Pith. sign in

REVIEW 4 major objections 7 minor 77 references

Walking the Tightrope: Disentangling Beneficial and Detrimental Drifts in Non-Stationary Custom-Tuning

T0 review · 4 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read Counterfactual Preference Optimization keeps medical AI reasoning stable during reinforcement fine-tuning.

desk verdict A useful counterfactual reasoning dataset and a DPO variant that looks strong empirically, but the causal disentanglement story is not actually enforced by the loss. read the letter →

arxiv 2505.13081 v1 pith:E24EJJHU submitted 2025-05-19 cs.LG cs.CV

classification cs.LGcs.CV
keywords conceptdriftreinforcementfine-tuningcounterfactualpreferenceoptimizationchain-of-thoughtreasoningmultimodallargelanguagemodelschestX-raydiagnosiscausalinferenceMIMIC-CXR
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

During reinforcement fine-tuning (RFT), the chain-of-thought reasoning of a multimodal large language model (MLLM) can drift in ways that look harmless locally but corrupt the final diagnosis. This paper claims that this is a form of concept drift: the token-by-token distribution of the reasoning stream changes over the course of tuning, and the change can be either beneficial domain adaptation or detrimental bias. To separate the two, the authors propose Counterfactual Preference Optimization (CPO), which builds a hierarchical graph of chest-radiology concepts, uses a medical language model to generate plausible counterfactual reasoning trajectories from that graph, and optimizes a preference loss that pulls the model toward the true report and away from the counterfactual one. On chest X-ray benchmarks, CPO is reported to improve classification accuracy, report-generation metrics, and zero-shot generalization relative to supervised fine-tuning and existing baselines, and the paper releases the CXR-CounterFact (CCF) dataset of 320,416 counterfactual reasoning trajectories.

What carries the argument

The central object is the counterfactual preference optimization loss $\mathcal{L}_{\mathrm{CPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}_{(v,l,t^+,t^-)}[\log\sigma(\beta\log\frac{\pi_\theta(t^+|v,l)}{\pi_{\mathrm{ref}}(t^+|v,l)} - \beta\log\frac{\pi_\theta(t^-|v,l)}{\pi_{\mathrm{ref}}(t^-|v,l)})]$, where $t^+$ is the preferred radiologist report and $t^-$ is a counterfactual trajectory. The negatives are generated by a medical-domain large language model expert that consults a hierarchical concept graph of 12 pulmonary diseases, 53 attributes, and association, irrelevance, and exclusion relations, perturbing attributes while keeping the narrative radiologically plausible. A structural causal graph with input $X$, reasoning $T$, prediction $Z$, and latent drift $D$ supplies the interpretation: the counterfactual trajectories are interventions on the mediator $T$ that block the backdoor path $X \leftarrow D \to Z$. This object carries the argument because it turns 'detrimental drift' from a metaphor into a training signal.

What would settle it

Run CPO against a control that uses equally fluent negative reports generated without the concept graph (for example, rewriting the same report with a different diagnosis by prompting the same LLM without the graph), keeping all training settings identical. If the graph-free control matches CPO's accuracy on MS-CXR-T and report metrics on MIMIC-CXR, the concept-graph counterfactual mechanism is not carrying the result; if the control falls short, the causal constraints are responsible.

Watch

Extended reading notes

Core claim

The paper's central claim is that the autoregressive chain-of-thought of an MLLM is a non-stationary distribution, so reinforcement fine-tuning should be understood as adaptation to concept drift rather than simple reward maximization. It models each cognitive state as the tokens generated so far together with a latent predicted-outcome distribution, and defines drift as a change in the joint distribution $P_i(t,z)$ across positions. To keep the useful part of adaptation while removing the harmful part, the paper builds a structural causal graph $X \to T \to Z$ with a latent drift node $D$ confounding $X$ and $Z$, and uses the counterfactual quantity $\psi = \mathbb{E}[Z_{T\leftarrow t, D\leftarrow d} - Z_{T\leftarrow t', D\leftarrow d}]$ to justify training against concept-graph-constrained counterfactual trajectories. The resulting CPO loss is a direct-preference-optimization-style objective whose negative examples are not disliked human outputs but causal interventions on the reasoning path. The paper claims that this disentangles beneficial distribution adaptation from detrimental concept drift and that its experiments demonstrate improved robustness, generalization, and coordination in RFT.

Load-bearing premise

The load-bearing assumption is that the counterfactual trajectories produced by the medical language model under the concept graph are faithful examples of detrimental concept drift, so that training the model away from them removes harmful bias without also suppressing useful domain adaptation—if they are simply fluent alternative reports, CPO collapses into ordinary preference optimization and the claimed disentanglement is not demonstrated.

Editorial extensions

If this is right

  • CPO makes reinforced fine-tuning more robust on the chest X-ray classification task: on MS-CXR-T it reports 81.8% top-1 accuracy against 73.8% for supervised fine-tuning alone, with the largest gains on consolidation and pneumonia, the two conditions the paper identifies as drift-prone.
  • The counterfactual preference signal also improves report generation: CPO reports BLEU-4 of 0.155, ROUGE-L of 0.321, and METEOR of 0.236 on MIMIC-CXR, all above the prior baselines listed in the paper, which the authors read as more coherent and complete clinical reasoning.
  • Models tuned with CPO generalize zero-shot: across six chest X-ray benchmarks, CPO's AUC is reported to be highest on every dataset, including Open-I, PadChest, ChestXray14, ChestXpert, and ChestXDet10.
  • The ablations indicate that chain-of-thought and counterfactual preference are complementary: adding only CoT to RFT yields +1.1%, adding only CPO yields +4.5%, and adding both yields +8.0% over SFT on the MS-CXR-T test split.

Reading between the lines

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

  • The same recipe—build a domain concept graph, generate counterfactual reasoning paths, and use them as negative preferences—should transfer to any safety-critical domain where a structured ontology of plausible alternatives exists, such as pathology reports, legal reasoning, or autonomous-driving scene descriptions; the paper demonstrates it only for chest radiology.
  • A direct way to test whether the concept graph is doing the causal work is to ablate it: generate the same number of negative trajectories from the same LLM without graph constraints. If performance does not drop, CPO may simply be a stronger form of preference optimization; if it drops, the graph's relational constraints are what separate beneficial adaptation from detrimental drift.
  • The formal equivalence between concept drift and CoT token streams suggests that drift detectors could monitor reasoning during fine-tuning directly, flagging the moment when $P_i(t,z)\neq P_{i+1}(t,z)$ along a trajectory; the paper does not build such a detector, but its definition makes one straightforward to construct.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper studies reinforcement fine-tuning (RFT) of multimodal large language models (MLLMs) in non-stationary medical settings. It formalizes autoregressive chain-of-thought (CoT) generation as a concept-drift process, introduces a structural causal model {X, Z, T, D} in which a latent drift D confounds inputs and predictions through a backdoor path, and defines a counterfactual effect ψ (Eq. 4) that would isolate the front-door path X→T→Z. To operationalize this, the authors construct a concept graph of 12 disease entities and 53 attributes from MIMIC-CXR using Med-PaLM, generate 320,416 counterfactual reasoning trajectories (the CCF dataset), and train Qwen2.5-VL 7B with Counterfactual Preference Optimization (CPO, Eq. 8), which is a DPO-style loss over positive human trajectories t+ and counterfactual negatives t−. Experiments report gains over prior art on MS-CXR-T classification, MIMIC-CXR report generation, and zero-shot transfer to six chest X-ray benchmarks, plus an ablation isolating CoT and CPO.

Significance. If the disentanglement claim were established, the paper would offer a practical way to stabilize RFT for safety-critical domains, and the CCF dataset would be a reusable resource. The empirical results are large and consistent across several external benchmarks, which is encouraging and goes beyond circular evaluation. However, the paper's central scientific claim—that CPO separates beneficial domain adaptation from detrimental concept drift—is not presently supported: the optimized loss is standard DPO over generated negatives, the defined causal effect is never estimated, and comparisons to prior work are not controlled. The dataset construction is also not validated for clinical plausibility. With additional estimation or validation and controlled baselines, the contribution could become significant; in its current form it is an interesting but not yet demonstrated causal claim.

major comments (4)
  1. [Sec. 2.2, Sec. 2.4, Eq. (4), Eq. (8)] The defined causal effect ψ is never estimated or used. The loss L_CPO in Eq. (8) is exactly the standard DPO loss with t− produced by Med-PaLM; it contains no do-operator, no backdoor adjustment, and no term involving ψ. Thus the decoupling of beneficial and detrimental drift is not enforced by the optimization objective; it is delegated entirely to the data-generation step. The paper must either estimate and use ψ, or provide independent evidence that the generated t− remove harmful drift while preserving useful domain adaptation; without this, the central claim of the paper is unsupported.
  2. [Sec. 3, Tables 2-5, Appendix C] All reported results are from single training runs; there are no error bars, confidence intervals, or significance tests. The comparisons mix architectures and training data: Table 2 compares a 7B MLLM against specialist vision-language encoders with different backbones and corpora, Table 4 compares against CLIP-based zero-shot models, and Table 3 compares against report generators that do not use CoT. Without same-backbone and same-data baselines (e.g., Qwen2.5-VL SFT/RFT/DPO under identical settings), the large observed margins could be confounded by base-model strength rather than by CPO. The ablation in Table 5 is the most informative comparison but also lacks seed variance and a clear specification of what negatives are used in the DPO-only condition.
  3. [Sec. 2.3, Appendix B] The validity of the counterfactual trajectories is the load-bearing assumption of the method, yet it is not verified. There is no clinician evaluation, no inter-annotator agreement, no automated coherence or plausibility metric, and no analysis of how often the concept-graph perturbation corresponds to actual radiological contradictions. If these trajectories are merely plausible alternate reports, the method reduces to DPO with model-generated hard negatives, and the stated disentanglement is unsupported. The authors should validate the counterfactual generation process, for example by human expert ratings or by measuring whether the generated negatives systematically flip the diagnosis while preserving the original image evidence.
  4. [Sec. 3.2, Table 3] The report generation metrics improve, but the relation to drift disentanglement remains correlational. The METEOR gain of 34.8% may reflect lexical differences between counterfactual and positive texts rather than reasoning stability; a direct measure of drift—such as token distribution shift during training, or diagnostic consistency of the CoT before and after RFT—is missing. The paper should quantify concept drift directly to substantiate the central phenomenon (Observation 1.1) and its mitigation, since the current metrics only measure output quality, not drift disentanglement.
minor comments (7)
  1. [Abstract, Sec. 3] The abstract and Sec. 3 claim 'statistically significant improvements', but no statistical tests are reported anywhere in the manuscript.
  2. [Sec. 2.2, Eq. (4)] The symbol ψ is defined but never used again in the paper; clarify its role in the derivation or remove it to avoid implying that the causal effect is estimated.
  3. [Sec. 2.4] There is a notation inconsistency: the state is defined as s_j, but the text later says 'a0 is usually the token <think>'; this should be s0 or the state-action notation should be unified.
  4. [Table 3] The venue for AlignTrans is listed as 'MICCAI’2'; this should be 'MICCAI’21' (or the full citation).
  5. [Author affiliation] The affiliation contains a typo: 'Faulty of Engineering' should be 'Faculty of Engineering'.
  6. [Appendix B] The prompt text says 'chest DR examination'; this is likely intended as 'chest X-ray (DR)' or 'chest radiograph', but the expression is ambiguous.
  7. [References] Several references are incomplete: for example, references [56] and [57] lack venue information, and some reference strings are cut off.

Circularity Check

1 steps flagged · score 4.0 of 10

CPO's disentanglement claim reduces to DPO against self-generated counterfactual negatives; the causal effect ψ in Eq. 4 is never used.

  1. self definitional [Section 2.4, Eq. (8); Introduction]
    "the generated counterfactual CoT is represented by t−. ... LCPO(πθ;πref) = −E(v,l,t+,t−) [log σ(β log πθ(t+|v,l)/πref(t+|v,l) − β log πθ(t−|v,l)/πref(t−|v,l))] ... it culminates in counterfactual reinforced custom-tuning, an adaptive framework that effectively differentiates between advantageous domain adaptation and harmful concept drift."

    The CPO loss is structurally identical to DPO: it only increases likelihood of t+ and decreases likelihood of t−. The counterfactual causal effect ψ from Eq. 4 never appears in the objective, and no do-operator or backdoor adjustment is applied. The only CPO-specific ingredient is the selection of t−, generated by Med-PaLM under the authors' concept graph and implicitly treated as 'detrimental concept drift.' Because the paper equates 'harmful concept drift' with these generated counterfactual trajectories and then trains the model to reject them, the claimed 'decoupling between beneficial domain adaptation and detrimental concept drift' is true by construction relative to that operationalization rather than by measurement or causal estimation.

full rationale

The empirical evaluation is genuinely external: MS-CXR-T classification, MIMIC-CXR report generation, and zero-shot benchmarks (Open-I, PadChest, ChestXray14, etc.) compare against published baselines, so the headline performance claims are not circular. There is no load-bearing self-citation: the authors' prior concept-drift works appear mainly in related work, and the method does not depend on an unverified uniqueness theorem. The circularity concern is narrower but real: the paper's central conceptual claim—that CPO disentangles beneficial adaptation from detrimental drift—is asserted immediately after writing down a DPO loss whose only novel input is the counterfactual negative trajectories. Since those trajectories are called 'detrimental concept drift' and the loss simply lowers their probability, the disentanglement result reduces to a definitional identification between the generated negatives and the harmful drift, rather than to the causal framework in Eq. 4. This warrants a moderate score: the empirical results stand independently, but the paper's advertised mechanism is not enforced by the objective and is instead imported through the data-generation assumption.

Assumptions & free parameters 2 free parameters · 4 assumptions · 2 invented entities

The central contribution relies on a modeling choice (token streams as drifting distributions), an asserted causal graph, an assumed validity of Med-PaLM-generated counterfactuals, and standard DPO assumptions. The only explicit free hyperparameter is β; the concept graph structure is a design choice. No parameters are fitted to the evaluation data, so the empirical benchmark results are not circular in the narrow sense.

free parameters (2)
  • β (DPO temperature) = not stated in main text
    Controls the strength of deviation from the reference policy in Eq.6 to Eq.8 and therefore scales the preference loss. No value or ablation is reported, so the reported results may depend on an unstated choice.
  • Concept graph structure (12 disease entities, 53 attributes, 3 relation types) = not stated as a fitted value
    The counterfactual generation pipeline in Section 2.3 uses these procedurally chosen entities and relations to constrain negative samples. The counts are outputs of semantic parsing by Med-PaLM, and no analysis shows sensitivity of CPO gains to this structure.
assumptions (4)
  • domain assumption Autoregressive decoding of CoT can be modeled as a token stream with a joint distribution P_i(t,z) over tokens and predicted results (Definition 2.1).
    The concept drift formalism in Eq.2 assumes this stream representation and that consecutive steps differ. This is a modeling choice, not proven.
  • ad hoc to paper The structural causal graph {X,Z,T,D} with D as a confounder and T as a mediator is correct (Section 2.2, Fig.3).
    No identifiability conditions or estimation of the counterfactual effect ψ in Eq.4 are provided; the causal role of D is asserted.
  • ad hoc to paper Med-PaLM with the extracted concept graph produces counterfactual trajectories that isolate detrimental concept drift while preserving plausible radiology (Section 2.3, Appendix B).
    The validity of the negative samples is assumed; there is no external clinical validation or measurement of drift removal.
  • domain assumption Bradley-Terry preference model and DPO assumptions hold for pairs (t+, t-) of original versus counterfactual reasoning (Section 2.4, Eq.8).
    The CPO loss is DPO applied to counterfactual pairs, requiring the usual preference optimization assumptions.
invented entities (2)
  • Latent concept drift D in the structural causal model
    purpose: Explains distributional shift in CoT during RFT and motivates the counterfactual intervention.
    D is introduced as a latent confounder and mediator in Fig.3 but is never directly measured or independently detected; all evidence is indirect through generated counterfactuals.
  • CXR-CounterFact (CCF) dataset independent evidence
    purpose: Provides 320,416 counterfactual reasoning trajectories for training and benchmarking.
    Public release of the dataset would be an independent resource, but the main text does not give a stable commit hash or data link, and generation relies on the proprietary Med-PaLM model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Walking the Tightrope: Disentangling Beneficial and Detrimental Drifts in Non-Stationary Custom-Tuning." pith.science (2026). https://pith.science/paper/E24EJJHU

@misc{pith2026250513081,
  author       = {Pith},
  title        = {Pith review of: Walking the Tightrope: Disentangling Beneficial and Detrimental Drifts in Non-Stationary Custom-Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E24EJJHU}},
  note         = {Machine review of arXiv:2505.13081}
}
read the original abstract

This paper uncovers a critical yet overlooked phenomenon in multi-modal large language models (MLLMs): detrimental concept drift within chain-of-thought (CoT) reasoning during non-stationary reinforcement fine-tuning (RFT), where reasoning token distributions evolve unpredictably, thereby introducing significant biases in final predictions. To address this, we are pioneers in establishing the theoretical bridge between concept drift theory and RFT processes by formalizing CoT's autoregressive token streams as non-stationary distributions undergoing arbitrary temporal shifts. Leveraging this framework, we propose a novel counterfact-aware RFT that systematically decouples beneficial distribution adaptation from harmful concept drift through concept graph-empowered LLM experts generating counterfactual reasoning trajectories. Our solution, Counterfactual Preference Optimization (CPO), enables stable RFT in non-stationary environments, particularly within the medical domain, through custom-tuning of counterfactual-aware preference alignment. Extensive experiments demonstrate our superior performance of robustness, generalization and coordination within RFT. Besides, we also contributed a large-scale dataset CXR-CounterFact (CCF), comprising 320,416 meticulously curated counterfactual reasoning trajectories derived from MIMIC-CXR. Our code and data are public.

Figures

Figures reproduced from arXiv: 2505.13081 by the authors.

Figure 1
Figure 1. Concept Drift in RFT’s rea￾soning for chest diagnosis. Despite analogous occurrence probabilities of "lung opacity" (in red) and "opacity" (in blue) tokens during the CoT, non￾stationarity induces significant bad dis￾tributional drift in clinical conclusions, especially the opposite diagnosis of at￾electasis, cardiomegaly and pneumonia. Intuitively, we provide a representative case study that demonstrates this pheno… view at source ↗
Figure 2
Figure 2. The main contributions of our methods. (a) By formalizing autoregressive CoT generation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Structural Causal Graph. X: Inputs, Z: Prediction Results, T: Chain￾of-Thought, and D: Latent Concept Drift within CoT under Non-stationary Rein￾forced Custom-Tuning. The optimization objective formalized Eq.3 necessitates disentanglement of two competing goals: advantageous policy-induced domain adaptation and versus pathological concept drift arising from suboptimal policy execution, which are both sampled from po… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Non-stationarity of MIMIC￾CXR with its percentage of diseases. Blue signifies patients with clinically confirmed diagnoses showing the long￾tailed characteristic, while red demar￾cates suspected cases emphasizing the inherent uncertainty within medicine. MIMIC-CXR[11] …
Figure 5
Figure 5. Figure 5: Samples of CXR-CounterFact (CCF) Dataset. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

77 extracted references · 55 canonical work pages

  1. [1]

    Zhang, Z

    Trung, L., X. Zhang, Z. Jie, et al. ReFT: Reasoning with Reinforced Fine-Tuning. In L.-W. Ku, A. Martins, V . Srikumar, eds.,Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume1: Long Papers), pages 7601–7614. Association for Computational Linguistics, Bangkok, Thailand, 2024

  2. [2]

    Liu, Z., Z. Sun, Y . Zang, et al. Visual-RFT: Visual Reinforcement Fine-Tuning, 2025

  3. [3]

    Donahue, P

    Alayrac, J.-B., J. Donahue, P. Luc, et al. Flamingo: A Visual Language Model for Few-Shot Learning. Advances in Neural Information Processing Systems, 35:23716–23736, 2022

  4. [4]

    Bai, J., S. Bai, S. Yang, et al. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023

  5. [5]

    Dai, W., J. Li, D. Li, et al. InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning, 2023

  6. [6]

    Lu, J., A. Liu, F. Dong, et al. Learning under Concept Drift: A Review. 31(12):2346–2363, 2019

  7. [7]

    Yang, X., J. Lu, E. Yu. Adapting multi-modal large language model to concept drift from pre-training onwards. arXiv preprint arXiv:2405.13459, 2025

  8. [8]

    Causal diagrams for empirical research

    Pearl, J. Causal diagrams for empirical research. Biometrika, 82(4):669–688, 1995

Show all 77 references
  1. [9]

    Causal inference in statistics: a primer

    —. Causal inference in statistics: a primer. John Wiley & Sons, 2016

  2. [10]

    Yang, X., J. Lu, E. Yu. Causal-informed contrastive learning: Towards bias-resilient pre-training under concept drift. arXiv preprint arXiv:2502.07620, 2025

  3. [11]

    Johnson, A. E., T. J. Pollard, S. J. Berkowitz, et al. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019

  4. [12]

    Direct and indirect effects

    Pearl, J. Direct and indirect effects. In Probabilistic and causal inference: the works of Judea Pearl, pages 373–392. 2022

  5. [13]

    Zhang, X., C. Wu, Y . Zhang, et al. Knowledge-enhanced visual-language pre-training on chest radiology images. 14(1):4542, 2023

  6. [14]

    Zhang, C

    Zhou, X., X. Zhang, C. Wu, et al. Knowledge-enhanced Visual-Language Pretraining for Computational Pathology, 2024

  7. [15]

    Azizi, T

    Singhal, K., S. Azizi, T. Tu, et al. Large language models encode clinical knowledge. 620(7972):172–180, 2023

  8. [16]

    Singhal, K., T. Tu, J. Gottweis, et al. Toward expert-level medical question answering with large language models. 31(3):943–950, 2025

  9. [17]

    Sharma, E

    Rafailov, R., A. Sharma, E. Mitchell, et al. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. 36:53728–53741, 2023

  10. [18]

    Qwen2.5-vl, 2025

    Team, Q. Qwen2.5-vl, 2025

  11. [19]

    Hyland, Q

    Bannur, S., S. Hyland, Q. Liu, et al. Learning to exploit temporal structure for biomedical vision-language processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15016–15027. 2023

  12. [20]

    Karwande, G., A. B. Mbakwe, J. T. Wu, et al. Chexrelnet: An anatomy-aware model for tracking longitudinal relationships between chest x-rays. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 581–591. Springer, 2022. 10

  13. [21]

    Usuyama, S

    Boecking, B., N. Usuyama, S. Bannur, et al. Making the most of text semantics to improve biomedical vision–language processing. In European conference on computer vision, pages 1–21. Springer, 2022

  14. [22]

    Yang, J., B. Su, X. Zhao, et al. Unlocking the power of spatial and temporal information in medical multimodal pre-training. In Forty-first International Conference on Machine Learning. 2024

  15. [23]

    Yang, Z., L. Shen. Tempa-vlp: Temporal-aware vision-language pretraining for longitudinal exploration in chest x-ray image. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV), pages 4625–4634. 2025

  16. [24]

    Chen, Y ., S. Xu, A. Sellergren, et al. Coca-cxr: Contrastive captioners learn strong temporal structures for chest x-ray vision-language understanding, 2025

  17. [25]

    Hyland, F

    Bannur, S., S. Hyland, F. Liu, et al. Learning to exploit temporal structure for biomedical vision-language processing. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2023

  18. [26]

    Wu, J. T., N. N. Agu, I. Lourentzou, et al. Chest imagenome dataset for clinical reasoning. arXiv preprint arXiv:2108.00316, 2021

  19. [27]

    Song, T.-H

    Chen, Z., Y . Song, T.-H. Chang, et al. Generating radiology reports via memory-driven trans- former. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1439–1449. 2020

  20. [28]

    Liu, F., X. Wu, S. Ge, et al. Exploring and distilling posterior and prior knowledge for radiology report generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13753–13762. 2021

  21. [29]

    You, D., F. Liu, S. Ge, et al. Aligntransformer: Hierarchical alignment of visual regions and dis- ease tags for medical report generation. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 2...

  22. [30]

    Liu, F., S. Ge, X. Wu. Competence-based multimodal curriculum learning for medical report generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (V olume1...

  23. [31]

    Yan, B., M. Pei. Clinical-bert: Vision-language pre-training for radiograph diagnosis and reports generation. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, pages 2982–2990. 2022

  24. [32]

    Wang, Z., L. Liu, L. Wang, et al. Metransformer: Radiology report generation by trans- former with multiple learnable expert tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11558–11567. 2023

  25. [33]

    Li, M., B. Lin, Z. Chen, et al. Dynamic graph enhanced contrastive learning for chest x- ray report generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3334–3343. 2023

  26. [34]

    Wang, Z., L. Liu, L. Wang, et al. R2gengpt: Radiology report generation with frozen llms. Meta-Radiology, 1(3):100033, 2023

  27. [35]

    Jin, H., H. Che, Y . Lin, et al. Promptmrg: Diagnosis-driven prompts for medical report generation. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pages 2607–2615. 2024

  28. [36]

    Liu, C., Y . Tian, W. Chen, et al. Bootstrapping large language models for radiology report generation. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pages 18635–18643. 2024

  29. [37]

    Wang, X., F. Wang, Y . Li, et al. CXPMRG-Bench: Pre-training and Benchmarking for X-ray Medical Report Generation on CheXpert Plus Dataset, 2024

  30. [38]

    Wang, Z., Z. Wu, D. Agarwal, et al. Medclip: Contrastive learning from unpaired medical images and text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing, vol. 2022, page 3876. 2022. 11

  31. [39]

    Zhang, S., Y . Xu, N. Usuyama, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023

  32. [40]

    Huang, S.-C., L. Shen, M. P. Lungren, et al. Gloria: A multimodal global-local representa- tion learning framework for label-efficient medical image recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3942–3951. 2021

  33. [41]

    Talius, P

    Tiu, E., E. Talius, P. Patel, et al. Expert-level detection of pathologies from unannotated chest x-ray images via self-supervised learning. Nature biomedical engineering, 6(12):1399–1406, 2022

  34. [42]

    Zhang, Y

    Wu, C., X. Zhang, Y . Zhang, et al. Medklip: Medical knowledge enhanced language-image pre-training in radiology, 2023

  35. [43]

    Zhang, X., C. Wu, Y . Zhang, et al. Knowledge-enhanced visual-language pre-training on chest radiology images. Nature Communications, 14(1):4542, 2023

  36. [44]

    Lai, H., Q. Yao, Z. Jiang, et al. Carzero: Cross-attention alignment for radiology zero-shot classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11137–11146. 2024

  37. [45]

    Antani, M

    Demner-Fushman, D., S. Antani, M. Simpson, et al. Design and development of a multimodal biomedical information retrieval system. Journal of Computing Science and Engineering, 6(2):168–177, 2012

  38. [46]

    Pertusa, J.-M

    Bustos, A., A. Pertusa, J.-M. Salinas, et al. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis, 66:101797, 2020

  39. [47]

    Wang, X., Y . Peng, L. Lu, et al. Chestx-ray8: Hospital-scale chest x-ray database and bench- marks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017

  40. [48]

    Rajpurkar, M

    Irvin, J., P. Rajpurkar, M. Ko, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, vol. 33, pages 590–597. 2019

  41. [49]

    Liu, J., J. Lian, Y . Yu. Chestx-det10: Chest x-ray dataset on detection of thoracic abnormalities, 2020

  42. [50]

    Lu, J., A. Liu, Y . Song, et al. Data-driven decision support under concept drift in streamed big data. Complex & intelligent systems, 6(1):157–163, 2020

  43. [51]

    Xiong, A

    Wang, K., L. Xiong, A. Liu, et al. A self-adaptive ensemble for user interest drift learning. 577:127308, 2024

  44. [52]

    Jiao, B., Y . Guo, D. Gong, et al. Dynamic Ensemble Selection for Imbalanced Data Streams With Concept Drift. 35(1):1278–1291, 2024

  45. [53]

    Cerqueira, V ., H. M. Gomes, A. Bifet, et al. STUDD: A student–teacher method for unsupervised concept drift detection. 112(11):4351–4378, 2023

  46. [54]

    Yang, X., Y . Chen, X. Yue, et al. T-distributed Spherical Feature Representation for Imbalanced Classification. Proceedings of the AAAI Conference on Artificial Intelligence, 37(9):10825– 10833, 2023

  47. [55]

    Yu, E., J. Lu, B. Zhang, et al. Online boosting adaptive learning under concept drift for multistream classification. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pages 16522–16530. 2024

  48. [56]

    Yu, E., Y . Song, G. Zhang, et al. Learn-to-adapt: Concept drift adaptation for hybrid multiple streams. 496:121–130, 2022

  49. [57]

    Yu, H., W. Liu, J. Lu, et al. Detecting group concept drift from multiple data streams. 134:109113, 2023

  50. [58]

    Li, W., X. Yang, W. Liu, et al. DDG-DA: Data Distribution Generation for Predictable Concept Drift Adaptation. 36(4):4092–4100, 2022-06-28

  51. [59]

    Zhang, J

    Yang, X., H. Zhang, J. Cai. Deconfounded Image Captioning: A Causal Retrospect. 45(11):12996–13010, 2023. 12

  52. [60]

    Liu, B., D. Wang, X. Yang, et al. Show, Deconfound and Tell: Image Captioning With Causal Inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18041–18050. 2022

  53. [61]

    Zhang, D

    Zhang, C., L. Zhang, D. Zhou. Causal Walk: Debiasing Multi-Hop Fact Verification with Front-Door Adjustment, 2024

  54. [62]

    Jeong, H

    Choi, S., M. Jeong, H. Han, et al. C2L: Causally Contrastive Learning for Robust Text Classification. 36(10):10526–10534, 2022

  55. [63]

    Rohekar, R. Y ., Y . Gurwicz, S. Nisimov. Causal Interpretation of Self-Attention in Pre-Trained Transformers. 36:31450–31465, 2023

  56. [64]

    Zhang, G

    Yang, X., H. Zhang, G. Qi, et al. Causal Attention for Vision-Language Tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9847–9857. 2021

  57. [65]

    Yang, X., L. Xu, H. Li, et al. Masked Image Contrastive Learning for Efficient Visual Conceptual Pre-training. arXiv preprint arXiv:2411.09858, 2024

  58. [66]

    Christiano, P. F., J. Leike, T. Brown, et al. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017

  59. [67]

    Ouyang, L., J. Wu, X. Jiang, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744, 2022

  60. [68]

    Kalai, A

    Jaech, A., A. Kalai, A. Lerer, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

  61. [69]

    Kadavath, S

    Bai, Y ., S. Kadavath, S. Kundu, et al. Constitutional ai: Harmlessness from ai feedback.arXiv preprint arXiv:2212.08073, 2022

  62. [70]

    Guo, D., D. Yang, H. Zhang, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

  63. [71]

    Shao, Z., P. Wang, Q. Zhu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

  64. [72]

    Gulcehre, C., T. L. Paine, S. Srinivasan, et al. Reinforced self-training (rest) for language modeling, 2023

  65. [73]

    Yuan, Z., H. Yuan, C. Li, et al. Scaling relationship on learning mathematical reasoning with large language models, 2024

  66. [74]

    Huang, L

    Zeng, W., Y . Huang, L. Zhao, et al. B-STar: Monitoring and balancing exploration and exploitation in self-taught reasoners. In The Thirteenth International Conference on Learning Representations. 2025

  67. [75]

    Zheng, Y

    Zhang, Z., C. Zheng, Y . Wu, et al. The lessons of developing process reward models in mathematical reasoning. arXiv preprint arXiv:2501.07301, 2025

  68. [76]

    Yang, X., L. Xu, H. Sun, et al. Enhancing visual grounding and generalization: A multi-task cycle training approach for vision-language models. arXiv preprint arXiv:2311.12327, 2024. 13 Appendix A Related Works A.1 Concept Drift In their survey spanning multiple studies, Lu et...

  69. [77]

    A.2 Causal Inference Recently, increasing researchers have incorporated causal inference into deep-learning models, espe- cially in large models

    establishes a teacher-student discrepancy framework that leverages predictive consistency analysis to enable label-agnostic drift identification while maintaining detection sensitivity, thereby addressing practical deployment constraints in evolving data environments. A.2 Caus...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.