Paper deep dive

Thinking Fast, Thinking Wrong: Intuitiveness Modulates LLM Counterfactual Reasoning in Policy Evaluation

Yanjie He

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 27

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%

Last extracted: 4/14/2026, 2:18:55 AM

Summary

This paper introduces a benchmark of 40 empirical policy evaluation cases to assess the counterfactual reasoning capabilities of Large Language Models (LLMs). The study identifies a 'Chain-of-Thought (CoT) paradox' where CoT prompting significantly improves performance on intuitive cases but fails on counter-intuitive ones, suggesting that LLMs perform 'slow talking' rather than genuine 'slow thinking' (System 2 reasoning). The findings indicate that intuitiveness is the dominant factor in model performance, and that models often fail to override intuitive priors even when they possess the relevant causal knowledge.

Entities (7)

Chain-of-thought · prompting-strategy · 100%Claude Opus 4.6 · llm · 100%Claude Sonnet 4 · llm · 100%GPT-4.1 · llm · 100%GPT-5.2 · llm · 100%Yanjie He · researcher · 100%Dual-process theory · cognitive-theory · 95%

Relation Signals (3)

Chain-of-thought → failson → Counter-intuitive cases

confidence 98% · this benefit is nearly eliminated on counter-intuitive ones

Chain-of-thought → improvesperformanceon → Obvious cases

confidence 98% · chain-of-thought prompting dramatically improves performance on obvious cases

GPT-5.2 → outperforms → GPT-4.1

confidence 95% · GPT-5.2 substantially outperforms all other models, particularly on two-sided counter-intuitive cases

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Large language models (LLMs) are increasingly used for causal and counterfactual reasoning, yet their reliability in real-world policy evaluation remains underexplored. We construct a benchmark of 40 empirical policy evaluation cases drawn from economics and social science, each grounded in peer-reviewed evidence and classified by intuitiveness -- whether the empirical finding aligns with (obvious), is unclear relative to (ambiguous), or contradicts (counter-intuitive) common prior expectations. We evaluate four frontier LLMs across five prompting strategies with 2,400 experimental trials and analyze the results using mixed-effects logistic regression. Our findings reveal three key results: (1) a chain-of-thought (CoT) paradox, where chain-of-thought prompting dramatically improves performance on obvious cases but this benefit is nearly eliminated on counter-intuitive ones (interaction OR = 0.053, $p < 0.001$); (2) intuitiveness as the dominant factor, explaining more variance than model choice or prompting strategy (ICC = 0.537); and (3) a knowledge-reasoning dissociation, where citation-based familiarity is unrelated to accuracy ($p = 0.53$), suggesting models possess relevant knowledge but fail to reason with it when findings contradict intuition. We frame these results through the lens of dual-process theory (System 1 vs. System 2) and argue that current LLMs' "slow thinking" may be little more than "slow talking" -- they produce the form of deliberative reasoning without the substance.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

27,056 characters extracted from source content.

Expand or collapse full text

Thinking Fast, Thinking Wrong: Intuitiveness Modulates LLM Counterfactual Reasoning in Policy Evaluation Yanjie He Independent Researcher heyanjie0@outlook.com Abstract Large language models (LLMs) are increasingly used for causal and counterfactual reasoning, yet their reliability in real-world policy evaluation remains underexplored. We construct a benchmark of 40 empirical policy evaluation cases drawn from economics and social science, each grounded in peer-reviewed evidence and classified by intuitiveness—whether the empirical finding aligns with (obvious), is unclear relative to (ambiguous), or contradicts (counter-intuitive) common prior expectations. We evaluate four frontier LLMs across five prompting strategies with 2,400 experimental trials and analyze the results using mixed-effects logistic regression. Our findings reveal three key results: (1) a chain-of-thought (CoT) paradox, where chain-of-thought prompting dramatically improves performance on obvious cases but this benefit is nearly eliminated on counter-intuitive ones (interaction OR = 0.053, p<0.001p<0.001); (2) intuitiveness as the dominant factor, explaining more variance than model choice or prompting strategy (ICC = 0.537); and (3) a knowledge–reasoning dissociation, where citation-based familiarity is unrelated to accuracy (p=0.53p=0.53), suggesting models possess relevant knowledge but fail to reason with it when findings contradict intuition. We frame these results through the lens of dual-process theory (System 1 vs. System 2) and argue that current LLMs’ “slow thinking” may be little more than “slow talking”—they produce the form of deliberative reasoning without the substance. Thinking Fast, Thinking Wrong: Intuitiveness Modulates LLM Counterfactual Reasoning in Policy Evaluation Yanjie He Independent Researcher heyanjie0@outlook.com 1 Introduction Large language models (LLMs) have demonstrated impressive capabilities across a wide range of reasoning tasks (Brown et al., 2020; OpenAI, 2023). Recent work suggests they can perform causal reasoning with high accuracy, surpassing existing methods on standard benchmarks (Kıcıman et al., 2023). These findings have fueled optimism about using LLMs as assistants for causal inference and policy analysis. However, most existing evaluations rely on synthetic causal graphs (Jin et al., 2024), commonsense scenarios, or abstract logical problems (Huang et al., 2024). The real-world domain of policy evaluation—where economists and social scientists use quasi-experimental methods to estimate causal effects—presents a fundamentally different challenge. Policy effects are messy, contested, and, critically, sometimes counter-intuitive: the data show the opposite of what most people would expect. We introduce a benchmark of 40 empirical policy evaluation cases drawn from peer-reviewed economics and social science research. Each case is classified along an intuitiveness dimension: • Obvious: The empirical result aligns with common prior expectations (e.g., Medicaid expansion increases healthcare utilization). • Ambiguous: Reasonable arguments exist for multiple directions (e.g., class size reduction may or may not improve test scores). • Counter-intuitive: The empirical result contradicts naïve expectations (e.g., introducing fines for late daycare pickup increases lateness). This classification is inspired by dual-process theory from cognitive psychology (Kahneman, 2011; Stanovich and West, 2000; Evans, 2003): obvious cases can be solved by fast, intuitive “System 1” processing, while counter-intuitive cases require slow, deliberative “System 2” reasoning that overrides the initial heuristic response. We evaluate four frontier LLMs—GPT-5.2, Claude Sonnet 4, Claude Opus 4.6, and GPT-4.1—across five prompting strategies (naïve, expert persona, chain-of-thought, structured reasoning, and adversarial framing) with 3 repetitions per condition, yielding 2,400 experimental trials. We analyze the results using mixed-effects logistic regression to properly account for the nested structure of the data. Our key findings are: (1) a CoT paradox—chain-of-thought prompting dramatically improves accuracy on obvious cases (main-effect OR ≈ 26) but this benefit is nearly eliminated on counter-intuitive ones (interaction OR = 0.053, p<0.001p<0.001); (2) intuitiveness as the dominant factor, with case-level random intercepts explaining 53.7% of variance (ICC = 0.537); and (3) a knowledge–reasoning dissociation, where citation-based familiarity is unrelated to accuracy (p=0.53p=0.53), suggesting models possess relevant knowledge but fail to reason with it when findings contradict intuition. Overall accuracy of 81.3% masks critical failures precisely where counterfactual reasoning matters most. 2 Related Work Causal Reasoning in LLMs. Kıcıman et al. (2023) report 92% accuracy on counterfactual reasoning but rely on commonsense and synthetic scenarios. Jin et al. (2024) propose CLadder using synthetic causal graphs with oracle-derived ground truth. Zečević et al. (2023) argue LLMs are “causal parrots” reciting memorized knowledge. Our work bridges these perspectives: LLMs can leverage causal knowledge, but fail when doing so requires overriding intuitive priors—a distinction invisible in synthetic benchmarks. Counterfactual Reasoning Benchmarks. Huang et al. (2024) propose CLOMO for generative counterfactual modification; Mu and Li (2024) use VAE-based counterfactual reasoning in narratives; Weinzierl and Harabagiu (2024) develop Tree-of-Counterfactual prompting. Unlike these works, we focus on evaluative counterfactual reasoning in real-world policy evaluation, where ground truth is established by peer-reviewed empirical research. Cognitive Parallels in LLM Behavior. Ying et al. (2024) classify LLMs’ response styles as “intuitive” or “dependent”; their “intuitive” refers to behavioral style, whereas ours refers to cognitive accessibility of task content. Mondorf and Plank (2024) find that LLM accuracy does not necessarily reflect reasoning validity—a conclusion our findings reinforce. Shen et al. (2025) show that LLMs follow the Yerkes-Dodson law under stress-inducing prompts. Chain-of-Thought Reasoning. CoT prompting (Wei et al., 2022; Kojima et al., 2022) is widely assumed to improve reasoning, but Jacovi et al. (2024) show that reasoning chains are only as strong as their weakest step. Our CoT paradox extends this: the weakest step may be the very first one—the intuitive prior that anchors subsequent reasoning. Snell et al. (2025) show that test-time compute scaling depends critically on prompt difficulty, consistent with our finding. 3 Benchmark Design 3.1 Case Selection We curate 40 empirical policy evaluation cases from economics/social science, each satisfying three criteria: (1) at least one peer-reviewed publication establishing the causal effect via a credible identification strategy (difference-in-differences, regression discontinuity, instrumental variables, or randomized controlled trial); (2) an unambiguous directional finding (increase, decrease, or no significant change); and (3) sufficient prominence that the study is likely present in LLM training data. Cases span 8 countries, 12 policy domains, and publication years from 1974 to 2022. Table 1 summarizes the distribution. Intuitiveness One-sided Two-sided Obvious 0 13 Ambiguous 1 6 Counter-intuitive 12 8 Total 13 27 Table 1: Distribution of 40 benchmark cases by intuitiveness and test type. One-sided cases ask a YES/NO question; two-sided cases ask for directional prediction (INCREASE / DECREASE / NO SIGNIFICANT CHANGE). 3.2 Intuitiveness Classification Each case is classified as obvious, ambiguous, or counter-intuitive based on whether the established empirical finding aligns with, is unclear relative to, or contradicts the prior that an informed layperson would hold. This classification was performed by the author and validated against the pattern of LLM responses: obvious cases have high naïve-prompt accuracy across all models, while counter-intuitive cases have low naïve accuracy (see §5). Examples of counter-intuitive cases include: • Introducing fines for late daycare pickup increases lateness (Gneezy and Rustichini, 2000): monetary penalties replace moral obligation with a market transaction. • “Right-to-carry” concealed handgun laws increase violent crime (Donohue et al., 2019): contrary to deterrence theory. • Violent movie releases decrease violent crime (Dahl and DellaVigna, 2009): voluntary incapacitation outweighs any media-effects priming. 3.3 Test Types We employ two test types. One-sided tests present a directional hypothesis and ask the model to answer YES or NO (e.g., “Did the minimum wage increase decrease employment? YES or NO”). Two-sided tests ask the model to predict the direction of the counterfactual (INCREASE / DECREASE / NO SIGNIFICANT CHANGE). Two-sided tests are strictly harder, as the model must both identify whether an effect exists and determine its sign. 3.4 Prompt Design We design five prompting strategies to probe different aspects of reasoning: 1. P1 – Naïve: A direct question with no guidance, capturing the model’s default “System 1” response. 2. P2 – Expert Persona: The model is told it is an “experienced econometrician” to test whether domain framing improves accuracy. 3. P3 – Chain-of-Thought (CoT): The model is asked to “think step by step,” listing causal mechanisms and confounders before answering. 4. P4 – Structured Reasoning: A more detailed template requiring explicit enumeration of causal channels, confounders, and finally a prediction. 5. P5 – Adversarial Framing: The prompt introduces a fictitious “analyst consensus” suggesting the policy had no effect, testing epistemic robustness against anchoring bias (Tversky and Kahneman, 1974). 4 Experimental Setup 4.1 Models We evaluate four frontier LLMs accessed via API: • GPT-5.2 (OpenAI) • Claude Sonnet 4 (Anthropic) • Claude Opus 4.6 (Anthropic) • GPT-4.1 (OpenAI) All models are accessed via default inference parameters, held constant across all experimental conditions. To prevent data contamination during evaluation, we sandbox each call: file-system tools are disabled, the working directory is set to a temporary folder, and custom instructions are suppressed. 4.2 Procedure Each of the 40×4×5×3=2,40040× 4× 5× 3=2,400 trials is executed independently. Responses are parsed using regular expressions to extract the final answer (INCREASE / DECREASE / NO SIGNIFICANT CHANGE / YES / NO), with a lightweight LLM fallback parser for the <1%<1\% of responses where regex extraction fails. 4.3 Statistical Analysis We analyze the binary outcome (correct/incorrect) using mixed-effects logistic regression (Bates et al., 2015): logit(P(correctij))=ij+ujlogit(P(correct_ij))=X_ij β+u_j (1) where i indexes observations and j indexes cases. The fixed effects β include model identity, prompt strategy, intuitiveness category, and their interactions (prompt × intuitiveness). The random intercept uj∼(0,σu2)u_j (0, _u^2) captures case-level heterogeneity. This specification follows recommendations in Barr et al. (2013) for confirmatory hypothesis testing. We report odds ratios (OR) with 95% confidence intervals. All analyses are implemented in R using lme4 (Bates et al., 2015). 5 Results 5.1 Overall Performance Across all 2,400 trials, overall accuracy is 81.3% (1,951/2,400). However, this aggregate masks substantial heterogeneity. One-sided tests yield 98.3% accuracy (near ceiling), while two-sided tests yield 73.1%. Within two-sided tests, accuracy on counter-intuitive cases drops to 68.8%, compared to 75.4% for obvious cases. 5.2 Model Comparison Model All Two-sided TS c-intuit. GPT-5.2 94.0% 91.1% 90.0% Claude Sonnet 4 82.5% 76.3% 66.7% Claude Opus 4.6 79.8% 70.4% 75.0% GPT-4.1 68.8% 54.6% 43.3% Table 2: Accuracy by model. TS c-intuit. = two-sided counter-intuitive cases only. GPT-5.2 maintains high performance even on counter-intuitive cases. Notably, Claude Opus 4.6 outperforms Claude Sonnet 4 on counter-intuitive cases despite lower overall accuracy. Table 2 shows that GPT-5.2 substantially outperforms all other models, particularly on two-sided counter-intuitive cases, where GPT-4.1 falls to 43.3%—only 10 percentage points above the 33% chance baseline. Notably, Claude Opus 4.6 outperforms Claude Sonnet 4 on counter-intuitive cases (75.0% vs. 66.7%) despite lower overall accuracy, suggesting different models may have different degrees of susceptibility to intuitive priors (see also Figure 3 in Appendix C). 5.3 Prompt Strategy Comparison Prompt All Two-sided TS c-intuit. P4 Structured 94.4% 92.3% 79.2% P3 CoT 92.1% 88.9% 75.0% P1 Naïve 84.0% 76.9% 70.8% P2 Expert 78.3% 69.1% 72.9% P5 Adversarial 57.7% 38.3% 45.8% Table 3: Accuracy by prompt strategy. The expert persona (P2) underperforms the naïve baseline (P1), while the adversarial prompt (P5) severely degrades overall accuracy. On two-sided counter-intuitive cases, the gap between strategies narrows substantially. Structured reasoning (P4) and CoT (P3) dominate overall, but these aggregates conceal the critical interaction with intuitiveness (§5.4). 5.4 The CoT Paradox The central finding of this paper emerges from the interaction between prompting strategy and intuitiveness. Table 4 presents the key interaction terms from our mixed-effects model. Interaction OR 95% CI p P3 CoT × c-intuit. 0.053 [0.01, 0.28] <<0.001 P4 Struct. × c-intuit. 0.034 [0.00, 0.29] 0.002 P5 Advers. × c-intuit. 2.97 [1.22, 7.23] 0.017 Random effects σu2σ^2_u (case intercept) 3.815 ICC 0.537 Table 4: Key interaction terms from the mixed-effects logistic regression (Eq. 1). CoT and structured prompting show a dramatic reduction of benefit on counter-intuitive cases (OR ≪ 1). The P5 positive interaction (p=0.017p=0.017) is confounded by test-type imbalance (see text). The CoT × counter-intuitive interaction OR of 0.053 means that CoT’s large benefit on obvious cases (main-effect OR ≈ 26) is almost entirely negated on counter-intuitive cases. Structured reasoning shows an even more pronounced reduction (OR = 0.034). The adversarial prompt (P5) shows a significant positive interaction with counter-intuitive cases (OR = 2.97, p=0.017p=0.017). However, this result is confounded by test type: 12 of 20 counter-intuitive cases are one-sided with ground truth NO, and P5’s “no effect” anchor coincidentally aligns with the correct answer. On two-sided counter-intuitive cases—where the ground truth is INCREASE or DECREASE—P5 reduces accuracy from 70.8% (P1) to 45.8%, the worst of any prompt. The apparent positive interaction thus reflects accidental alignment between the anchor and the correct answer, not genuine debiasing. Figure 1: The CoT paradox: change in accuracy from naïve to CoT prompting by intuitiveness category (two-sided cases only). CoT improves accuracy by +17 percentage points on obvious cases but only +4 points on counter-intuitive cases, with this gap driven by CoT reinforcing rather than overriding intuitive priors. Figure 1 illustrates the CoT paradox: the gain from chain-of-thought prompting diminishes dramatically as intuitiveness decreases. 5.5 Self-Consistency Analysis Majority-vote accuracy (Wang et al., 2023) across three repetitions improves marginally from 81.3% to 82.4% overall, but provides minimal improvement on two-sided counter-intuitive cases (68.8% → 70.0%). Errors are systematic, not stochastic: models consistently give the same wrong answer, reflecting stable but incorrect causal beliefs. 6 Analysis 6.1 Error Patterns Among incorrect two-sided responses where the ground truth is INCREASE, 75% of errors predict DECREASE (directional reversal) rather than NO SIGNIFICANT CHANGE—consistent with models applying an intuitive causal model that predicts the opposite direction. The expert persona (P2) shows a distinctive pattern: 70% of its errors are NO SIGNIFICANT CHANGE, suggesting that “econometrician” framing induces excessive hedging. 6.2 Case Difficulty The random intercepts (uju_j) reveal substantial case-level heterogeneity. The hardest cases are Case 23 (daycare fines; 42%) and Case 36 (right-to-carry laws; 42%), both counter-intuitive and two-sided. Figure 2: Heatmap of accuracy by model and case across all prompt strategies and repetitions. Red cells indicate systematic failures; cases are sorted by intuitiveness. Notable failure clusters appear on Cases 23 (daycare fines) and 36 (right-to-carry laws). Figure 2 reveals the full pattern: while most cases are solved reliably by at least one model, the hardest counter-intuitive cases produce systematic failures across all models. 6.3 Familiarity as a Non-Factor Adding log(citations) as a moderator yields a non-significant effect (OR = 1.271, p=0.53p=0.53). Models are not simply “looking up” well-known results: Case 23 (daycare late pickup fines), based on one of the most widely cited behavioral economics experiments (Gneezy and Rustichini, 2000), achieves only 42% accuracy. This speaks to the “causal parrots” debate (Zečević et al., 2023): if LLMs merely recited memorized facts, accuracy should correlate with citation frequency. The null result suggests models have absorbed causal knowledge broadly, but deploying it correctly requires a reasoning step that fails when intuitions conflict. 7 Discussion 7.1 Slow Talking, Not Slow Thinking Our CoT paradox challenges the assumption that CoT elicits deeper reasoning. On counter-intuitive cases, CoT appears to reinforce rather than override intuitive priors: the reasoning chain’s first step is anchored by the same prior that produces the wrong answer, and subsequent steps rationalize this anchor. This is consistent with Jacovi et al. (2024), who show that reasoning chains are only as strong as their weakest step. From the dual-process perspective (Kahneman, 2011), effective System 2 reasoning requires inhibiting System 1. Our results suggest that LLMs’ CoT performs the form of System 2 processing without the function—“slow talking” rather than “slow thinking.” 7.2 Implications for Policy Applications The overall accuracy of 81.3% is inflated by obvious cases. On counter-intuitive cases—where careful analysis matters most—accuracy drops substantially. Policymakers relying on LLM-assisted analysis may receive confident but incorrect guidance precisely where it is most consequential. 7.3 Limitations Our benchmark of 40 cases is sufficient for mixed-effects modeling but limited in breadth. The intuitiveness classification involves authorial judgment, though validated against LLM response patterns. We evaluate models via API at a single point in time (March 2026); performance may vary across versions. 8 Conclusion We introduce a benchmark for evaluating LLM counterfactual reasoning in policy evaluation, organized around an intuitiveness dimension inspired by dual-process theory. The CoT paradox reveals that chain-of-thought prompting helps primarily when the correct answer aligns with intuition; its benefit is nearly eliminated on counter-intuitive cases. Combined with the failure of self-consistency and the null effect of familiarity, this points to a fundamental limitation: current LLMs can retrieve causal knowledge but struggle to reason with it when doing so requires overriding intuitive priors. Improving counterfactual reasoning may require innovations that enable genuine inhibitory control—the ability to recognize when one’s first instinct is wrong. Ethics Statement This study evaluates proprietary LLMs via their public API interfaces and does not involve human subjects experimentation. The benchmark cases are drawn entirely from published, peer-reviewed research in economics and social science; no private or sensitive data were collected or used. All experimental prompts and model outputs are stored locally and do not contain personally identifiable information. We acknowledge that LLM-assisted policy analysis carries risks if deployed without appropriate scrutiny. Our finding that LLMs systematically fail on counter-intuitive cases—precisely where careful analysis is most valuable—underscores the importance of human oversight in policy-relevant applications. We caution against using LLM outputs as a substitute for rigorous empirical analysis and expert judgment in policy evaluation. References Barr et al. (2013) Dale J Barr, Roger Levy, Christoph Scheepers, and Harry J Tily. 2013. Random Effects Structure for Confirmatory Hypothesis Testing: Keep It Maximal. Journal of Memory and Language, 68(3):255–278. Bates et al. (2015) Douglas Bates, Martin Mächler, Ben Bolker, and Steve Walker. 2015. Fitting Linear Mixed-Effects Models Using lme4. Journal of Statistical Software, 67(1):1–48. Brown et al. (2020) Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 1877–1901. Dahl and DellaVigna (2009) Gordon Dahl and Stefano DellaVigna. 2009. Does Movie Violence Increase Violent Crime? Quarterly Journal of Economics, 124(2):677–734. Donohue et al. (2019) John J. Donohue, Abhay Aneja, and Kyle D. Weber. 2019. Right‐to‐Carry Laws and Violent Crime: A Comprehensive Assessment Using Panel Data and a State‐Level Synthetic Control Analysis. Journal of Empirical Legal Studies, 16(2):198–247. Evans (2003) Jonathan St BT Evans. 2003. In Two Minds: Dual-Process Accounts of Reasoning. Trends in Cognitive Sciences, 7(10):454–459. Gneezy and Rustichini (2000) Uri Gneezy and Aldo Rustichini. 2000. A Fine is a Price. The Journal of Legal Studies, 29(1):1–17. Huang et al. (2024) Yinya Huang, Ruixin Hong, Hongming Zhang, Wei Shao, Zhicheng Yang, Dong Yu, Changshui Zhang, Xiaodan Liang, and Linqi Song. 2024. CLOMO: Counterfactual Logical Modification with Large Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). Jacovi et al. (2024) Alon Jacovi, Yonatan Bitton, Bernd Bohnet, Jonathan Herzig, Or Honovich, Michael Tseng, Michael Collins, Roee Aharoni, and Mor Geva. 2024. A Chain-of-Thought Is as Strong as Its Weakest Link: A Benchmark for Verifiers of Reasoning Chains. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). Jin et al. (2024) Zhijing Jin, Yuen Chen, Felix Leber, Luigi Gresele, Ojasv Kamath, and 1 others. 2024. CLadder: A Benchmark to Assess Causal Reasoning Capabilities of Language Models. Advances in Neural Information Processing Systems (NeurIPS), 36. Kahneman (2011) Daniel Kahneman. 2011. Thinking, Fast and Slow. Farrar, Straus and Giroux. Kıcıman et al. (2023) Emre Kıcıman, Robert Ness, Amit Sharma, and Chenhao Tan. 2023. Causal Reasoning and Large Language Models: Opening a New Frontier for Causality. arXiv preprint arXiv:2305.00050. Kojima et al. (2022) Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large Language Models are Zero-Shot Reasoners. In Advances in Neural Information Processing Systems (NeurIPS), volume 35. Mondorf and Plank (2024) Philipp Mondorf and Barbara Plank. 2024. Comparing Inferential Strategies of Humans and Large Language Models in Deductive Reasoning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). Mu and Li (2024) Feiteng Mu and Wenjie Li. 2024. A Causal Approach for Counterfactual Reasoning in Narratives. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). OpenAI (2023) OpenAI. 2023. GPT-4 Technical Report. arXiv preprint arXiv:2303.08774. Shen et al. (2025) Guobin Shen, Dongcheng Zhao, Aorigele Bao, Xiang He, Yiting Dong, and Yi Zeng. 2025. StressPrompt: Does Stress Impact Large Language Models and Human Performance Similarly? In Proceedings of the 39th AAAI Conference on Artificial Intelligence (AAAI). Snell et al. (2025) Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. Scaling LLM Test-Time Compute Optimally Can be More Effective than Scaling Parameters for Reasoning. In Proceedings of the 13th International Conference on Learning Representations (ICLR). Stanovich and West (2000) Keith E Stanovich and Richard F West. 2000. Individual Differences in Reasoning: Implications for the Rationality Debate? Behavioral and Brain Sciences, 23(5):645–665. Tversky and Kahneman (1974) Amos Tversky and Daniel Kahneman. 1974. Judgment Under Uncertainty: Heuristics and Biases. Science, 185(4157):1124–1131. Wang et al. (2023) Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-Consistency Improves Chain of Thought Reasoning in Language Models. In Proceedings of the 11th International Conference on Learning Representations (ICLR). Wei et al. (2022) Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems (NeurIPS), volume 35. Weinzierl and Harabagiu (2024) Maxwell Weinzierl and Sanda Harabagiu. 2024. Tree-of-Counterfactual Prompting for Zero-Shot Stance Detection. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). Ying et al. (2024) Jiahao Ying, Yixin Cao, Kai Xiong, Long Cui, Yidong He, and Yongbin Liu. 2024. Intuitive or Dependent? Investigating LLMs’ Behavior Style to Conflicting Prompts. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). Zečević et al. (2023) Matej Zečević, Moritz Willig, Devendra Singh Dhami, and Kristian Kersting. 2023. Causal Parrots: Large Language Models May Talk Causality But Are Not Causal. Transactions on Machine Learning Research. Appendix A Full Case List The complete list of 40 benchmark cases, including policy descriptions, countries, years, key papers, intuitiveness classifications, test types, and ground truth answers, is available from the author upon request. All case DOIs have been verified against the Crossref API. Appendix B Prompt Templates The five prompt templates (P1–P5) are available from the author upon request. For two-sided tests, the response options are INCREASE / DECREASE / NO SIGNIFICANT CHANGE. For one-sided tests, the options are YES / NO. Appendix C Additional Results Additional analyses including per-case accuracy tables, confusion matrices, and BLUP estimates are available from the author upon request. Figure 3: Accuracy by model and intuitiveness category on two-sided cases. GPT-5.2 maintains high accuracy across categories, while GPT-4.1 drops to 43.3% on counter-intuitive cases.

Instant research discovery