pith. sign in

arxiv: 2606.12578 · v1 · pith:6XGTGBE7new · submitted 2026-06-10 · 💻 cs.CL

MARD: Mirror-Augmented Reasoning Distillation for Mechanism-Level Drug-Drug Interaction Prediction

Mohammadreza Riyazat , Vian Lelo , Rameen Jafri , Yumna Khan , Abeer Badawi This is my paper

Pith reviewed 2026-06-27 09:43 UTC · model grok-4.3

classification 💻 cs.CL
keywords drug-drug interactionmechanism predictionreasoning distillationcold-split evaluationprocess reward modelDDI taxonomypharmacological reasoningsmall language model
0
0 comments X

The pith

A 7B model trained with reasoning distillation predicts drug interaction mechanisms on unseen pairs more accurately than GPT-4o.

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

The paper develops a reproducible protocol for mechanism-level drug-drug interaction prediction that labels not only whether two drugs interact but which enzyme or axis is involved, the direction, and the supporting evidence. It trains a 7B model called MARD using three specific techniques: single-token KL divergence to anchor direction predictions, process-reward-weighted DPO with hard negatives, and a leakage-safe retrieval channel. On a cold-split evaluation from the April-2026 DrugBank release, this model is the only one among 32 systems whose accuracy holds for novel drug pairs, exceeding the strongest baseline by 13.9 points and GPT-4o by 6.7 points at roughly one percent of the API cost. An accompanying analysis shows accuracy rising on rarely encountered drugs, which the authors interpret as evidence of structured pharmacological reasoning instead of frequency-based memorization.

Core claim

MARD-7B is the only system in a 32-system comparison whose accuracy survives drug-pair novelty, beating the best baseline by +13.9 pp and GPT-4o by +6.7 pp at ~1% of frontier API cost, achieved by combining single-token KL divergence on direction tags, per-loss PRM-weighted DPO with programmatic hard negatives, and leakage-safe mechanism-aware retrieval, with process-reward labels automatically verifiable against DrugBank fields.

What carries the argument

Mirror-Augmented Reasoning Distillation pipeline that ties model predictions to direction tags via KL divergence, weights DPO steps by verifiable process rewards, and adds a leakage-safe retrieval channel over the taxonomy.

If this is right

  • Accuracy improves on rarely seen drugs, indicating the gains come from structured pharmacological reasoning rather than drug-frequency memorization.
  • Process-reward step labels can be produced automatically from DrugBank-structured fields without human or LLM judges.
  • The method supports auditable reasoning metrics that go beyond flat binary interaction classification.
  • Performance holds under leakage-safe cold-split protocols designed to test novelty in drug pairs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Specialized distillation on domain taxonomies may let smaller models handle precise biomedical reasoning tasks more reliably than general frontier models.
  • The anti-memorization pattern could be tested on other structured prediction problems where new entities appear after training.
  • If the retrieval channel proves essential, similar leakage-safe indexes might improve generalization in additional scientific domains with evolving entity sets.

Load-bearing premise

The leakage-safe cold-split protocols and the 7-family/147-subtype taxonomy together ensure that performance gains reflect genuine generalization to unseen drug pairs rather than residual data leakage or taxonomy-specific artifacts.

What would settle it

Running the same 32-system comparison on a random split that permits drug-pair overlap with training data, or on a later DrugBank release, and finding that MARD-7B no longer leads or that its accuracy on rarely seen drugs stops improving.

Figures

Figures reproduced from arXiv: 2606.12578 by Abeer Badawi, Mohammadreza Riyazat, Rameen Jafri, Vian Lelo, Yumna Khan.

Figure 1
Figure 1. Figure 1: MARD-7B pipeline for an unseen drug pair. The system retrieves structured evidence and similar labelled pairs, trains with mirror-augmented reasoning and PRM-weighted DPO, generates a citation-grounded reasoning trace, and produces a verifiable DDI prediction. MFS and HR are corpus-level evaluation metrics. conclusion. Appendix B expands the full schema and the step-role vocabulary. Mirror requirement. Fli… view at source ↗
Figure 2
Figure 2. Figure 2: Reliability diagram on RANDOM￾SPLIT (WARM). Greedy decode (red) vs. PRM_VOTE_CONSENSUS + conformal (blue); dotted line is perfect calibration. ECE 0.068 vs. 0.231. category difference, not a model gap: the MLP emits a single 7-way family label and cannot pre￾dict subtype, direction, or abstention (App. D). Trace quality and joint score. Beyond the flat family label, the distilled MARD is jointly mirror￾sta… view at source ↗
Figure 3
Figure 3. Figure 3: Where MARD-7B wins. (a) Family Macro-F1 across three test protocols (the cold-split robustness advantage). (b) Per-decile accuracy as a function of training-pair frequency (the anti-memorisation sign flip) [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The deployed MARD-7B stack already captures the bulk of available signal. Gap between our learned router and the all-candidate oracle is only +0.064 on Drug-Cold and +0.094 on Pair-Cold — the bottleneck is candidate selection, not generation, moti￾vating the PRM-weighted reranker (§4.6). Limitations L1. The selection–generation gap is the work this paper does not do. The candidate-pool or￾acle (Tab. 14) re… view at source ↗
Figure 5
Figure 5. Figure 5: End-to-end case study for pair DB00582|DB06626 (Voriconazole + Axitinib). Row 1 lists the raw DrugBank fields read for each drug; the same UniProt P08684 (CYP3A4) appears with opposite roles (inhibitor vs. substrate), which is the mechanistic signal. Row 2 shows what the MARD-7B actually receives — the binary PK-flag table, the four pair-level similarity scalars (all low, so structure / drug-class similari… view at source ↗
Figure 6
Figure 6. Figure 6: Inference-time depth curve. PRM-argmax macro-F1 (blue circles) and rare-F1 (orange squares) saturate at N=4; voting at the same N=8 compute budget unlocks an additional +0.014 macro-F1 and – on the right-hand axis – a +0.092 jump in mirror family stability (green star) over the best PRM-argmax MFS curve (green dashed). Numbers in [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Coverage–accuracy Pareto frontier across 15 aggregator variants on RANDOM-SPLIT (WARM). The dashed green step line is the Pareto envelope. The deployed PRM_VOTE_CONSENSUS+conformal stack (blue star, top-right) trades −45% coverage for +14.4 pp Macro-F1 over greedy (lower-left cluster). The parameter-free PRM_VOTE_CONSENSUS (blue tri￾angle) sits on the envelope at 0.601 Macro-F1 with no learned threshold, c… view at source ↗
Figure 8
Figure 8. Figure 8: Per-split reliability diagrams for the deployed PRM_VOTE_CONSENSUS+conformal stack. The calibration fit transfers from RANDOM-SPLIT (WARM) (left) to the two cold generalisation splits (centre, right) within ±0.02 ECE, showing that the conformal thresholds calibrated on RANDOM-SPLIT (WARM).VAL remain valid when the deployment distri￾bution shifts to held-out drugs or novel pairs. Three-split ITS-layer ablat… view at source ↗
Figure 9
Figure 9. Figure 9: Per-family Macro-F1 across all three splits. Three stacked panels (RANDOM-SPLIT (WARM) / DRUG￾COLD / PAIR-COLD) on the common 497-pair stratified slice. Gold border marks the column winner; the rightmost column is the per-row Macro-F1 summary. Top row of each panel is MARD-7B. The bottom annotation tracks MARD-7B’s Macro-F1 trajectory across the three regimes: MARD-7B is the only system whose score does no… view at source ↗
read the original abstract

Mechanism-level drug-drug interaction (DDI) prediction requires identifying which enzyme or pharmacodynamic axis is implicated, in which direction, and with which evidence -- not merely whether two drugs interact. We introduce a reproducible mechanism-level DDI labelling and evaluation protocol with a structured 7-family/147-subtype taxonomy, leakage-safe cold-split protocols, and auditable reasoning metrics for evaluating pharmacological prediction beyond flat interaction classification. We propose a pipeline that produces a 7B reasoning MARD (Mirror-Augmented Reasoning Distillation), combining three training innovations: a single-token KL divergence on direction tag that ties the model's prediction, per-loss PRM-weighted DPO with programmatic hard negatives, and a leakage-safe mechanism-aware retrieval channel. Process-reward step labels are automatically verifiable against DrugBank-structured fields, requiring no human or LLM judges. On the April-2026 DrugBank release, our MARD-7B is the only system in a 32-system comparison whose accuracy survives drug-pair novelty, beating the best baseline by +13.9 pp and GPT-4o by +6.7 pp at ~1% of frontier API cost. Further analysis reveals an anti-memorisation signature where accuracy improves on rarely seen drugs, suggesting that gain comes from structured pharmacological reasoning rather than drug-frequency memorisation. We release corpus, DDI-PRM, retrieval index, and training code.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper introduces a reproducible mechanism-level DDI prediction protocol using a 7-family/147-subtype taxonomy, leakage-safe cold splits, and auditable reasoning metrics. It proposes MARD-7B, a 7B model trained via mirror-augmented reasoning distillation (single-token KL on direction tags, PRM-weighted DPO with programmatic hard negatives, and leakage-safe mechanism-aware retrieval). On the April-2026 DrugBank release, MARD-7B is claimed to be the sole system among 32 whose accuracy survives drug-pair novelty, outperforming the best baseline by +13.9 pp and GPT-4o by +6.7 pp at ~1% API cost, with an anti-memorization signature where accuracy rises on rarely seen drugs. Code, corpus, DDI-PRM, and retrieval index are released.

Significance. If the cold-split protocols and taxonomy truly eliminate leakage on drugs, pairs, and subtypes while the gains reflect structured pharmacological reasoning, the work would advance DDI prediction beyond flat classification toward mechanism-level, generalizable, and low-cost models. The automatic verifiability of process rewards against DrugBank fields and the release of training artifacts are strengths supporting reproducibility.

major comments (1)
  1. [Abstract and evaluation protocol section] Abstract and evaluation protocol section: the central claim that MARD-7B is the only system whose accuracy survives drug-pair novelty (with +13.9 pp gain) rests on the unverified assertion that the 7-family/147-subtype taxonomy plus cold-split protocols ensure zero shared drugs, zero shared pairs, zero shared mechanism subtypes, and a retrieval index built strictly on train-only data. No explicit audit, split statistics, or code snippet confirming these properties is provided, leaving open the possibility that performance exploits taxonomy-level patterns or retrieval leakage rather than genuine generalization.

Simulated Author's Rebuttal

1 responses · 0 unresolved

Thank you for your constructive review. We agree that explicit verification of the leakage-safe properties is necessary to support the central claims and will strengthen the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract and evaluation protocol section] Abstract and evaluation protocol section: the central claim that MARD-7B is the only system whose accuracy survives drug-pair novelty (with +13.9 pp gain) rests on the unverified assertion that the 7-family/147-subtype taxonomy plus cold-split protocols ensure zero shared drugs, zero shared pairs, zero shared mechanism subtypes, and a retrieval index built strictly on train-only data. No explicit audit, split statistics, or code snippet confirming these properties is provided, leaving open the possibility that performance exploits taxonomy-level patterns or retrieval leakage rather than genuine generalization.

    Authors: We agree this is a valid concern and that the current manuscript lacks the requested quantitative audit. In the revised version we will add a new subsection under Evaluation Protocol that reports: (i) exact counts of unique drugs, pairs, and mechanism subtypes in each split with explicit zero-overlap confirmation; (ii) a self-contained code snippet implementing the cold-split logic; and (iii) a statement plus index-construction log verifying that the retrieval index contains only training-set mechanisms. These additions will directly rule out taxonomy-level or retrieval leakage as alternative explanations for the reported gains. revision: yes

Circularity Check

0 steps flagged

No circularity: claims rest on external DrugBank labels and claimed leakage-safe splits without reduction to fitted inputs by construction

full rationale

The abstract describes a structured taxonomy, leakage-safe cold-split protocols, and programmatic hard negatives derived from DrugBank fields for DPO training, with process-reward labels verifiable against those fields. However, no equations or steps are shown that make the reported accuracy (or the +13.9 pp gain) equivalent to the training inputs by definition. The central generalization claim is presented as depending on the cold-split and taxonomy design rather than being forced by self-definition or self-citation. No load-bearing self-citation chain or ansatz smuggling is visible in the provided text. This is the normal case of a paper whose performance metric is externally benchmarked against held-out data.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central performance claim rests on the validity of the new taxonomy and cold-split protocol plus the assumption that DrugBank structured fields provide ground-truth process rewards without annotation error; no free parameters are explicitly named but loss weighting in DPO and retrieval hyperparameters are implicit; no new physical entities are postulated.

free parameters (1)
  • DPO loss weights and PRM scaling factors
    The per-loss PRM-weighted DPO requires choosing relative weights between the KL term, the preference term, and the retrieval channel; these are fitted or tuned to achieve the reported accuracy.
axioms (2)
  • domain assumption DrugBank structured fields supply accurate, complete mechanism annotations that can serve as automatic process-reward labels without human verification
    The paper states that process-reward step labels are automatically verifiable against DrugBank-structured fields.
  • ad hoc to paper The 7-family/147-subtype taxonomy partitions pharmacological mechanisms without overlap or omission that would bias the cold-split evaluation
    The taxonomy is introduced as part of the reproducible labelling protocol.

pith-pipeline@v0.9.1-grok · 5795 in / 1714 out tokens · 23596 ms · 2026-06-27T09:43:13.257644+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

94 extracted references · 19 canonical work pages

  1. [2]

    Lukas Berglund, Meg Tong, Max Kaufmann, Mikita Balesni, Asa Cooper Stickland, Tomasz Korbak, and Owain Evans. 2024. https://arxiv.org/abs/2309.12288 The reversal curse: LLMs trained on `` A is B '' fail to learn `` B is A '' . In The Twelfth International Conference on Learning Representations

    arXiv 2024

  2. [3]

    Tianqi Chen and Carlos Guestrin. 2016. https://arxiv.org/abs/1603.02754 Xgboost: A scalable tree boosting system . In KDD

    Pith/arXiv arXiv 2016

  3. [5]

    Chi, Xuezhi Wang, and Denny Zhou

    Xinyun Chen, Ryan A. Chi, Xuezhi Wang, and Denny Zhou. 2024. https://proceedings.mlr.press/v235/chen24i.html Premise order matters in reasoning with large language models . In Proceedings of the 41st International Conference on Machine Learning, volume 235, pages 6596--6620. PMLR

    2024

  4. [7]

    Gabriele De Vito, Filomena Ferrucci, and Athanasios Angelakis. 2025. https://arxiv.org/abs/2502.06890 LLMs for drug-drug interaction prediction: A comprehensive comparison . arXiv preprint arXiv:2502.06890

    Pith/arXiv arXiv 2025

  5. [8]

    Yifan Deng, Xinran Xu, Yang Qiu, Jingbo Xia, Wen Zhang, and Shichao Liu. 2020. https://academic.oup.com/bioinformatics/article/36/15/4316/5837109 A multimodal deep learning framework for predicting drug-drug interaction events . Bioinformatics

    2020

  6. [10]

    Yao Fu, Hao Peng, Litu Ou, Ashish Sabharwal, and Tushar Khot. 2023. https://proceedings.mlr.press/v202/fu23d.html Specializing smaller language models towards multi-step reasoning . In Proceedings of the 40th International Conference on Machine Learning, volume 202, pages 10421--10430. PMLR

    2023

  7. [11]

    Ziabari, and Morteza Dehghani

    Parsa Hejabi, Elnaz Rahmati, Alireza S. Ziabari, and Morteza Dehghani. 2025. https://arxiv.org/abs/2510.14242 Flip-flop consistency: Unsupervised training for robustness to prompt perturbations in LLMs . arXiv preprint arXiv:2510.14242

    arXiv 2025

  8. [12]

    Yixin Hong, Pengyu Luo, Shuting Jin, and Xiangrong Liu. 2022. https://academic.oup.com/bioinformatics/article/38/24/5406/6769887 Lagat: Link-aware graph attention network for drug-drug interaction prediction . Bioinformatics

    2022

  9. [13]

    Glass, and Jimeng Sun

    Kexin Huang, Cao Xiao, Lucas M. Glass, and Jimeng Sun. 2020. https://ojs.aaai.org/index.php/AAAI/article/view/5412 Caster: Predicting drug interactions with chemical substructure representation . In AAAI

    2020

  10. [16]

    Minoru Kanehisa and Susumu Goto. 2000. https://doi.org/10.1093/nar/28.1.27 KEGG : Kyoto Encyclopedia of Genes and Genomes . Nucleic Acids Research, 28(1):27--30

    work page doi:10.1093/nar/28.1.27 2000

  11. [17]

    Sunyoung Kim, Hyeri Lee, Jaeyu Park, Jiseung Kang, Masoud Rahmati, Sang Youl Rhee, and Dong Keon Yon. 2024. https://doi.org/10.1016/j.archger.2024.105465 Global and regional prevalence of polypharmacy and related factors, 1997--2022: An umbrella review . Archives of Gerontology and Geriatrics, 122:105465

    work page doi:10.1016/j.archger.2024.105465 2024

  12. [18]

    Klinger, and 1 others

    Craig Knox, Michael Wilson, Christopher M. Klinger, and 1 others. 2024. https://doi.org/10.1093/nar/gkad976 Drugbank 6.0: the drugbank knowledgebase for 2024 . Nucleic Acids Research, 52(D1):D1265--D1275

    work page doi:10.1093/nar/gkad976 2024

  13. [21]

    Zimeng Li, Shichao Zhu, Bin Shao, Xiangxiang Zeng, Tong Wang, and Tie-Yan Liu. 2023. https://academic.oup.com/bib/article/24/1/bbac597/6966537 Dsn-ddi: An accurate and generalizable drug-drug interaction prediction network . In Briefings in Bioinformatics

    2023

  14. [22]

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. https://arxiv.org/abs/2305.20050 Let's verify step by step . In ICLR

    Pith/arXiv arXiv 2023

  15. [23]

    Guanying Liu, Yifan Zhang, Xuan Liu, and Quanming Yao. 2025. https://arxiv.org/abs/2505.23034 Case-based reasoning enhances the predictive power of LLM s in drug-drug interaction . In arXiv:2505.23034

    arXiv 2025

  16. [25]

    Manning, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. https://arxiv.org/abs/2305.18290 Direct preference optimization: Your language model is secretly a reward model . In Advances in Neural Information Processing Systems (NeurIPS)

    Pith/arXiv arXiv 2024

  17. [26]

    Yaniv Romano, Matteo Sesia, and Emmanuel Cand\` e s. 2020. https://proceedings.neurips.cc/paper_files/paper/2020/file/244edd7e85dc81602b7615cd705545f5-Paper.pdf Classification with valid and adaptive coverage . In NeurIPS

    2020

  18. [27]

    Jae Yong Ryu, Hyun Uk Kim, and Sang Yup Lee. 2018. https://doi.org/10.1073/pnas.1803294115 Deep learning improves prediction of drug-drug and drug-food interactions . In Proceedings of the National Academy of Sciences, volume 115, pages E4304--E4311

    work page doi:10.1073/pnas.1803294115 2018

  19. [28]

    Shuaijie She, Junxiao Liu, Yifeng Liu, Jiajun Chen, Xin Huang, and Shujian Huang. 2025. https://arxiv.org/abs/2503.21295 R-PRM : Reasoning-driven process reward modeling . In arXiv:2503.21295

    arXiv 2025

  20. [29]

    Karan Singhal, Shekoofeh Azizi, Tao Tu, S. Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, Perry Payne, Martin Seneviratne, Paul Gamble, Chris Kelly, Abubakr Babiker, Nathanael Sch\" a rli, Aakanksha Chowdhery, Philip Mansfield, Dina Demner-Fushman, and 13 others. 2023. https://doi.org/10.1038/s415...

    work page doi:10.1038/s41586-023-06291-2 2023

  21. [30]

    Zhaoyue Sun, Jiazheng Li, Gabriele Pergola, and Yulan He. 2024. https://arxiv.org/abs/2409.05592 ExDDI : Explaining drug-drug interaction predictions with natural language . In AAAI

    arXiv 2024

  22. [31]

    Yao Tian, Jiacai Yi, Ningning Wang, Chengkun Wu, Jinfu Peng, Shao Liu, Guoping Yang, and Dongsheng Cao. 2025. https://doi.org/10.1093/nar/gkae726 Ddinter 2.0: An enhanced drug interaction resource with expanded data coverage, new interaction types, and improved user interface . Nucleic Acids Research, 53(D1):D1356--D1362

    work page doi:10.1093/nar/gkae726 2025

  23. [33]

    Vladimir Vovk, Alex Gammerman, and Glenn Shafer. 2005. https://link.springer.com/book/10.1007/b106715 Algorithmic Learning in a Random World . Springer

    work page doi:10.1007/b106715 2005

  24. [34]

    Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. https://arxiv.org/abs/2203.11171 Self-consistency improves chain of thought reasoning in language models . In ICLR

    Pith/arXiv arXiv 2023

  25. [35]

    Ziyan Wang, Zhankun Xiong, Feng Huang, Xuan Liu, and Wen Zhang. 2024. https://arxiv.org/abs/2407.00891 ZeroDDI : A zero-shot drug-drug interaction event prediction method with semantic enhanced learning and dual-modal uniform alignment . In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI)

    arXiv 2024

  26. [36]

    Guangzhi Xiong, Qiao Jin, Zhiyong Lu, and Aidong Zhang. 2024. https://aclanthology.org/2024.findings-acl.372/ Benchmarking retrieval-augmented generation for medicine . In ACL Findings

    2024

  27. [37]

    Zhankun Xiong, Shichao Liu, Feng Huang, Ziyan Wang, Xuan Liu, Zhongfei Zhang, and Wen Zhang. 2023. https://doi.org/10.1609/aaai.v37i4.25665 Multi-relational contrastive learning graph neural network for drug-drug interaction event prediction . In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 5339--5347

    work page doi:10.1609/aaai.v37i4.25665 2023

  28. [38]

    Huimin Xu, Xin Mao, Feng-Lin Li, Xiaobao Wu, and 1 others. 2025. https://arxiv.org/abs/2502.14356 Full-step- DPO : Self-supervised preference optimization with step-wise rewards for mathematical reasoning . In arXiv:2502.14356

    arXiv 2025

  29. [39]

    Glass, Jimeng Sun, and Cao Xiao

    Yue Yu, Kexin Huang, Chao Zhang, Lucas M. Glass, Jimeng Sun, and Cao Xiao. 2021. https://doi.org/10.1093/bioinformatics/btab207 Sumgnn: Multi-typed drug interaction prediction via efficient knowledge graph summarization . In Bioinformatics

    work page doi:10.1093/bioinformatics/btab207 2021

  30. [40]

    Jaehoon Yun, Jiwoong Sohn, Jungwoo Park, Hyunjae Kim, Xiangru Tang, Daniel Shao, Yong Hoe Koo, Minhyeok Ko, Qingyu Chen, Mark Gerstein, Michael Moor, and Jaewoo Kang. 2025. https://arxiv.org/abs/2506.11474 Med-prm: Medical reasoning models with stepwise, guideline-verified process rewards . Preprint, arXiv:2506.11474

    arXiv 2025

  31. [42]

    AISTATS , year=

    A General Theoretical Paradigm to Understand Learning from Human Preferences , author=. AISTATS , year=

  32. [43]

    arXiv:2110.14168 , year=

    Training Verifiers to Solve Math Word Problems , author=. arXiv:2110.14168 , year=

    Pith/arXiv arXiv

  33. [44]

    arXiv:2211.14275 , year=

    Solving Math Word Problems with Process- and Outcome-Based Feedback , author=. arXiv:2211.14275 , year=

    Pith/arXiv arXiv

  34. [45]

    ICLR , year=

    Let's Verify Step by Step , author=. ICLR , year=

  35. [46]

    arXiv:2403.04706 , year=

    Common 7B Language Models Already Possess Strong Math Capabilities , author=. arXiv:2403.04706 , year=

    arXiv

  36. [47]

    arXiv:2501.19393 , year=

    S1: Simple Test-Time Scaling , author=. arXiv:2501.19393 , year=

    Pith/arXiv arXiv

  37. [48]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

    Teaching Small Language Models to Reason , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=. 2023 , address=. doi:10.18653/v1/2023.acl-short.151 , url=

    work page doi:10.18653/v1/2023.acl-short.151 2023

  38. [49]

    Proceedings of the 40th International Conference on Machine Learning , pages=

    Specializing Smaller Language Models towards Multi-Step Reasoning , author=. Proceedings of the 40th International Conference on Machine Learning , pages=. 2023 , publisher=

    2023

  39. [50]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Skip-Thinking: Chunk-wise Chain-of-Thought Distillation Enable Smaller Language Models to Reason Better and Faster , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=. 2025 , address=. doi:10.18653/v1/2025.emnlp-main.610 , url=

    work page doi:10.18653/v1/2025.emnlp-main.610 2025

  40. [51]

    Bioinformatics , year=

    A multimodal deep learning framework for predicting drug-drug interaction events , author=. Bioinformatics , year=

  41. [52]

    AAAI , year=

    CASTER: Predicting Drug Interactions with Chemical Substructure Representation , author=. AAAI , year=

  42. [53]

    Findings of EMNLP , year=

    ExDDI: Explainable Drug-Drug Interaction Prediction with Retrieval-Augmented LLMs , author=. Findings of EMNLP , year=

  43. [54]

    IJCAI , year=

    KGNN: Knowledge Graph Neural Network for Drug-Drug Interaction Prediction , author=. IJCAI , year=

  44. [55]

    IJCAI , year=

    GoGNN: Graph of Graphs Neural Network for Predicting Structured Entity Interactions , author=. IJCAI , year=

  45. [56]

    Bioinformatics , year=

    LaGAT: Link-aware Graph Attention Network for Drug-Drug Interaction Prediction , author=. Bioinformatics , year=

  46. [57]

    Bioinformatics , year=

    MUFFIN: Multi-Scale Feature Fusion for Drug-Drug Interaction Prediction , author=. Bioinformatics , year=

  47. [58]

    WWW , year=

    MIRACLE: Multi-View Graph Contrastive Representation Learning for Drug-Drug Interaction Prediction , author=. WWW , year=

  48. [59]

    Bioinformatics , year=

    SumGNN: Multi-Typed Drug Interaction Prediction via Efficient Knowledge Graph Summarization , author=. Bioinformatics , year=

  49. [60]

    KDD , year=

    TIGER: Transformer over Interaction Graphs for Drug-Drug Interaction , author=. KDD , year=

  50. [61]

    Briefings in Bioinformatics , year=

    DSN-DDI: An Accurate and Generalizable Drug-Drug Interaction Prediction Network , author =. Briefings in Bioinformatics , year=

  51. [62]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Multi-Relational Contrastive Learning Graph Neural Network for Drug-Drug Interaction Event Prediction , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2023 , doi =

    2023

  52. [63]

    IEEE Trans.\ Neural Netw.\ Learn.\ Syst

    MKG-FENN: A Multi-Knowledge-Graph Fingerprint Edge Neural Network for Drug-Drug Interaction , author=. IEEE Trans.\ Neural Netw.\ Learn.\ Syst. , year=

  53. [64]

    Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI) , year =

    Wang, Ziyan and Xiong, Zhankun and Huang, Feng and Liu, Xuan and Zhang, Wen , title =. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI) , year =

  54. [65]

    arXiv preprint arXiv:2602.00539 , year =

    Jin, Xinmo and Fan, Bowen and Li, Xunkai and Sun, Henan and others , title =. arXiv preprint arXiv:2602.00539 , year =

    arXiv

  55. [66]

    2025 , url=

    De Vito, Gabriele and Ferrucci, Filomena and Angelakis, Athanasios , journal=. 2025 , url=

    2025

  56. [67]

    Nature , volume=

    Large Language Models Encode Clinical Knowledge , author=. Nature , volume=. 2023 , doi=

    2023

  57. [68]

    and others , title =

    Knox, Craig and Wilson, Michael and Klinger, Christopher M. and others , title =. Nucleic Acids Research , volume =. 2024 , doi =

    2024

  58. [69]

    Nucleic Acids Research , volume=

    DDInter 2.0: An enhanced drug interaction resource with expanded data coverage, new interaction types, and improved user interface , author=. Nucleic Acids Research , volume=. 2025 , doi=

    2025

  59. [70]

    2025 , eprint=

    Med-PRM: Medical Reasoning Models with Stepwise, Guideline-Verified Process Rewards , author=. 2025 , eprint=

    2025

  60. [71]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  61. [72]

    Qwen2.5 Technical Report , author=

  62. [73]

    DeepSeek-R1: Incentivizing Reasoning Capability via Reinforcement Learning , author=

  63. [74]

    KDD , year=

    XGBoost: A Scalable Tree Boosting System , author=. KDD , year=

  64. [75]

    ACL Findings , year=

    Benchmarking Retrieval-Augmented Generation for Medicine , author =. ACL Findings , year=

  65. [76]

    EMNLP , year=

    Knowledge-Augmented Multimodal Clinical Rationale Generation for Disease Diagnosis with Small Language Models , author=. EMNLP , year=

  66. [77]

    The Reversal Curse:

    Berglund, Lukas and Tong, Meg and Kaufmann, Max and Balesni, Mikita and Stickland, Asa Cooper and Korbak, Tomasz and Evans, Owain , booktitle=. The Reversal Curse:. 2024 , url=

    2024

  67. [78]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    Premise Order Matters in Reasoning with Large Language Models , author=. Proceedings of the 41st International Conference on Machine Learning , pages=. 2024 , publisher=

    2024

  68. [79]

    Findings of the Association for Computational Linguistics:

    Striking a Balance: Alleviating Inconsistency in Pre-trained Models for Symmetric Classification Tasks , author=. Findings of the Association for Computational Linguistics:. 2022 , address=. doi:10.18653/v1/2022.findings-acl.148 , url=

    work page doi:10.18653/v1/2022.findings-acl.148 2022

  69. [80]

    and Dehghani, Morteza , journal=

    Hejabi, Parsa and Rahmati, Elnaz and Ziabari, Alireza S. and Dehghani, Morteza , journal=. Flip-Flop Consistency: Unsupervised Training for Robustness to Prompt Perturbations in. 2025 , url=

    2025

  70. [81]

    ICLR , year=

    Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval , author=. ICLR , year=

  71. [82]

    arXiv:2407.13856 , year=

    Hard Negative Sample-Augmented DPO Post-Training for Small Language Models , author=. arXiv:2407.13856 , year=

    arXiv

  72. [83]

    Proceedings of the 41st International Conference on Machine Learning (ICML) , year =

    Yuan, Weizhe and Pang, Richard Yuanzhe and Cho, Kyunghyun and others , title =. Proceedings of the 41st International Conference on Machine Learning (ICML) , year =

  73. [84]

    NeurIPS , year=

    Adversarial Negatives for Direct Preference Optimization , author=. NeurIPS , year=

  74. [85]

    arXiv:2402.13228 , year=

    Symmetric Direct Preference Optimization , author=. arXiv:2402.13228 , year=

    Pith/arXiv arXiv

  75. [86]

    arXiv:2403.01469 , year=

    MedBench: A Large-Scale Chinese Medical Benchmark , author=. arXiv:2403.01469 , year=

    arXiv

  76. [87]

    ACL , year=

    Statistical Reward Models for Factual Generation , author=. ACL , year=

  77. [88]

    arXiv:2305.10425 , year=

    SLiC-HF: Sequence Likelihood Calibration with Human Feedback , author=. arXiv:2305.10425 , year=

    arXiv

  78. [89]

    ICLR , year=

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author=. ICLR , year=

  79. [90]

    2005 , url =

    Algorithmic Learning in a Random World , author=. 2005 , url =

    2005

  80. [91]

    NeurIPS , year=

    Classification with Valid and Adaptive Coverage , author=. NeurIPS , year=

Showing first 80 references.