Research map / Large language models / Instruction and fine-tuning
Reasoning Elicitation in Language Models via Counterfactual Feedback
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes causal-consistency metrics (N-IR, S-IR, AN-IR, AS-IR) and fine-tuning with counterfactual feedback (SFT, DPO, and causal consistency feedback DPO+CCF) to improve causal reasoning of Phi-3-mini, evaluated in-domain, across generalization modes, and on three real-world causal problems.
Problem
Language models answer factual questions much better than counterfactual ones, so their causal reasoning (necessity and sufficiency) is weak and standard error-rate metrics do not capture consistency between factual and counterfactual answers.
Approach
Define unit-wise necessity/sufficiency inconsistency rates that score factual and counterfactual answers jointly per context, and fine-tune with preference data built from dialogues pairing factual and counterfactual questions (causal consistency feedback) so that DPO targets these rates rather than per-question correctness.
Key results
- In-domain, DPO+CCF reduces Avg-ER and Avg-IR to 0.48 and 0.47 of base (normalized), versus 0.42/0.44 for SFT-F&CF and 0.86/0.72 for OnlyF DPO.
- Across generalization modes, the common-cause mode shows errors increasing for F&CF methods (DPO+CCF normalized Avg-ER 2.04), indicating no transfer there.
- Inductive generalization yields reductions (SFT-F&CF Avg-ER 0.60, DPO+CCF 0.69 normalized) while deductive effect-based is near base for most methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add paired factual and counterfactual examples to in-domain fine-tuning | Phi-3 mini · Candy party puzzle, in-domain | S-IR | Helped |
| Add causal consistency feedback to paired factual-counterfactual DPO | Phi-3 mini · Candy party puzzle, in-domain | S-IR | Helped |
| Use factual-only examples instead of paired factual-counterfactual examples (ablation) | SFT and DPO fine-tuning · Candy party puzzle, in-domain | S-IR | Hurt |
| Use counterfactual-only examples instead of paired factual-counterfactual examples (ablation) | SFT and DPO fine-tuning · Candy party puzzle, in-domain | S-IR | Hurt |
| Add paired factual-counterfactual examples | Fine-tuning methods · Real-world in-domain average | Avg-ER: 1.00 -> 0.42 | Helped |
| Add paired factual-counterfactual examples | Fine-tuning methods · Real-world in-domain average | Avg-IR: 1.00 -> 0.44 | Helped |
| Add paired factual-counterfactual examples | Fine-tuning methods · Real-world common-cause generalization | Avg-ER: 1.00 -> 1.17 | Hurt |
| Add paired factual-counterfactual examples | Fine-tuning methods · Real-world common-effect generalization | Avg-ER: 1.00 -> 0.53 | Helped |
| Add paired factual-counterfactual examples | Fine-tuning methods · Real-world inductive generalization | Avg-IR: 1.00 -> 0.50 | Helped |
| Add paired factual-counterfactual examples | Fine-tuning methods · Real-world deductive, cause-based generalization | Avg-ER: 1.00 -> 0.57 | Helped |
| Add paired factual-counterfactual examples | Fine-tuning methods · Real-world deductive, effect-based generalization | Avg-IR: 1.00 -> 0.76 | Helped |
| Use causal consistency preferences rather than correctness preferences | DPO with factual-counterfactual feedback · Real-world average across modes | Avg-IR | Mixed |
Surprising findings
- Fine-tuning on only the target cause or effect reduces N-IR but does not reduce S-IR, and common-cause generalization often increases error relative to base.
- Phi-3-mini's factual error rate is much lower than its counterfactual error rate, showing a recall-reasoning gap.
Limitations
- Causes and effects restricted to binary variables.
- Generalization results vary strongly across modes; common-cause and effect-based deduction do not improve.
- Evaluation limited to a single small model (Phi-3-mini) and handcrafted/synthetic problems.
Experimental setup
Datasets: synthetic candy party puzzle breast cancer treatment causal model transmission line fault detection GSM8K
Benchmarks: candy party puzzle (in-domain) generalization modes (common-cause, common-effect, inductive, deductive) breast cancer treatment transmission line protection GSM8K-based math scenarios
Metrics: F-ER CF-ER Avg-ER N-IR S-IR AN-IR AS-IR Avg-IR
Baselines: Phi-3-mini base SFT-OnlyF DPO-OnlyF SFT-OnlyCF DPO-OnlyCF SFT-OnlyFx2 DPO-OnlyFx2
Builds on: DPO probabilities of causation
Closest papers by meaning
- What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models (CVPR 2024 · similarity 0.89)
- AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions (NeurIPS 2025 · similarity 0.89)
- Counterfactual reasoning: an analysis of in-context emergence (NeurIPS 2025 · similarity 0.88)
- Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement (ICLR 2024 · similarity 0.88)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.88)
- Linguini: A benchmark for language-agnostic linguistic reasoning (NeurIPS 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.