When Many Answers Are Valid, Voting Fails: Symbolic Verification for Best-of-K Causal Reasoning in LLMs

Analysis

Why This Paper Matters

This paper addresses a critical flaw in the common practice of using self-consistency (majority voting) to improve LLM reasoning. While voting works well for tasks with a single correct answer, it fails in causal reasoning where multiple answers can be valid and errors are often systematic. The authors show that on CLEAR find-one-valid queries, plurality voting, reward models, LLM judges, and even model confidence all plateau around 30% accuracy, while a symbolic verifier (CALVER) achieves 42.1%. This is a significant finding because it suggests that current evaluation and aggregation methods are inadequate for causal reasoning tasks, and that scaling model size alone does not solve the problem.

The paper's significance lies in its proposal of a training-free, symbolic approach that leverages Pearl's causal criteria to verify reasoning traces. This is a departure from purely neural methods and highlights the potential of hybrid neuro-symbolic systems. By showing that a lightweight symbolic verifier can outperform much larger models, the authors make a strong case for incorporating causal structure into LLM reasoning pipelines.

Technical Contributions

• CALVER: A training-free symbolic verifier that scores each candidate reasoning trace against Pearl's causal axioms (d-separation, backdoor adjustment, intervention) given a causal structure.
• No reference answer needed: CALVER selects the highest-scoring trace without consulting a gold answer, making it suitable for open-ended reasoning.
• Efficiency: Runs in milliseconds on CPU, making it practical for real-time applications.
• Generalization: Works across multiple Bayesian networks, model families, and settings where the causal graph is built from text.
• Extensions: Applies to thresholded average-treatment-effect (ATE) decisions and logic tasks with a truth-table checker.

Results

On CLEAR find-one-valid queries, CALVER achieves 42.1% accuracy, while plurality voting, a reward model, an LLM judge, and model confidence all remain near 30% on identical frozen pools. Scaling the judge to 72B does not close the gap. In an audited clean-core subset, 11 of 21 graph-valid CALVER selections differ from the benchmark's listed answer but still satisfy the requested predicate, suggesting that the benchmark may have incomplete ground truth. The advantage widens with sampling budget and reproduces across ten published Bayesian networks and a second model family. CALVER also improves thresholded ATE decisions against exact ground truth and generalizes to logic under a truth-table checker.

Significance

This work challenges the assumption that majority voting is reliable for causal reasoning and demonstrates the value of symbolic verification for selecting valid answers. It suggests that future LLM reasoning systems should incorporate causal structure and verification mechanisms to handle tasks with multiple valid answers. The findings also have implications for benchmark design, as they reveal potential issues with existing ground truth labels. CALVER's efficiency and training-free nature make it an attractive addition to LLM pipelines, potentially improving reliability in causal reasoning and decision-making applications.