Research map / Large language models / Language, safety and interpretability
Ice Cream Doesn’t Cause Drowning: Benchmarking LLMs Against Statistical Pitfalls in Causal Inference
Jin Du, Li Chen, Xun Xian, An Luo, Fangqiao Tian, Ganghua Wang, Charles Doss, Xiaotong Shen, Jie Ding
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use code-assisted prompting instead of direct prompting | GPT-4.1 · CausalPitfalls | causal reliability: 25.24 -> 37.32 | Improved |
| Use code-assisted prompting instead of direct prompting | Gemini-2.0-flash · CausalPitfalls | causal reliability: 24.36 -> 31.80 | Improved |
| Use code-assisted prompting instead of direct prompting | GPT-o4-mini · CausalPitfalls | causal reliability: 40.72 -> 43.03 | Improved |
| Use code-assisted prompting instead of direct prompting | Mistral-7b · CausalPitfalls | causal reliability: 14.43 -> 7.65 | Hurt |
| Use code-assisted prompting instead of direct prompting | Llama3.1-8b · CausalPitfalls | causal reliability: 17.86 -> 12.86 | Hurt |
| Use code-assisted prompting instead of direct prompting | Gemma2-9b · CausalPitfalls | causal reliability: 13.69 -> 13.24 | No clear effect |
| Use code-assisted prompting instead of direct prompting | GPT-o4-mini · CausalPitfalls very hard questions | causal reliability: 17.75 -> 32.84 | Improved |
| Use code-assisted prompting instead of direct prompting | GPT-o4-mini · CausalPitfalls very easy questions | causal reliability: 60.72 -> 56.73 | Hurt |
| Remove hints from question (very easy to very hard) | GPT-o4-mini direct prompting · CausalPitfalls | causal reliability: 60.72 -> 17.75 | Hurt |
| Change brand label from HealthPlus to UltraSugar with identical data | GPT-4o and Gemini-2.0-flash · Branding bias test | conclusion correctness | Hurt |
| Allow one debugging attempt for code-assisted prompting | Mistral-7B and Llama-8B · CausalPitfalls | causal reliability | Improved |
Closest papers by meaning
- Can Large Language Models Infer Causation from Correlation? (ICLR 2024 · similarity 0.92)
- Unveiling Causal Reasoning in Large Language Models: Reality or Mirage? (NeurIPS 2024 · similarity 0.91)
- Causality can systematically address the monsters under the bench(marks) (NeurIPS 2026 · similarity 0.90)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.89)
- COLD: Causal reasOning in cLosed Daily activities (NeurIPS 2024 · similarity 0.88)
- Causal Order: The Key to Leveraging Imperfect Experts in Causal Inference (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
