pith. sign in

arxiv: 2606.06715 · v1 · pith:IWELVL6Qnew · submitted 2026-06-04 · 💻 cs.CL · cs.AI· cs.LG

Does Topic Sentiment Cause Perceived Ideology? Comparing Human and LLM Annotations in Political News Articles

Upasana Chatterjee This is my paper

Pith reviewed 2026-06-28 01:15 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.LG
keywords topic sentimentperceived ideologyLLM annotationsshortcut learningmediation analysisdouble machine learningpolitical newscausal effects
0
0 comments X

The pith

Fine-tuned GPT-4o-mini alone shows a causal effect of topic sentiment on perceived ideology that human annotators do not.

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

The paper tests whether topic sentiment exerts a causal influence on how readers perceive the political ideology of news articles, and whether the answer changes with the source of the ideology labels. It pairs AllSides articles with sentiment scores from one shared LLM and then runs the same Double Machine Learning and mediation analyses on ideology labels supplied by human experts, a baseline LLM, a fine-tuned LLM, and a second LLM. Human labels produce no detectable community-level treatment effects or natural direct effects. Only the fine-tuned GPT-4o-mini produces statistically significant effects on both measures, which the authors read as the model having internalized a spurious sentiment-ideology link during fine-tuning.

Core claim

Fine-tuned GPT-4o-mini is the only annotator paradigm that produces significant community-level treatment effects and significant natural direct effects (NDEs) in mediation. We interpret this as evidence of shortcut learning: fine-tuning on ideology-labeled data causes the model to internalise a spurious sentiment-ideology coupling not operative in human judgment for this task.

What carries the argument

Double Machine Learning combined with community-level mediation analysis, applied separately to each of the four ideology annotation sources while holding sentiment labels fixed.

If this is right

  • Human ideology annotations exhibit no causal response to topic sentiment.
  • Standard classification accuracy metrics such as F1 miss the spurious coupling that appears in causal estimators.
  • Fine-tuned LLM labels can function as silver labels for classification yet still distort downstream causal conclusions.
  • The sentiment-ideology coupling is invisible to accuracy-based validation but surfaces under mediation analysis.

Where Pith is reading between the lines

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

  • Researchers who substitute fine-tuned LLM labels for human labels in causal studies risk importing training artifacts that only become visible under mediation or DML.
  • One practical check would be to compare causal estimates obtained from fine-tuned versus zero-shot LLM labels on the same task before treating the labels as interchangeable with human data.
  • The same shortcut mechanism could appear in other paired annotation tasks where one variable is used both for fine-tuning and for later causal modeling.

Load-bearing premise

The shared sentiment annotations are independent of how ideology labels were produced and the estimators recover true causal effects without unmeasured confounding between sentiment and ideology.

What would settle it

Re-running the identical DML and mediation pipeline on the same articles but with sentiment scores supplied by human raters instead of the LLM, and finding no significant NDE for the fine-tuned model, would falsify the shortcut-learning claim.

Figures

Figures reproduced from arXiv: 2606.06715 by Upasana Chatterjee.

Figure 1
Figure 1. Figure 1: Complete causal diagram for article ideology [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Expanded setup: sentiment variables as media [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Community-level ATE. • Community 3 (Social Issues: topics include immigration, LGBTQ, abortion, religion): Positive sentiment shifts fine-tuned GPT ideol￾ogy predictions left (ATE = −0.023; ≈2.3% of one category step toward Left) • Community 9 (Geopolitics and Foreign Af￾fairs: topics include foreign policy, climate, global affairs): Positive sentiment shifts fine￾tuned GPT ideology predictions left (ATE =… view at source ↗
Figure 4
Figure 4. Figure 4: Community 9 → 11 mediation results. Error bars show 95% bootstrap CIs; filled markers indicate significance [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ideology prediction distribution for Commu [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overall ideology prediction distribution across [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ideology prediction distribution for Commu [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 10
Figure 10. Figure 10: Ideology prediction distribution for Commu [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Donald Trump → Politics Mediation Effects Given the small article counts, it is not possible to determine whether the absence of significant effects reflects a genuine lack of causal coupling or insufficient statistical power. A larger topic￾level corpus would be required to distinguish these explanations. D Community-level Mediation Results Six treatment-mediator pairs were evaluated across communities {… view at source ↗
Figure 11
Figure 11. Figure 11: ATE for top five topics. The topic-level dataset is substantially smaller than the community-level dataset, as individual top￾ics appear in far fewer articles than aggregated com￾munities. This limits statistical power and makes it difficult to draw reliable conclusions from topic￾level estimates alone. With a larger corpus covering more diverse topics and time periods, topic-level analysis would become m… view at source ↗
read the original abstract

We ask whether topic sentiment has a causal effect on perceived political ideology, and whether the answer depends on who assigns the ideology label. Using articles from AllSides, paired with shared sentiment annotations from Llama-3.3-70b-versatile, we compare ideology labels from expert human annotators, GPT-4o-mini (baseline and finetuned), and Llama-3.3-70B. We apply Double Machine Learning (DML) and community-level mediation analysis across all four annotation paradigms. Human annotations yield no significant causal effects at the community level. Fine-tuned GPT-4o-mini achieves the highest classification accuracy (F1=72.48) and is the only annotator paradigm that produces significant community-level treatment effects and significant natural direct effects (NDEs) in mediation. We interpret this as evidence of shortcut learning: fine-tuning on ideology-labeled data causes the model to internalise a spurious sentiment--ideology coupling not operative in human judgment for this task. This coupling is structurally invisible to F1-based evaluation, with implications for the use of LLM annotations as silver labels and as proxies for human judgment in downstream causal analyses.

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

2 major / 2 minor

Summary. The manuscript examines whether topic sentiment has a causal effect on perceived political ideology in news articles and whether this effect varies by annotator type (human experts, baseline and fine-tuned GPT-4o-mini, Llama-3.3-70B). Using shared Llama-3.3-70b sentiment annotations on AllSides articles, the authors apply Double Machine Learning (DML) and community-level mediation analysis. They report no significant causal effects for human annotations, but significant treatment effects and natural direct effects (NDEs) only for the fine-tuned GPT-4o-mini model (F1=72.48), interpreting this as evidence of shortcut learning induced by fine-tuning.

Significance. If the causal identification holds, the finding that fine-tuning introduces a spurious sentiment-ideology link not present in human annotations is significant for the field of LLM evaluation and causal inference using machine annotations. It suggests that accuracy metrics like F1 may miss important biases in downstream causal analyses, with implications for using LLMs as silver labels. The use of shared sentiment annotations across paradigms is a strength that allows direct comparison.

major comments (2)
  1. [DML implementation (methods section)] DML implementation (methods section): The paper provides no details on the covariate set used in the nuisance models for DML (e.g., whether article embeddings, topic indicators, or other text features beyond the Llama sentiment treatment are included). This is load-bearing for the central claim because both treatment (sentiment) and outcome (ideology label) derive from the same article text; omitted variables can induce confounding whose magnitude may differ across annotator paradigms, particularly for the fine-tuned model.
  2. [Mediation analysis results] Mediation analysis results: The claim that only the fine-tuned GPT-4o-mini yields significant community-level treatment effects and NDEs (supporting the shortcut-learning interpretation) rests on the no-unmeasured-confounding assumption without reported sensitivity analyses (e.g., e-values or alternative covariate specifications). The shared Llama sentiment is treated as exogenous, but the manuscript does not test or discuss whether this holds under the observed covariates alone.
minor comments (2)
  1. [Abstract] Abstract: The F1=72.48 for fine-tuned GPT-4o-mini is reported without the corresponding F1 scores for the other three annotation paradigms, making it impossible to assess the relative accuracy advantage.
  2. [Results] The manuscript would benefit from an explicit table reporting sample sizes, annotation agreement rates, and the exact number of articles per community for each paradigm.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The two major comments identify important gaps in the reporting of our identification strategy. We address each below and commit to revisions that strengthen the manuscript without altering its core claims or conclusions.

read point-by-point responses

  1. Referee: [DML implementation (methods section)] DML implementation (methods section): The paper provides no details on the covariate set used in the nuisance models for DML (e.g., whether article embeddings, topic indicators, or other text features beyond the Llama sentiment treatment are included). This is load-bearing for the central claim because both treatment (sentiment) and outcome (ideology label) derive from the same article text; omitted variables can induce confounding whose magnitude may differ across annotator paradigms, particularly for the fine-tuned model.

    Authors: We agree that explicit documentation of the nuisance-model covariates is required for readers to evaluate the plausibility of the conditional ignorability assumption. The current manuscript describes the DML procedure at a high level but does not enumerate the feature set passed to the outcome and treatment nuisance functions. In the revision we will add a dedicated subsection that lists all covariates (article-level embeddings from the shared Llama encoder, topic indicators, length, and publication metadata) together with the cross-fitting procedure and hyper-parameter choices. This addition will allow direct assessment of whether the same covariate specification is used across all four annotation paradigms. revision: yes

  2. Referee: [Mediation analysis results] Mediation analysis results: The claim that only the fine-tuned GPT-4o-mini yields significant community-level treatment effects and NDEs (supporting the shortcut-learning interpretation) rests on the no-unmeasured-confounding assumption without reported sensitivity analyses (e.g., e-values or alternative covariate specifications). The shared Llama sentiment is treated as exogenous, but the manuscript does not test or discuss whether this holds under the observed covariates alone.

    Authors: The referee correctly notes the absence of formal sensitivity checks for the no-unmeasured-confounding assumption that underpins both the DML and mediation results. While the shared Llama sentiment annotations are held fixed across paradigms, we did not report e-values, placebo tests, or alternative covariate sets. In the revised version we will add (i) e-value calculations for the significant NDE estimates obtained with the fine-tuned model and (ii) a brief discussion of how the results change when the covariate set is restricted to the observed text features alone. These additions will quantify the robustness of the reported community-level effects without requiring new data collection. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical causal estimates rely on external AllSides data and standard DML/mediation methods without self-referential reduction

full rationale

The paper applies Double Machine Learning and community-level mediation analysis to compare ideology labels across human and LLM annotators, using shared Llama-3.3-70b sentiment annotations as the treatment variable. No equations or derivations within the paper reduce the reported treatment effects or NDEs to quantities defined by parameters fitted inside the study itself. The analysis depends on external article data from AllSides and off-the-shelf models; the interpretation of shortcut learning in the fine-tuned GPT-4o-mini is an empirical claim resting on the validity of the no-unmeasured-confounding assumption rather than any definitional or self-citational loop. No self-citations are load-bearing for the central results, and no ansatz, renaming, or fitted-input-as-prediction patterns appear.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard causal inference assumptions rather than new free parameters or invented entities. No new particles, forces, or dimensions are postulated.

axioms (2)
  • domain assumption Double Machine Learning assumptions (unconfoundedness, overlap, correct nuisance models) hold for the observed covariates
    Invoked implicitly when applying DML to recover treatment effects of sentiment on ideology labels
  • domain assumption Sentiment annotations from Llama-3.3-70b are independent of the ideology annotation process
    Required for treating sentiment as the treatment variable across all four labeling paradigms

pith-pipeline@v0.9.1-grok · 5736 in / 1517 out tokens · 25371 ms · 2026-06-28T01:15:15.853150+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

44 extracted references · 16 canonical work pages

  1. [1]

    Proceedings of the 9th workshop on computational approaches to subjectivity, sentiment and social media analysis , pages=

    Topic-specific sentiment analysis can help identify political ideology , author=. Proceedings of the 9th workshop on computational approaches to subjectivity, sentiment and social media analysis , pages=

  2. [2]

    , author=

    A Preliminary Investigation into Sentiment Analysis of Informal Political Discourse. , author=. AAAI spring symposium: computational approaches to analyzing weblogs , pages=

  3. [3]

    Political Analysis , volume=

    Sentiment is not stance: Target-aware opinion classification for political text analysis , author=. Political Analysis , volume=. 2023 , publisher=

    2023

  4. [4]

    Discourse & Communication , volume=

    Ideology through sentiment analysis: A changing perspective on Russia and Islam in NYT , author=. Discourse & Communication , volume=. 2017 , publisher=

    2017

  5. [5]

    , title =

    Rubin, Donald B. , title =. Journal of Educational Psychology , year =

  6. [6]

    The Econometrics Journal , year =

    Chernozhukov, Victor and Chetverikov, Denis and Demirer, Mert and Duflo, Esther and Hansen, Christian and Newey, Whitney and Robins, James , title =. The Econometrics Journal , year =

  7. [7]

    Journal of the American Statistical Association , year =

    Wager, Stefan and Athey, Susan , title =. Journal of the American Statistical Association , year =

  8. [8]

    Sensitivity Analysis for Causal Mediation through Text: an Application to Political Polarization

    Tierney, Graham and Volfovsky, Alexander. Sensitivity Analysis for Causal Mediation through Text: an Application to Political Polarization. Proceedings of the First Workshop on Causal Inference and NLP. 2021. doi:10.18653/v1/2021.cinlp-1.5

    work page doi:10.18653/v1/2021.cinlp-1.5 2021

  9. [9]

    and Taddy, Matt , title =

    Gentzkow, Matthew and Shapiro, Jesse M. and Taddy, Matt , title =. Econometrica , volume =. doi:https://doi.org/10.3982/ECTA16566 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.3982/ECTA16566 , abstract =

    work page doi:10.3982/ecta16566

  10. [10]

    An E mbedding M odel for E stimating L egislative P references from the F requency and S entiment of T weets

    Spell, Gregory and Guay, Brian and Hillygus, Sunshine and Carin, Lawrence. An E mbedding M odel for E stimating L egislative P references from the F requency and S entiment of T weets. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020. doi:10.18653/v1/2020.emnlp-main.46

    work page doi:10.18653/v1/2020.emnlp-main.46 2020

  11. [11]

    Text as Causal Mediators: Research Design for Causal Estimates of Differential Treatment of Social Groups via Language Aspects

    Keith, Katherine and Rice, Douglas and O ' Connor, Brendan. Text as Causal Mediators: Research Design for Causal Estimates of Differential Treatment of Social Groups via Language Aspects. Proceedings of the First Workshop on Causal Inference and NLP. 2021. doi:10.18653/v1/2021.cinlp-1.2

    work page doi:10.18653/v1/2021.cinlp-1.2 2021

  12. [12]

    Conference on uncertainty in artificial intelligence , pages=

    Adapting text embeddings for causal inference , author=. Conference on uncertainty in artificial intelligence , pages=. 2020 , organization=

    2020

  13. [13]

    Investigating Gender Bias in Language Models Using Causal Mediation Analysis , url =

    Vig, Jesse and Gehrmann, Sebastian and Belinkov, Yonatan and Qian, Sharon and Nevo, Daniel and Singer, Yaron and Shieber, Stuart , booktitle =. Investigating Gender Bias in Language Models Using Causal Mediation Analysis , url =

  14. [14]

    Making reasoning matter: Measuring and improving faithfulness of chain-of-thought reasoning

    Paul, Debjit and West, Robert and Bosselut, Antoine and Faltings, Boi. Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.882

    work page doi:10.18653/v1/2024.findings-emnlp.882 2024

  15. [15]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =

    Stolfo, Alessandro and Belinkov, Yonatan and Sachan, Mrinmaya. A Mechanistic Interpretation of Arithmetic Reasoning in Language Models using Causal Mediation Analysis. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.435

    work page doi:10.18653/v1/2023.emnlp-main.435 2023

  16. [16]

    Beyond Surface Structure: A Causal Assessment of

    Yujin Han and Lei Xu and Sirui Chen and Difan Zou and Chaochao Lu , booktitle=. Beyond Surface Structure: A Causal Assessment of. 2025 , url=

    2025

  17. [17]

    Keith and Emaad Manzoor and Reid Pryzant and Dhanya Sridhar and Zach Wood-Doughty and Jacob Eisenstein and Justin Grimmer and Roi Reichart and Margaret E

    Amir Feder and Katherine A. Keith and Emaad Manzoor and Reid Pryzant and Dhanya Sridhar and Zach Wood-Doughty and Jacob Eisenstein and Justin Grimmer and Roi Reichart and Margaret E. Roberts and Brandon M. Stewart and Victor Veitch and Diyi Yang , title =. 2022 , journal =

    2022

  18. [18]

    Probabilistic and Causal Inference , year=

    Direct and Indirect Effects , author=. Probabilistic and Causal Inference , year=

  19. [19]

    Neural Media Bias Detection Using Distant Supervision With BABE - Bias Annotations By Experts

    Spinde, Timo and Plank, Manuel and Krieger, Jan-David and Ruas, Terry and Gipp, Bela and Aizawa, Akiko. Neural Media Bias Detection Using Distant Supervision With BABE - Bias Annotations By Experts. Findings of the Association for Computational Linguistics: EMNLP 2021. 2021. doi:10.18653/v1/2021.findings-emnlp.101

    work page doi:10.18653/v1/2021.findings-emnlp.101 2021

  20. [20]

    MBIC--A Media Bias Annotation Dataset Including Annotator Characteristics , author=

  21. [21]

    We Can Detect Your Bias: Predicting the Political Ideology of News Articles

    Baly, Ramy and Da San Martino, Giovanni and Glass, James and Nakov, Preslav. We Can Detect Your Bias: Predicting the Political Ideology of News Articles. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020. doi:10.18653/v1/2020.emnlp-main.404

    work page doi:10.18653/v1/2020.emnlp-main.404 2020

  22. [22]

    Moral Framing and Ideological Bias of News , ISBN=

    Mokhberian, Negar and Abeliuk, Andrés and Cummings, Patrick and Lerman, Kristina , year=. Moral Framing and Ideological Bias of News , ISBN=. doi:10.1007/978-3-030-60975-7_16 , booktitle=

    work page doi:10.1007/978-3-030-60975-7_16

  23. [23]

    Entman , title =

    Robert M. Entman , title =. Journalism , volume =. 2010 , doi =

    2010

  24. [24]

    Wagner and Mike Gruszczynski , title =

    Michael W. Wagner and Mike Gruszczynski , title =. Journalism & Communication Monographs , volume =. 2016 , doi =

    2016

  25. [25]

    IEEE Transactions on Computational Social Systems , volume=

    NewsSlant: Analyzing political news and its influence through a moral lens , author=. IEEE Transactions on Computational Social Systems , volume=. 2024 , publisher=

    2024

  26. [26]

    I ndi V ec: An Exploration of Leveraging Large Language Models for Media Bias Detection with Fine-Grained Bias Indicators

    Lin, Luyang and Wang, Lingzhi and Zhao, Xiaoyan and Li, Jing and Wong, Kam-Fai. I ndi V ec: An Exploration of Leveraging Large Language Models for Media Bias Detection with Fine-Grained Bias Indicators. Findings of the Association for Computational Linguistics: EACL 2024. 2024. doi:10.18653/v1/2024.findings-eacl.70

    work page doi:10.18653/v1/2024.findings-eacl.70 2024

  27. [27]

    Algorithmic Fidelity of Large Language Models in Generating Synthetic G erman Public Opinions: A Case Study

    Ma, Bolei and Yoztyurk, Berk and Haensch, Anna-Carolina and Wang, Xinpeng and Herklotz, Markus and Kreuter, Frauke and Plank, Barbara and A enmacher, Matthias. Algorithmic Fidelity of Large Language Models in Generating Synthetic G erman Public Opinions: A Case Study. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics ...

    work page doi:10.18653/v1/2025.acl-long.90 2025

  28. [28]

    Large language models can be used to estimate the latent positions of politicians , author=

  29. [29]

    Nature human behaviour , volume=

    Testing theory of mind in large language models and humans , author=. Nature human behaviour , volume=. 2024 , publisher=

    2024

  30. [30]

    Marketing Science , volume=

    Frontiers: Determining the validity of large language models for automated perceptual analysis , author=. Marketing Science , volume=. 2024 , publisher=

    2024

  31. [31]

    2023 , institution=

    Large language models as simulated economic agents: What can we learn from homo silicus? , author=. 2023 , institution=

    2023

  32. [32]

    Proceedings of the National Academy of Sciences122, 24 (June 2025), e2501660122

    Yuan Gao and Dokyun Lee and Gordon Burtch and Sina Fazelpour , title =. Proceedings of the National Academy of Sciences , volume =. 2025 , doi =. https://www.pnas.org/doi/pdf/10.1073/pnas.2501660122 , abstract =

    work page doi:10.1073/pnas.2501660122 2025

  33. [33]

    Reading Psychology , volume =

    Marcy Zipke , title =. Reading Psychology , volume =. 2007 , publisher =. doi:10.1080/02702710701260615 , URL =. https://doi.org/10.1080/02702710701260615 , abstract =

    work page doi:10.1080/02702710701260615 2007

  34. [34]

    Vocabulary acquisition: Implications for reading comprehension , pages=

    Metalinguistic awareness and the vocabulary-comprehension connection , author=. Vocabulary acquisition: Implications for reading comprehension , pages=. 2007 , publisher=

    2007

  35. [35]

    Metalinguistic awareness in children: Theory, research, and implications , pages=

    The development of metalinguistic awareness: A conceptual overview , author=. Metalinguistic awareness in children: Theory, research, and implications , pages=. 1984 , publisher=

    1984

  36. [36]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Do llms overcome shortcut learning? an evaluation of shortcut challenges in large language models , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  37. [37]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Explore spurious correlations at the concept level in language models for text classification , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  38. [38]

    Communications of the ACM , volume=

    Shortcut learning of large language models in natural language understanding , author=. Communications of the ACM , volume=. 2023 , publisher=

    2023

  39. [39]

    npj Digital Medicine , author=

    Shortcut learning in medical AI hinders generalization: Method for estimating AI model generalization without external data , volume=. npj Digital Medicine , author=. 2024 , month=. doi:10.1038/s41746-024-01118-4 , number=

    work page doi:10.1038/s41746-024-01118-4 2024

  40. [40]

    LLM s Are Prone to Fallacies in Causal Inference

    Joshi, Nitish and Saparov, Abulhair and Wang, Yixin and He, He. LLM s Are Prone to Fallacies in Causal Inference. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653/v1/2024.emnlp-main.590

    work page doi:10.18653/v1/2024.emnlp-main.590 2024

  41. [41]

    2025 , eprint=

    Correlation or Causation: Analyzing the Causal Structures of LLM and LRM Reasoning Process , author=. 2025 , eprint=

    2025

  42. [42]

    Proceedings of the International Conference on Learning Representations (ICLR) , year =

    Can Large Language Models Infer Causation from Correlation? , author =. Proceedings of the International Conference on Learning Representations (ICLR) , year =

  43. [43]

    SIGKDD Explor

    Moraffah, Raha and Karami, Mansooreh and Guo, Ruocheng and Raglin, Adrienne and Liu, Huan , title =. SIGKDD Explor. Newsl. , month = may, pages =. 2020 , issue_date =. doi:10.1145/3400051.3400058 , abstract =

    work page doi:10.1145/3400051.3400058 2020

  44. [44]

    Selçuk and Liu, Huan , journal=

    Cheng, Lu and Guo, Ruocheng and Moraffah, Raha and Sheth, Paras and Candan, K. Selçuk and Liu, Huan , journal=. Evaluation Methods and Measures for Causal Learning Algorithms , year=