📊 statistics

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

The paper introduces CALVER, a training-free symbolic verifier that outperforms traditional voting and reward-based methods in causal reasoning by scoring structured reasoning traces against Pearl's causal criteria to identify valid answers even when multiple correct solutions exist.

Original authors: Omatharv Bharat Vaidya, Connor Thomas Jerzak, Zayne Rea Sprague, Fangcong Yin, Nhat Ho

Published 2026-08-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Omatharv Bharat Vaidya, Connor Thomas Jerzak, Zayne Rea Sprague, Fangcong Yin, Nhat Ho

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a mystery, like figuring out why a plant in your room is dying. You might ask a super-smart AI detective to look at the clues. The AI doesn't just give you one answer; it tries to think through the problem ten different times, like a detective trying out ten different theories. Usually, if the AI is smart, most of those ten theories will be the same, and the one that appears most often is probably the right one. This is called "voting" or "self-consistency," and it works great for math problems or simple logic puzzles where there is only one right answer.

But what happens when the mystery has many right answers? Imagine the plant is dying because of three different possible causes: too much sun, not enough water, or a bug. All three are valid reasons. If you ask the AI to think ten times, it might guess "sun" three times, "water" three times, and "bug" three times. The votes are split! Meanwhile, the AI might accidentally guess "the plant is hungry" four times. Even though "hungry" is a silly, wrong answer, it wins the vote just because it was the most popular guess. This is the tricky problem this paper tackles: when there are many correct solutions, the usual "majority rules" method can actually pick the wrong one.

The researchers, working in the field of artificial intelligence and causal reasoning (which is just a fancy way of saying "figuring out cause and effect"), discovered that this "split vote" problem is a major headache for AI. They found that when an AI is asked to find any valid way to fix a problem, the correct answers often get scattered across many different options, while a single wrong answer can accidentally become the most popular.

To fix this, the team invented a new tool called CALVER (Causal Axiom-Level VERification). Think of CALVER as a strict, rule-following referee instead of a popularity contest. Instead of counting how many times an answer appears, CALVER checks every single guess against a set of unbreakable rules of cause and effect. It asks, "Does this answer actually make sense according to the laws of physics and logic?" If an answer follows the rules, it gets a high score, even if it's the only one of its kind. If an answer breaks the rules, it gets a zero, even if it was the most popular guess.

The paper shows that this referee approach works much better than just counting votes. In their tests, CALVER was able to find the right answer about 42.1% of the time, while the old "vote for the most popular" method only got it right about 30% of the time. This gap got even bigger when they let the AI try more times (up to 32 tries), with CALVER pulling ahead by a huge margin. They also proved that this works even when the AI has to read a messy story to figure out the rules, not just when the rules are given clearly.

The authors are very careful to say that this isn't magic; it's a specific fix for a specific problem. They proved mathematically that when there are many valid answers, voting will often fail, and they showed through experiments that checking the rules works. They even tested this on logic puzzles that have nothing to do with plants or causes, and the same "check the rules" idea still worked. However, they also noted that if the problem is simple and has only one right answer, the old voting method is still fine. But for those tricky situations where there are many right ways to solve a puzzle, CALVER is the new champion that stops the AI from getting tricked by popularity.

Technical Summary: When Many Answers Are Valid, Voting Fails

Problem Statement
Current inference-time scaling methods for Large Language Models (LLMs), such as Self-Consistency, rely on the assumption that the most frequent answer among sampled reasoning traces is the most reliable. This approach fails in causal reasoning tasks where multiple distinct outputs can be fully correct (e.g., different valid adjustment sets satisfying Pearl's backdoor criterion). In such "many-satisfier" scenarios, valid probability mass fragments across multiple answer strings, while a single invalid answer may dominate the vote simply by being the most common error. Consequently, exact plurality voting can select an invalid answer even when the majority of sampled traces contain valid reasoning, leading to inconsistent and incorrect selections.

Methodology: CALVER
The authors introduce CALVER (Causal Axiom-Level VERification), a training-free, symbolic verifier designed to select the best candidate from a pool of $K$ sampled traces without consulting a reference answer.

Trace Contract: CALVER requires models to output reasoning in a typed six-slot schema: (1) Graph, (2) Query, (3) Strategy, (4) Derivation Record, (5) Computed Result, and (6) Answer.
Symbolic Verification: Instead of aggregating answer frequencies, CALVER scores each trace deterministically against an executable causal specification. The verifier performs six checks:
1. Graph Parsing: Validates the graph structure (DAG or ADMG).
2. Query Binding: Ensures the query variables match the problem.
3. Strategy Validity: Checks if the proposed set satisfies graphical criteria (e.g., d-separation, m-separation, backdoor adjustment, intervention reachability).
4. Derivation Record: Verifies the syntax and provenance of the reasoning steps.
5. Computed Result: Independently recomputes the functional (e.g., Average Treatment Effect) to ensure numerical consistency.
6. Answer Consistency: Confirms the final answer aligns with the computed result.
Selection Rule: The selector chooses the trace with the highest score (maximum deployed score of 6). Ties are broken by selecting the earliest trace.
Modes: CALVER operates in two modes:
- Supplied-Graph: The source graph is provided; traces are scored against it.
- Constructed-Graph: The model must extract a graph from text; the trace is scored against its own reconstructed graph, while final grading uses the hidden source graph.

Key Contributions

Formulation: The paper formulates best-of- $K$ selection under answer multiplicity as aggregation over a decidable validity class. CALVER instantiates this with a six-component graph-dependent trace score for DAGs and ADMGs.
Theory:
- Strict ATE Certificate Soundness: The authors prove that a maximum-score trace satisfying specific guards (treatment exclusion and decision margin) guarantees a correct threshold decision for Average Treatment Effect (ATE) queries.
- Validity Fragmentation Theorem: They prove that exact plurality can become inconsistent when valid probability mass is divided among multiple valid answers, allowing an invalid answer with slightly higher individual mass to win.
Experiments: Extensive evaluation across multiple model families (Qwen, Mistral), graph families (CLEAR, bnlearn), and settings (supplied graphs, text-to-graph construction, logic puzzles).

Results

Superiority over Baselines: On the CLEAR clean-core benchmark (126 problems with multiple valid answers), CALVER achieves 42.1% accuracy, significantly outperforming exact plurality (~30.9%), a set-aware medoid, model confidence, a learned reward model (Skywork Reward V2), and an LLM judge (even when scaled to 72B parameters).
Scaling with $K$ : The performance gap between CALVER and plurality widens as the sampling budget ( $K$ ) increases. While plurality performance plateaus, CALVER continues to improve, reaching 57.9% at $K=32$ .
Robustness to Graph Recovery: In text-to-graph settings where exact graph recovery is low (e.g., 34% at naturalistic narrative levels), CALVER still outperforms plurality by 17.6 percentage points. This demonstrates that candidate-wise verification can succeed even when the global graph structure is imperfectly reconstructed, provided the query-local relations are preserved.
Generalization: The method transfers to independent benchmarks (bnlearn, CausalGraph2LLM), different model families, and non-causal domains (Knights and Knaves logic puzzles) when the verification predicate is adapted (e.g., truth-table consistency).
Efficiency: Symbolic verification adds minimal overhead (~7% over generation cost), scoring each trace in 1–8 ms on a single CPU thread.

Significance and Claims
The paper claims that CALVER provides a stronger selection signal than surface frequency or learned preference models in regimes where multiple valid answers exist. It establishes that:

Voting Fails: Exact plurality is fundamentally flawed for causal reasoning when valid answers fragment the vote.
Executable Validity: Aggregating based on the satisfaction of a task-defined validity class (via symbolic checking) is more reliable than aggregating based on answer string frequency.
Regime Bound: The method is most effective in the "intermediate regime" where many answers are valid and no single extraction is dependable. When answers are near-unique, standard voting suffices; when text-to-graph extraction is highly reliable, extracting one graph and solving directly is superior.
Independence: The selection mechanism relies on executable predicates (causal axioms or logic rules) rather than learned heuristics, making it robust across different model architectures and training objectives.

The authors emphasize that CALVER does not claim to determine scientific truth about real-world causal graphs but rather selects the most formally valid reasoning trace given the provided structure or text. The method's advantage is strictly tied to the ability to execute a validity check on the candidate's reasoning.

Technical Summary: When Many Answers Are Valid, Voting Fails

More like this