Reasoning Elicitation in Language Models via Counterfactual Feedback

Alihan Hüyük, Xinnuo Xu, Jacqueline R. M. A. Maasch, Aditya V. Nori, Javier González

Abstract

Large Language Models (LLMs) often exhibit a significant "reasoning-recall gap," where they perform well on factual queries but fail on counterfactual interventions. This paper introduces a framework for reasoning elicitation that prioritizes causal consistency during fine-tuning. By training models on paired factual and counterfactual questions and rewarding logical coherence—specifically necessity and sufficiency—the authors demonstrate substantial improvements in causal reasoning across synthetic and real-world domains. The results suggest that training on the relationships between states, rather than individual answer correctness, is key to bridging the reasoning gap.

Introduction: The Reasoning-Recall Gap

While LLMs demonstrate high accuracy on many benchmarks, it remains unclear whether they are genuinely reasoning or simply recalling statistical correlations from their training data. This ambiguity is most evident in causal tasks: a model might correctly state a factual relationship (e.g., "A causes B") but fail to predict the outcome of a counterfactual intervention (e.g., "What if A had not occurred?").

Figure 2: The reasoning-recall gap illustrated through a divisibility problem. While the model handles the factual question relatively well, the error rate spikes for the counterfactual variant.

The authors identify this as the "reasoning-recall gap." Standard fine-tuning often focuses on individual answer correctness, which fails to teach the model the underlying causal mechanism. This research aims to bridge this gap by forcing models to coordinate their factual and counterfactual "beliefs" through structured feedback.

Methodology

The researchers propose a framework for Reasoning Elicitation built on three pillars:

1. Causal Consistency Metrics

Beyond simple accuracy, the study introduces Necessity and Sufficiency Inconsistency Rates (N-IR & S-IR). These metrics evaluate whether a model's answers across factual and counterfactual scenarios are logically consistent. For example, if a model identifies factor $X$ as necessary for outcome $Y$, it must consistently predict that removing $X$ results in the absence of $Y$.

2. Fine-tuning with Counterfactual Feedback

The study compares three primary fine-tuning approaches:

  • Supervised Counterfactual Feedback (SFT-F&CF): Direct training on ground-truth pairs of factual and counterfactual questions.
  • Preference-based Counterfactual Feedback (DPO-F&CF): Utilizing Direct Preference Optimization (DPO) to favor correct answers over incorrect ones on a per-question basis.
  • Preference-based Causal Consistency Feedback (DPO+CCF): A novel approach that applies DPO to multi-turn dialogues. It rewards the model when its joint answers across factual and counterfactual prompts satisfy causal consistency.

3. Generalization Taxonomy

To understand how reasoning transfers, the authors categorize problem structures into four modes: Common-Cause, Common-Effect, Inductive (transitive reasoning), and Deductive.

Figure 4: Taxonomy of causal generalization modes used to evaluate how models transfer reasoning capabilities across different logical structures.

Experimental Setup and Datasets

The methods were evaluated using a diverse set of synthetic and real-world-inspired problems:

  • Candy Party Puzzle: A synthetic logical puzzle testing pure causal structures.
  • Healthcare (Breast Cancer Treatment): A model based on clinical guidelines for determining surgery vs. therapy.
  • Engineering (Fault Detection): An algorithm for identifying electrical transmission line faults.
  • Math Benchmarking (GSM8K): A modified version of grade-school math problems framed as causal systems.

Key Results and Insights

  • DPO+CCF Superiority: In the "Candy Party" puzzle, the DPO+CCF model achieved the most substantial improvements across all metrics, outperforming models trained only on isolated factual or counterfactual data.
  • Real-World Efficacy: In the healthcare domain, SFT-F&CF reduced the Average Error Rate (Avg-ER) from 1.0 (Base) to 0.42, demonstrating that counterfactual training translates to practical decision-making tasks.
  • The Inductive Advantage: Fine-tuning generalizes most effectively in an inductive mode (learning $A \to B$ and $B \to C$ to infer $A \to C$). Conversely, deductive generalization (inferring the middle of a chain from its ends) remains a significant challenge for current LLMs.
  • Risks of Factual-Only Training: In certain structures like Common-Cause, fine-tuning exclusively on factual data actually increased error rates. This suggests that factual-only training can reinforce simple recall at the expense of structural understanding.

Conclusion

This research demonstrates that correctness is not a proxy for reasoning. By shifting the focus of fine-tuning from individual answers to the logical relationships between factual and counterfactual states, models develop more robust internal mechanisms. Future work will look to scale these techniques to more complex, non-binary causal webs and explore the intersection of causal and symbolic reasoning.

Summary updated

BibTeX
@inproceedings{huyuk2025reasoning,
  author = {Alihan Hüyük and Xinnuo Xu and Jacqueline R. M. A. Maasch and Aditya V. Nori and Javier González},
  title = {{Reasoning Elicitation in Language Models via Counterfactual Feedback}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=VVixJ9QavY},
}
Figure