Pith. sign in

REVIEW 41 references

Truth or Twist? Optimal Model Selection for Reliable Label Flipping Evaluation in LLM-based Counterfactuals

T0 review · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Independent, non-fine-tuned judge models align most closely with human label-flip judgments, yet all automated judges fall well short of human evaluation.

arxiv 2505.13972 v3 pith:3W5OQ2UE submitted 2025-05-20 cs.CL

Qianli Wang , Van Bach Nguyen , Nils Feldhus , Luis Felipe Villa-Arenas , Christin Seifert , Sebastian Möller , Vera Schmitt This is my paper · ORCID
classification cs.CL
keywords modelsjudgemodelgeneratorcounterfactualflippinglabeluser
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

Counterfactual examples are slightly modified versions of a text that change its correct answer, like turning a positive movie review into a negative one. Many teams use them to train more robust AI models, a process called counterfactual data augmentation. Before using a counterfactual, the team must check whether the ground-truth label actually flipped. Because checking by hand is expensive, teams often ask a large language model (LLM) to act as judge. This paper asks which judge model gives the most reliable answers.

The authors defined four relationships between the model that generates the counterfactual and the judge model: the same model, a model from the same family, an independent model, and a distilled model. For each of four generators and three datasets, they computed the label flip rate reported by 15 judge models and compared it with labels from 90 human annotators. The clearest result: judge models that are independent of the generator and not fine-tuned on the target dataset agreed best with humans. Judge models that were fine-tuned on the dataset, or that came from the same family as the generator, agreed worst. The authors also show that a judge model with high classification accuracy is not necessarily a reliable label-flip judge.

However, even the best judge models missed the human label by about 23 percentage points on average. The authors argue that fully automated counterfactual validation is not yet good enough and that human oversight is still needed. The study also reports that adding counterfactuals with judge-provided labels can improve model performance, though this part of the evidence is mixed.

Extended reading notes

Core claim

The central claim is that 'judge models with an independent, non-fine-tuned relationship to the generator model provide the most reliable label flipping evaluations' (Abstract). In the experiments, the independent non-fine-tuned relationship (Rimwo) achieves the highest human alignment with average rank 4.15, most-to-least ratio 3.5, and Pearson correlation 0.47, beating same-model, same-family, distilled, and fine-tuned independent judges. If the paper is correct, CDA practitioners should choose an independent non-fine-tuned judge model to validate counterfactual label flips.

Load-bearing premise

The paper's ranking of judge relationships rests on the user study as ground truth: 45 counterfactuals per dataset (two annotators each, Cohen's kappa = 0.55) are used to compute the human label flip rate, and all alignment metrics are differences from this rate. If those 45 items are unrepresentative of the full counterfactual distribution, or if the moderate human agreement makes the gold standard too noisy, the conclusion that Rimwo is best could change. The Appendix G.3 sanity check compares LFR on subsets but does not establish statistical equivalence.

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.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper fits no free parameters; its conclusions rest on the user study as external benchmark, on the stipulated relationship taxonomy, and on a small sample of 45 counterfactuals per dataset. No new entities are postulated.

assumptions (3)
  • domain assumption Human annotations from the user study are an unbiased gold standard for label flipping.
    All alignment metrics (rank, rm/l, rho) are defined relative to the user-study LFR, so the entire result inherits from the quality of these annotations. Inter-annotator agreement is only moderate (Cohen's kappa = 0.55), which limits the reliability of this gold standard. See Section 4.4.
  • domain assumption The four relationship categories (same, same-family, independent with/without fine-tuning, distilled) are the relevant factors for judge model selection.
    The taxonomy is stipulated in Section 3.2 and not derived from data; if a relevant fifth factor governs judge reliability, the conclusions may change.
  • domain assumption The 45 counterfactuals sampled per dataset are representative of all generated counterfactuals.
    User study LFR and the resulting rankings are computed from 45 items per generator-dataset pair. The sanity check in Appendix G.3 compares subset LFR to full LFR but does not provide a statistical test of representativeness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Truth or Twist? Optimal Model Selection for Reliable Label Flipping Evaluation in LLM-based Counterfactuals." pith.science (2026). https://pith.science/paper/3W5OQ2UE

@misc{pith2026250513972,
  author       = {Pith},
  title        = {Pith review of: Truth or Twist? Optimal Model Selection for Reliable Label Flipping Evaluation in LLM-based Counterfactuals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3W5OQ2UE}},
  note         = {Machine review of arXiv:2505.13972}
}
read the original abstract

Counterfactual examples are widely employed to enhance the performance and robustness of large language models (LLMs) through counterfactual data augmentation (CDA). However, the selection of the judge model used to evaluate label flipping, the primary metric for assessing the validity of generated counterfactuals for CDA, yields inconsistent results. To decipher this, we define four types of relationships between the counterfactual generator and judge models: being the same model, belonging to the same model family, being independent models, and having an distillation relationship. Through extensive experiments involving two state-of-the-art LLM-based methods, three datasets, four generator models, and 15 judge models, complemented by a user study (n = 90), we demonstrate that judge models with an independent, non-fine-tuned relationship to the generator model provide the most reliable label flipping evaluations. Relationships between the generator and judge models, which are closely aligned with the user study for CDA, result in better model performance and robustness. Nevertheless, we find that the gap between the most effective judge models and the results obtained from the user study remains considerably large. This suggests that a fully automated pipeline for CDA may be inadequate and requires human intervention.

Figures

Figures reproduced from arXiv: 2505.13972 by the authors.

Figure 1
Figure 1. A counterfactual generated by Llama3-8B, with its label evaluated by judge models with different relationships, complemented by human evaluation. The revised words are highlighted in red. between the generator and judge models that are most aligned with the user study, lead to improved model performance and robustness. Additionally, there remains a considerable gap between the performance of the best judge models an… view at source ↗
Figure 2
Figure 2. The average ranking of judge-generator model [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Average ranking and most-to-least ratio for [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Counterfactual Data Augmentation pipeline. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Label distributions of AG News, SNLI and SST2. creasing number of demonstrations does not yield significant benefits and, in some cases, even de￾grades performance for the SST2 dataset. There￾fore, zero-shot prompting is employed for all other decoder-only LLMs as well…
Figure 6
Figure 6. Figure 6: Dataset Examples from AG News, SST2 and SNLI. Such discrepancies may explain observed patterns in the automatic evaluation ( [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Classification performance of models on the [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Classification performance of models on the [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Classification performance of models on the [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Prompt instruction for classification on [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Annotation guidelines for the user study [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: An example from AG News, its corresponding counterfactual generated by FIZLE, and the chain-of￾thought from DeepSeek-R1-Distill-Llama-70B. Underlines indicate the insertion of new words, compared to the original input. Model Relation. AG News SST2 SNLI Avg. Llama3-8B …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 15 canonical work pages

  1. [1]

    Hewett, Mojan Javaheripi, Piero Kauffmann, James R

    Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J. Hewett, Mojan Javaheripi, Piero Kauffmann, James R. Lee, Yin Tat Lee, Yuanzhi Li, Weishung Liu, Caio C. T. Mendes, Anh Nguyen, Eric Price, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Xin Wang, Rachel Ward, Yue Wu, Dingli Yu,...

    arXiv 2024

  2. [2]

    Aryan Agrawal, Lisa Alazraki, Shahin Honarvar, Thomas Mensink, and Marek Rei. 2025. https://openreview.net/forum?id=abllmCsDp8 Enhancing LLM robustness to perturbed instructions: An empirical study . In ICLR 2025 Workshop on Building Trust in Language Models and Applications

    work page 2025

  3. [3]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  4. [4]

    Ananth Balashankar, Xuezhi Wang, Yao Qin, Ben Packer, Nithum Thain, Ed Chi, Jilin Chen, and Alex Beutel. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.10 Improving classifier robustness through active generative counterfactual data augmentation . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 127--139, Singapore. A...

    work page doi:10.18653/v1/2023.findings-emnlp.10 2023

  5. [5]

    Amrita Bhattacharjee, Raha Moraffah, Joshua Garland, and Huan Liu. 2024 a . https://doi.org/10.1109/BigData62323.2024.10825537 Zero-shot LLM-guided Counterfactual Generation: A Case Study on NLP Model Evaluation . In 2024 IEEE International Conference on Big Data (BigData), pages 1243--1248, Los Alamitos, CA, USA. IEEE Computer Society

    arXiv 2024

  6. [6]

    Amrita Bhattacharjee, Raha Moraffah, Joshua Garland, and Huan Liu. 2024 b . Towards llm-guided causal explainability for black-box text classifiers. In AAAI 2024 Workshop on Responsible Language Models, Vancouver, BC, Canada

    work page 2024

  7. [7]

    Bowman, Gabor Angeli, Christopher Potts, and Christopher D

    Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. https://doi.org/10.18653/v1/D15-1075 A large annotated corpus for learning natural language inference . In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632--642, Lisbon, Portugal. Association for Computational Linguistics

    doi:10.18653/v1/d15-1075 2015

  8. [8]

    Zeming Chen, Qiyue Gao, Antoine Bosselut, Ashish Sabharwal, and Kyle Richardson. 2023. https://doi.org/10.18653/v1/2023.acl-long.302 DISCO : Distilling counterfactuals with large language models . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5514--5528, Toronto, Canada. Associati...

    doi:10.18653/v1/2023.acl-long.302 2023

Show all 41 references
  1. [9]

    DeepSeek-AI. 2025. https://arxiv.org/abs/2501.12948 Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning . Preprint, arXiv:2501.12948

    2025 arXiv

  2. [10]

    Zahra Dehghanighobadi, Asja Fischer, and Muhammad Bilal Zafar. 2025. https://arxiv.org/abs/2502.18156 Can llms explain themselves counterfactually? Preprint, arXiv:2502.18156

    2025

  3. [11]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

    2019 doi

  4. [12]

    Tanay Dixit, Bhargavi Paranjape, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.216 CORE : A retrieve-then-edit framework for counterfactual data generation . In Findings of the Association for Computational Linguistics: EMNLP ...

    2022 doi

  5. [13]

    Marharyta Domnich, Julius Välja, Rasmus Moorits Veski, Giacomo Magnifico, Kadi Tulver, Eduard Barbu, and Raul Vicente. 2025. https://doi.org/10.1609/aaai.v39i15.33791 Towards unifying evaluation of counterfactual explanations: Leveraging large language models for human-centric...

    2025 doi

  6. [14]

    Nils Feldhus, Qianli Wang, Tatiana Anikina, Sahil Chopra, Cennet Oguz, and Sebastian M \"o ller. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.359 I nterro L ang: Exploring NLP models and datasets through dialogue-based explanations . In Findings of the Association for...

    2023 doi

  7. [15]

    Xiaoli Fern and Quintin Pope. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.452 Text counterfactuals via latent optimization and S hapley-guided search . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5578--5593, Online and...

    2021 doi

  8. [16]

    Matt Gardner, Yoav Artzi, Victoria Basmova, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, et al. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.117 Evaluating models' local decision boundaries via contrast sets . ...

    2020 doi

  9. [17]

    Yingqiang Ge, Shuchang Liu, Zelong Li, Shuyuan Xu, Shijie Geng, Yunqi Li, Juntao Tan, Fei Sun, and Yongfeng Zhang. 2021. https://arxiv.org/abs/2109.01962 Counterfactual evaluation for explainable ai . Preprint, arXiv:2109.01962

    2021 arXiv

  10. [18]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

    2023 arXiv

  11. [19]

    Divyansh Kaushik, Eduard Hovy, and Zachary Lipton. 2020. https://openreview.net/forum?id=Sklgs0NFvr Learning the difference that makes a difference with counterfactually-augmented data . In International Conference on Learning Representations

    2020

  12. [20]

    Dawei Li, Renliang Sun, Yue Huang, Ming Zhong, Bohan Jiang, Jiawei Han, Xiangliang Zhang, Wei Wang, and Huan Liu. 2025. https://arxiv.org/abs/2502.01534 Preference leakage: A contamination problem in llm-as-a-judge . Preprint, arXiv:2502.01534

    2025

  13. [21]

    Yongqi Li, Mayi Xu, Xin Miao, Shen Zhou, and Tieyun Qian. 2024. https://aclanthology.org/2024.lrec-main.1156/ Prompting large language models for counterfactual generation: An empirical study . In Proceedings of the 2024 Joint International Conference on Computational Linguist...

    2024

  14. [22]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2020. https://openreview.net/forum?id=SyxS0T4tvS Ro \ bert \ a: A robustly optimized \ bert \ pretraining approach

    2020

  15. [23]

    Nishtha Madaan, Inkit Padhi, Naveen Panwar, and Diptikalyan Saha. 2021. https://doi.org/10.1609/aaai.v35i15.17594 Generate your counterfactuals: Towards controlled counterfactual generation for text . Proceedings of the AAAI Conference on Artificial Intelligence, 35(15):13516--13524

    2021 doi

  16. [24]

    Andreas Madsen, Siva Reddy, and Sarath Chandar. 2022. https://doi.org/10.1145/3546577 Post-hoc interpretability for neural nlp: A survey . ACM Comput. Surv., 55(8)

    2022 doi

  17. [25]

    Tim Miller. 2019. Explanation in artificial intelligence: Insights from the social sciences. Artificial intelligence, 267:1--38

    2019

  18. [26]

    o rg Schl \

    Van Bach Nguyen, Christin Seifert, and J \"o rg Schl \"o tterer. 2024 a . https://aclanthology.org/2024.inlg-main.6/ CE val: A benchmark for evaluating counterfactual text generation . In Proceedings of the 17th International Natural Language Generation Conference, pages 55--6...

    2024

  19. [27]

    o rg Schl \

    Van Bach Nguyen, Paul Youssef, Christin Seifert, and J \"o rg Schl \"o tterer. 2024 b . https://doi.org/10.18653/v1/2024.findings-emnlp.870 LLM s for generating and evaluating counterfactuals: A comprehensive study . In Findings of the Association for Computational Linguistics...

    2024 doi

  20. [28]

    Qwen. 2024. https://arxiv.org/abs/2412.15115 Qwen2.5 technical report . Preprint, arXiv:2412.15115

    2024 arXiv

  21. [29]

    Marcel Robeer, Floris Bex, and Ad Feelders. 2021. https://doi.org/10.18653/v1/2021.findings-emnlp.306 Generating realistic natural language counterfactuals . In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3611--3625, Punta Cana, Dominican Repub...

    2021 doi

  22. [30]

    Alexis Ross, Ana Marasovi \'c , and Matthew Peters. 2021. https://doi.org/10.18653/v1/2021.findings-acl.336 Explaining NLP models via minimal contrastive editing ( M i CE ) . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3840--3852, Onlin...

    2021 doi

  23. [31]

    Manning, Andrew Ng, and Christopher Potts

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. https://www.aclweb.org/anthology/D13-1170 Recursive deep models for semantic compositionality over a sentiment treebank . In Proceedings of the 2013 Conferenc...

    2013

  24. [32]

    Hwanjun Song, Minseok Kim, Dongmin Park, Yooju Shin, and Jae-Gil Lee. 2023. https://doi.org/10.1109/tnnls.2022.3152527 Learning from noisy labels with deep neural networks: A survey . IEEE transactions on neural networks and learning systems, 34(11):8135—8153

    2023

  25. [33]

    Sowmya Vajjala and Shwetali Shimangaud. 2025. https://arxiv.org/abs/2502.11830 Text classification in the llm era - where do we stand? Preprint, arXiv:2502.11830

    2025 arXiv

  26. [34]

    Qianli Wang, Tatiana Anikina, Nils Feldhus, Josef Genabith, Leonhard Hennig, and Sebastian M \"o ller. 2024. https://doi.org/10.18653/v1/2024.hcinlp-1.9 LLMC heckup: Conversational examination of large language models via interpretability tools and self-explanations . In Proce...

    2024 doi

  27. [35]

    Qianli Wang, Nils Feldhus, Simon Ostermann, Luis Felipe Villa-Arenas, Sebastian Möller, and Vera Schmitt. 2025. https://arxiv.org/abs/2501.00777 Fitcf: A framework for automatic feature importance-guided counterfactual example generation . Preprint, arXiv:2501.00777

    2025 arXiv

  28. [36]

    Sarah Wiegreffe, Ana Marasovi \'c , and Noah A. Smith. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.804 M easuring association between labels and free-text rationales . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10266-...

    2021 doi

  29. [37]

    Tongshuang Wu, Marco Tulio Ribeiro, Jeffrey Heer, and Daniel Weld. 2021. https://doi.org/10.18653/v1/2021.acl-long.523 Polyjuice: Generating counterfactuals for explaining, evaluating, and improving models . In Proceedings of the 59th Annual Meeting of the Association for Comp...

    2021 doi

  30. [38]

    Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. https://proceedings.neurips.cc/paper_files/paper/2015/file/250cf8b51c773f3f8dc8b4be867a9a02-Paper.pdf Character-level convolutional networks for text classification . In Advances in Neural Information Processing Systems, volume 28...

    2015

  31. [39]

    Hedderich, Fangzhou Zhai, David Ifeoluwa Adelani, and Dietrich Klakow

    Dawei Zhu, Michael A. Hedderich, Fangzhou Zhai, David Ifeoluwa Adelani, and Dietrich Klakow. 2022. https://doi.org/10.18653/v1/2022.insights-1.8 Is BERT robust to label noise? a study on learning with noisy labels in text classification . In Proceedings of the Third Workshop o...

    2022 doi

  32. [40]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  33. [41]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.