Why Majority Voting Fails LLMs: ARBITER, Reasoning Trajectory Basins, and the Wrong-Majority Problem

Overview

This post reviews ARBITER: Reasoning Trajectory Basins and Majority Vote Failures in Test-Time Sampling (arXiv:2605.26172) by Meng Cai, Lars Kulik, and Farhana Choudhury (University of Melbourne, School of Computing and Information Systems; cs.LG).

The author opens with an analogy: copying 24 classmates' answers, where 20 pick A and 4 pick B—and A is wrong. The twist: the 20 'A' answers arrived via several entirely different reasoning paths that merely collided on the final answer. They were never independent votes.

Key points

1. The voting illusion

• Self-consistency generates multiple reasoning chains per question and picks the most frequent final answer. It reliably beats greedy decoding—but the chains are not independent.
• Chains cluster into a few 'reasoning basins', each defined by a normalized final answer (e.g., 18 chains in one basin, 4 in another, several singletons).
• Majority voting selects the largest basin, not the most coherent or accurate one. Shallow reasoning paths are easier for the model to reproduce, so wrong answers can win on volume. This is the 'wrong-majority' failure mode: the correct answer is in the pool but outvoted.
• Core insight: majority voting picks 'who is loudest,' not 'who is smartest.'

2. Structure ≠ truth

• The paper tests whether internal signals—chain coherence, hidden-state cluster structure, trajectory-graph topology—can identify the correct basin. They cannot.
• Re-scoring by the model, hidden-state basin scoring, graph-neural routing, and picking the most coherent basin all underperform plain majority vote.
• A wrong chain can be perfectly self-consistent; 'similar errors' are still errors. The paper's 'negative ladder' section cataloging failed methods is central, not filler: structure signals are real and informative, but surprisingly orthogonal to correctness.

3. Arbiter-Δ: layering evidence over consensus

• Since direct replacements hurt the strong consensus baseline, the authors treat consensus as a prior that can only be overturned by additional same-model evidence.
• Arbiter-Δ adds a 'challenger basin score' on top of majority vote using three evidence sources from the same model:
• Semantic frames: the model paraphrases each basin's core understanding and re-solves with that framing
• Frame panel (ablation only): side-by-side basin comparison for re-scoring
• Guided re-solve: 'assume the answer is X and re-derive it'
• Evidence is merged via a log-linear pooling formula (a log-ratio sum: prior vote ratio + reliability-weighted frame ratio + reliability-weighted guided ratio). If positive, switch to the challenger.
• No learnable parameters: α = 1.0 is a fixed Laplace smoothing constant; reliability weighting automatically down-weights evidence sources whose outputs fall outside the basin pair.

4. Results: slow and careful wins

• Across 3 models (Qwen3-4B, Llama-3.1-8B, Phi-4) × 3 benchmarks (GSM8K, MMLU-HS-Math, MATH-500): 8 of 9 cells gain accuracy, 1 unchanged, none negative.
• Largest single gain: Llama-3.1-8B on MATH-500, +3.0pp (51.60% → 54.60%). Average gains: Llama-3.1-8B +1.77pp, Qwen3-4B +0.23pp (ceiling effect, baseline >94%), Phi-4 +0.54pp.
• Only 168 arbitrations (<5% of questions): 78 correct corrections, 35 harmed correct consensus, net +43 recovered answers. High precision, low coverage—deliberately conservative, since against a 94% baseline, random perturbations cause far more correct→wrong regressions than fixes.

5. Basin Story Graph

• A visualization tool clusters sampled chains by answer basin and colors edges by each evidence source's support. Even when routing is disabled, it serves as a diagnostic for *why* consensus failed on a given question.
• Deeper message: LLM chains are not 24 unrelated random samples—they branch into a few river channels, each internally coherent; the fork points (where right and wrong diverge) are masked by stability signals.

6. Honest uncertainties

What the author is confident about:
• Reasoning basins are a genuine empirical phenomenon, with clear clustering visualizations and quantitative validation.
• Structure ≠ truth is the paper's most broadly applicable finding—a systematic, experimental rebuttal of the implicit 'coherence as proxy for correctness' assumption.
• Arbiter-Δ's simplicity (parameter-free, same-model, zero external information) makes it a pluggable module for any majority-vote system.

What remains unclear:

• Experiments cover only math reasoning (GSM8K, MATH-500, MMLU-HS-Math) and 4B–8B models; basin formation and correction windows at 70B+ scale are untested.
• 'Answer basins' may not be well-defined for non-normalized-answer tasks (code generation, commonsense reasoning, creative writing).
• No mechanistic explanation of why structure signals are orthogonal to correctness.
• Whether integrating Arbiter into an RL training loop as a reward signal could reduce wrong-majority basin formation at the source is unexplored.

Paper details

| Item | Content | |------|---------| | Title | ARBITER: Reasoning Trajectory Basins and Majority Vote Failures in Test-Time Sampling | | Authors | Meng Cai, Lars Kulik, Farhana Choudhury (University of Melbourne) | | arXiv ID | 2605.26172 | | Category | cs.LG | | Core contributions | (1) Discovery that sampled reasoning trajectories cluster into basins, with majority vote selecting the most stable rather than most accurate basin; (2) systematic proof that structure signals (coherence, hidden-state clusters, graph topology) do not equal correctness signals; (3) Arbiter-Δ, a parameter-free conservative arbitrator; (4) demonstration that reliable post-consensus recovery must be sparse, high-precision, and additive | | Scale | 3 models × 3 benchmarks = 9 model-dataset cells | | Limitations | Math-only, small-to-mid models; inapplicable to non-normalized-answer tasks; no mechanistic explanation of structure≠truth; RL integration unexplored |

References

1. Cai, Kulik & Choudhury, "ARBITER: Reasoning Trajectory Basins and Majority Vote Failures in Test-Time Sampling", arXiv:2605.26172, 2026. 2. Wang et al., "Self-Consistency Improves Chain of Thought Reasoning in Language Models", ICLR 2023. 3. Huang et al., "Large Language Models Cannot Self-Correct Reasoning Yet", ICLR 2024. 4. Snell et al., "Scaling LLM Test-Time Compute Optimally", ICLR 2025. 5. Vasudev et al., "Accurate Failure Prediction in Agents Does Not Imply Effective Failure Prevention", arXiv:2602.03338, 2026.