the paper never reports how many questions were flagged as weak. So the "difficult question sets" part of the claim is not quantified.

The absolute numbers are also small (MAP@10 0.0955 vs 0.0745), and the authors acknowledge their system is mid-table. That is fine for a competition system paper, but it caps the significance.

My take: this paper deserves peer review, not desk rejection, because it is a clean, honest systems description with a plausible new component that just needs proper ablation and a corrected table. But the 28% claim should not be cited as evidence for the recovery module until the ablation exists. I would not cite it in my own work yet, though I might bring it to a reading group as an example of how "integrated pipeline" comparisons can oversell a component.

Recommendation: send it to review, but flag the table inconsistency and require an isolated ablation or a revised claim.

Referee Report

4 major / 5 minor

Summary. The paper describes a BioASQ Task 14b retrieval-augmented QA pipeline integrating multi-source query expansion, PubMed retrieval, fine-tuned MiniLM reranking, Reciprocal Rank Fusion, a conditional weak-question recovery mechanism, post-retrieval pruning, and OpenBioLLM-based answer generation with output post-processing. The authors report official Phase A, A+, and B results across four BioASQ batches and a local proxy comparison on Batch 4. The headline empirical claim is that the final pipeline, including weak-question recovery and cleanup, improves MAP@10 by approximately 28% over a base retrieval configuration (0.0745 to 0.0955, Table 2), with Recall@10 rising from 0.1335 to 0.1539. The paper also compares several reranking variants and concludes that the lightweight MiniLM reranker combined with the full pipeline is more effective than larger or generic cross-encoder models.

Significance. If fully supported, the work would provide a useful engineering contribution to biomedical retrieval and RAG pipelines: a lightweight, modular retrieval system with competitive official BioASQ results and a targeted recovery mechanism for difficult questions. The paper has strengths: it reports official evaluation numbers, gives detailed operational descriptions of the pipeline, and provides a local proxy that closely matches the official Batch 4 MAP (0.0955 vs. 0.0956). However, the central causal claim about weak-question recovery and cleanup is not yet supported: the reported gain comes from a single batch, an integrated pipeline comparison with no component ablation, thresholds calibrated on the training set without a held-out split, and a numerical inconsistency between Table 2 and Table 3. The significance is therefore conditional on additional validation.

major comments (4)

[Section 5.2, Table 2] The central claim of a 28% MAP@10 improvement is attributed in the Abstract, Section 1, and Section 6 to 'weak-question recovery and cleanup strategies,' but Table 2 compares the full integrated pipeline against the base pipeline. As the authors state in Section 5.2, the improvement reflects 'the combined contribution of query expansion, neural reranking, weak-question recovery, retrieval fusion, and retrieval cleanup rather than the isolated effect of any individual component,' and Section 7 defers component-wise ablations. No run disables the recovery module while fixing all other components, so the causal claim is untested. This is load-bearing for the paper's main contribution.
[Section 5.2, Table 2] The evidence for the headline improvement rests on a single evaluation batch (Batch 4), with no error bars, no significance test, and no replication across batches. A one-batch difference of 0.021 in MAP@10 (0.0745 vs. 0.0955) is not sufficient to support a general claim about 'difficult question sets' or 'retrieval robustness.' The paper should report multiple batches and ideally a held-out development/test split for any configuration decisions.
[Section 4.1 and Section 5.2] The weak-question thresholds (composite lexical relevance score below 0.35 for weak, above 0.55 for strong) are described as 'empirically' determined, apparently on the training set, and then applied to the same Batch 4 on which the improvement is measured. There is no held-out validation that these thresholds transfer, and the paper never reports how many Batch 4 questions were classified as weak. Without this information, the claim that recovery specifically helps 'difficult question sets' is unquantified and the risk of overfitting to Batch 4 is unresolved.
[Table 2 vs. Table 3] There is a numerical inconsistency: Table 2 reports Precision@10 = 0.0367 for the 'Batch 4 Final + Rescue' configuration, while Table 3 reports Precision@10 = 0.0733 for the same 'Final MiniLM-based system' on the same Batch 4 evaluation split. The MAP values also differ slightly (0.0955 vs. 0.0956). Both tables cannot be correct for the same configuration and split; the authors must identify which values are official, which are local, and reconcile the precision discrepancy before the results can be considered reliable.

minor comments (5)

[Section 4.4.4, Eq. (1)] The macro-F1 formula is written as F_macro^1 = 1/2 (F_yes^1 + F_no^1). This is standard, but the superscript notation is confusing because '1' appears both as a superscript and in F1. Using F1_yes and F1_no would improve clarity.
[Section 5.2, Table 1] Table 1 includes a footnote that Batch 1 values were 'recalculated locally using the final pipeline after correcting a submission formatting issue.' This is not clearly labeled as a local recalculation in the table header, which says 'official BioASQ Phase A retrieval results.' Clarify which rows are official and which are local.
[Section 5.3, Table 4] In Table 4, several entries run together without spacing, e.g., '0.42110.3529' and '0.17070.1583'. This makes the table hard to read and should be fixed.
[Section 1, references [3-6]] References [3-6] are self-citations to papers on machine learning in healthcare and science communication. They are not clearly related to the BioASQ retrieval task or RAG pipeline discussed in the introduction; either integrate them into the relevant context or remove them.
[Section 4.2] The claim that coverage gains diminish after about two snippets is supported by Figure 2, but the precise measure of 'vocabulary coverage' and the aggregation method are not described. A brief definition would help reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the MAP@10 gain is an empirical pipeline comparison; the missing ablation weakens causal attribution to recovery/cleanup but is not a derivation-circle.

full rationale

The paper's derivation chain is not circular. The central claim — that weak-question recovery and cleanup improve MAP@10 on Batch 4 — is an empirical comparison of two retrieval configurations (Table 2: Batch 4 Base MAP@10 0.0745 vs Batch 4 Final+Rescue MAP@10 0.0955). The thresholds used to flag weak questions (0.35/0.55) were calibrated during system development on the training set, not fitted to the Batch 4 outcome, so the reported gain is not forced by construction. The reranker is trained on the BioASQ training collection and evaluated on an official batch; no equation uses its own output as an input. The self-citations [3–6] in Section 1 are contextual and not load-bearing. The paper explicitly concedes that the comparison reflects the combined contribution of query expansion, neural reranking, weak-question recovery, retrieval fusion, and retrieval cleanup rather than the isolated effect of any individual component (Section 5.2), and Section 7 defers component-wise ablation. That is a real evidentiary gap for the specific attribution to recovery/cleanup, but it is not circularity. The Precision@10 discrepancy between Table 2 (0.0367) and Table 3 (0.0733) for the same configuration is a data-quality concern, not a circularity. Overall, the derivation is self-contained against external evaluation batches and does not reduce to its inputs.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the generalization of training-set-derived thresholds and on the credibility of the local proxy evaluation. No new entity is introduced. The free parameters are empirical cutoffs, not fitted constants in a derivation.

free parameters (4)

Weak evidence threshold (composite lexical relevance score < 0.35) = 0.35
Empirical cutoff from training set to label snippets as weak evidence (Section 4.1).
Strong evidence threshold (composite lexical relevance score > 0.55) = 0.55
Empirical cutoff from training set to label snippets as strong evidence (Section 4.1).
Snippet pruning count = approximately 2 snippets per question
Derived from diminishing returns in training-set vocabulary coverage (Section 4.2, Figure 2), used to control pruning.
Reciprocal Rank Fusion constant k
The RRF constant is not specified in the paper, so the fusion behavior is underdetermined.

assumptions (4)

domain assumption The composite lexical relevance score correlates with true snippet relevance
Section 4.1 uses this score to decide which questions are weak and eligible for recovery, without validation against gold labels on the test set.
domain assumption The diminishing-returns coverage curve from the BioASQ training set (two snippets) transfers to the test batches
Section 4.2 and Figure 2 motivate pruning based on training-set coverage; no test-set verification.
domain assumption BioASQ Phase B released evidence snippets provide a valid gold-label proxy for local MAP evaluation
Section 5.2 uses Phase B evidence with a local evaluation script; the paper claims close match to official Batch 4 MAP=0.0956 but does not show per-question agreement.
domain assumption OpenBioLLM and MiniLM pretrained models behave as external black-box tools as documented
The system's performance depends on OpenBioLLM-70B (Section 4.4.1) and MiniLM fine-tuning (Section 4.3); these artifacts are not inspected.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Retrieval Augmented Biomedical Question Answering with Weak Question Recovery and Neural Reranking for BioASQ Task 14b." pith.science (2026). https://pith.science/paper/Y5IJMCA4

@misc{pith2026260801468,
  author       = {Pith},
  title        = {Pith review of: Retrieval Augmented Biomedical Question Answering with Weak Question Recovery and Neural Reranking for BioASQ Task 14b},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y5IJMCA4}},
  note         = {Machine review of arXiv:2608.01468}
}

read the original abstract

This work presents DS@GT ARC BioASQ team's work for a biomedical question answering pipeline, integrating multi-source query expansion, neural reranking, retrieval refinement, and OpenBioLLM-assisted answer generation. The system combines PubMed retrieval with fine-tuned MiniLM-based semantic reranking, Reciprocal Rank Fusion (RRF), and feature-based relevance scoring to improve document ranking quality. To address challenging queries with weak retrieval performance, we introduce a conditional weak-question recovery strategy that applies semantic expansion, relationship-aware augmentation, and selective result merging. A post-retrieval pruning stage further removes redundant or low-relevance snippets while preserving evidence coverage for downstream answer generation. Experimental results on BioASQ evaluation batches demonstrate that the proposed recovery and cleanup strategies substantially improve retrieval robustness and MAP@10 performance on difficult question sets. The final system also incorporates output validation and post-processing steps to ensure formatting consistency and submission reliability across BioASQ phases.

Figures

Figures reproduced from arXiv: 2608.01468 by the authors.

**Figure 1.** illustrates the overall retrieval and reranking framework used in our system. The retrieval pipeline begins by expanding each question using domain-specific strategies. These include training-memory expansion based on similar BioASQ questions, biomedical synonym expansion using OpenBioLLM, and relationship-aware augmentation that capture common associations such as disease–treatment or gene–pathway links. These co… view at source ↗

**Figure 2.** Relationship between the number of retrieved snippets and ideal-answer vocabulary coverage across the BioASQ training set. Coverage improves rapidly for the first few snippets but shows diminishing returns beyond approximately two snippets, motivating controlled snippet pruning and selective evidence filtering in the final retrieval pipeline. The shaded region represents one standard deviation across questions [PIT… view at source ↗

**Figure 3.** Training and validation loss curves for the fine-tuned MiniLM-based reranker across 10 training epochs. The training loss decreases steadily while the validation loss stabilizes after early epochs, indicating stable convergence with limited overfitting. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 17 canonical work pages

[1]

Tsatsaronis, G

G. Tsatsaronis, G. Balikas, P. Malakasiotis, I. Partalas, M. Zschunke, M. R. Alvers, D. Weissenborn, A. Krithara, S. Petridis, D. Polychronopoulos, Y. Almirantis, J. Pavlopoulos, N. Baskiotis, P. Gallinari, T. Artères, A.-C. N. Ngomo, N. Heino, E. Gaussier, J. Barrow-Green, X. Tannier, J. Mork, A. Nentidis, I. Androutsopoulos, An overview of the BIOASQ la...

work page 2015
[2]

Nentidis, K

A. Nentidis, K. Bougiatiotis, A. Krithara, G. Paliouras, Results of the seventh edition of the BioASQ challenge, in: Joint European Conference on Machine Learning and Knowledge Discovery in Databases, Springer, Cham, 2019, pp. 553–568

work page 2019
[3]

X. Zhao, W. Leo, A. Liu-Ferrara, Harnessing machine learning to identify breast cancer mortality risk factors and health care access inequities, European Journal of Cancer 237 (2026)

work page 2026
[4]

X. Zhao, B. Gao, AI and microfluidics: Unlocking cellular motility for bioengineering, Bioengi- neering 13 (2026) 172

work page 2026
[5]

X. Zhao, Y. Jiang, A. Radenbaugh, T. McCarthy, J. Moskwa, D. Jensen, Integrated artificial intelligence frameworks in single-cell multiomics: From intelligent automation to generative modeling, Artificial Intelligence in Health (2026) 025010119

work page 2026
[6]

Hayslett, X

M. Hayslett, X. Zhao, S. Cook, Beginning to define competencies in science communication training: a commentary, International Journal of Science Education, Part B 14 (2024) 465–469

work page 2024
[7]

Stuhlmann, M

L. Stuhlmann, M. A. Saxer, J. Furst, Efficient and reproducible biomedical question answering using retrieval augmented generation, arXiv preprint arXiv:2505.07917 (2025)

arXiv 2025
[8]

M. Hu, L. Zong, H. Wang, J. Zhou, J. Li, Y. Gao, K.-F. Wong, Y. Li, I. King, Serts: Self-rewarding tree search for biomedical retrieval-augmented generation, in: Findings of the Association for Computational Linguistics: EMNLP 2024, 2024, pp. 1321–1335

work page 2024

Show all 20 references

[9]

M. R. Rezaei, R. S. Fard, J. L. Parker, R. G. Krishnan, M. Lankarany, Agentic medical knowledge graphs enhance medical question answering: Bridging the gap between LLMs and evolving medical knowledge, arXiv preprint arXiv:2502.13010 (2025)

2025 arXiv
[10]

Ateia, U

S. Ateia, U. Kruschwitz, Can language models critique themselves? investigating self-feedback for retrieval augmented generation at bioasq 2025, arXiv preprint arXiv:2508.05366 (2025)

2025 arXiv
[11]

M. S. Ankit Pal, OpenBioLLMs: Advancing open-source large language models for healthcare and life sciences, 2024. URL: https://huggingface.co/aaditya/OpenBioLLM-Llama3-70B

2024
[12]

L. Fang, Q. Chen, C.-H. Wei, Z. Lu, K. Wang, Bioformer: an efficient transformer language model for biomedical text mining, ArXiv (2023) arXiv–2302

2023
[13]

Nogueira, Z

R. Nogueira, Z. Jiang, R. Pradeep, J. Lin, Document ranking with a pretrained sequence-to-sequence model, in: Findings of the Association for Computational Linguistics: EMNLP 2020, 2020, pp. 708–718

2020
[14]

Reimers, I

N. Reimers, I. Gurevych, Sentence-BERT: Sentence embeddings using siamese bert-networks, in: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2019, p...

2019
[15]

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

2023
[16]

Sharma, P

K. Sharma, P. Kumar, Y. Li, OG-RAG: Ontology-grounded retrieval-augmented generation for large language models, in: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 2025, pp. 32950–32969

2025
[17]

Sapkota, R

R. Sapkota, R. Shrestha, M. Rijal, M. Karkee, LangChain vs. LangGraph vs. LangSmith: Taxonomies of Agentic AI Toolchains for End-to-End Orchestration, 2025. Authorea Preprint

2025
[18]

J. Lin, J. Tang, H. Tang, S. Yang, W.-M. Chen, W.-C. Wang, G. Xiao, X. Dang, C. Gan, S. Han, AWQ: Activation-aware weight quantization for on-device llm compression and acceleration, Proceedings of Machine Learning and Systems 6 (2024) 87–100

2024
[19]

W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, I. Stoica, Efficient memory management for large language model serving with PagedAttention, in: Proceedings of the 29th Symposium on Operating Systems Principles (SOSP 2023), ACM, New York, NY,...

2023
[20]

URL: http://www

PACE, Partnership for an Advanced Computing Environment (PACE), 2017. URL: http://www. pace.gatech.edu

2017

Pith tools

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

[1] [1]

Tsatsaronis, G

G. Tsatsaronis, G. Balikas, P. Malakasiotis, I. Partalas, M. Zschunke, M. R. Alvers, D. Weissenborn, A. Krithara, S. Petridis, D. Polychronopoulos, Y. Almirantis, J. Pavlopoulos, N. Baskiotis, P. Gallinari, T. Artères, A.-C. N. Ngomo, N. Heino, E. Gaussier, J. Barrow-Green, X. Tannier, J. Mork, A. Nentidis, I. Androutsopoulos, An overview of the BIOASQ la...

work page 2015

[2] [2]

Nentidis, K

A. Nentidis, K. Bougiatiotis, A. Krithara, G. Paliouras, Results of the seventh edition of the BioASQ challenge, in: Joint European Conference on Machine Learning and Knowledge Discovery in Databases, Springer, Cham, 2019, pp. 553–568

work page 2019

[3] [3]

X. Zhao, W. Leo, A. Liu-Ferrara, Harnessing machine learning to identify breast cancer mortality risk factors and health care access inequities, European Journal of Cancer 237 (2026)

work page 2026

[4] [4]

X. Zhao, B. Gao, AI and microfluidics: Unlocking cellular motility for bioengineering, Bioengi- neering 13 (2026) 172

work page 2026

[5] [5]

X. Zhao, Y. Jiang, A. Radenbaugh, T. McCarthy, J. Moskwa, D. Jensen, Integrated artificial intelligence frameworks in single-cell multiomics: From intelligent automation to generative modeling, Artificial Intelligence in Health (2026) 025010119

work page 2026

[6] [6]

Hayslett, X

M. Hayslett, X. Zhao, S. Cook, Beginning to define competencies in science communication training: a commentary, International Journal of Science Education, Part B 14 (2024) 465–469

work page 2024

[7] [7]

Stuhlmann, M

L. Stuhlmann, M. A. Saxer, J. Furst, Efficient and reproducible biomedical question answering using retrieval augmented generation, arXiv preprint arXiv:2505.07917 (2025)

arXiv 2025

[8] [8]

M. Hu, L. Zong, H. Wang, J. Zhou, J. Li, Y. Gao, K.-F. Wong, Y. Li, I. King, Serts: Self-rewarding tree search for biomedical retrieval-augmented generation, in: Findings of the Association for Computational Linguistics: EMNLP 2024, 2024, pp. 1321–1335

work page 2024