Research map / Large language models / Agents, code and math
CounterBench: Evaluating and Improving Counterfactual Reasoning in Large Language Models
Yuefei Chen, Vivek K. Singh, Jing Ma, Ruixiang Tang
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 CoIn reasoning instead of CausalCoT | GPT-4o · CounterBench | accuracy: 75.8 -> 89.4 | Improved |
| use CoIn reasoning instead of CausalCoT | GPT-4o mini · CounterBench | accuracy: 59.8 -> 79.9 | Improved |
| use CoIn reasoning instead of CausalCoT | GPT-3 (Babbage-002) and Davinci-002 · CounterBench | accuracy: 49.3 -> 49.6 (Davinci-002) | No clear effect |
| add CausalCoT prompting over standard prompting | most LLMs · CounterBench | accuracy: 52.8 -> 75.8 (GPT-4o); 50.6 -> 49.1 (GPT-3.5) | Mixed |
| add tool-based Solver strategy (CausalTool re-implementation) | LLMs · CounterBench | accuracy: 50.4 -> 47.2 (GPT-4o mini) | Hurt |
| apply CoIn to Gemini-1.5-flash instead of standard prompting | Gemini-1.5-flash · CounterBench | accuracy: 68.0 -> 89.9 | Improved |
| apply CoIn on CLADDER counterfactual subset | Gemini-1.5 · CLADDER counterfactual | accuracy (%): 64.77 -> 78.98 | Improved |
| switch CLADDER queries from commonsense to anti-commonsense (ablation) | CoIn · CLADDER counterfactual | accuracy | No clear effect |
| increase number of events (variables) per question (ablation) | CausalCoT and CoIn · CounterBench | accuracy | Hurt |
| apply CoIn to small models (GPT-4o mini, Claude-3.5 Haiku, Gemini-1.5-flash-8b) | small LLMs · CounterBench | accuracy: 50.4 -> 79.9 (GPT-4o mini, standard vs CoIn) | Improved |
Closest papers by meaning
- Reasoning Elicitation in Language Models via Counterfactual Feedback (ICLR 2025 · similarity 0.91)
- What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models (CVPR 2024 · similarity 0.91)
- Counterfactual reasoning: an analysis of in-context emergence (NeurIPS 2025 · similarity 0.91)
- RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.89)
- Abstract Counterfactuals for Language Model Agents (NeurIPS 2025 · similarity 0.88)
- CofCA: A STEP-WISE Counterfactual Multi-hop QA benchmark (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.
