When five AI answers are worse than one

Can extra answers make a model less reliable?

A common way to improve a language model’s mathematical answers is to ask it for several step-by-step solutions, then select the most frequent final answer. Called self-consistency, this approach spends extra computation in the hope that correct answers will recur while mistakes vary.

A study by Wanrong Yang and colleagues tests where that bargain breaks down. Their arXiv preprint, which reports acceptance at ICONIP 2026, examines two small model families. The central finding is not that repeated sampling always fails: it usually improved average accuracy in these experiments. Rather, those averages concealed problems on which voting replaced a correct initial answer with a wrong one.

One deterministic answer versus five sampled answers

The researchers tested Qwen-2.5-0.5B-Instruct and Gemma-3-1B-it, with approximately half a billion and one billion parameters. Each ran in 16-bit precision and a four-bit, memory-saving format using Apple’s MLX framework.

They used the first 200 problems from each of two mathematical benchmarks: GSM8K, containing school-level word problems, and MATH-500, containing more advanced problems. That produced eight model–precision–task combinations. Combining the two tasks within each model–precision setting supplied four additional analyses, not independent new experiments.

For each problem, the team collected a deterministic “greedy” response, which chooses the highest-probability next token, and five randomly sampled responses. Self-consistency selected the most frequent extracted final answer among those five. This is a plurality vote: the winner need not have three votes, or even a majority.

The gains conceal some reversals

Voting improved accuracy over greedy answering in seven of the eight task-specific comparisons, by 1.5–7 percentage points. But in every comparison it also damaged some answers: 3–10.5% of all tested prompts went from correct under greedy answering to incorrect under voting.

The clearest negative result involved four-bit Qwen on MATH-500. Voting corrected 4% of prompts but spoiled 7%, reducing overall accuracy from 12.5% to 9.5%. That is a three-percentage-point decline, not evidence that all small models become worse when sampled repeatedly.

The sampled answers help explain the vulnerability. Across the four pooled settings, 80–90% of harmful reversals involved a winning answer supported by no more than two of the five samples. In 42–67% of harmful reversals, a correct answer was present among the samples but lost the vote. These observations implicate answer selection as well as answer generation; they do not establish what would happen with more than five samples.

An ideal selector is not a deployable system

The paper first considers an oracle: a hypothetical selector that knows whether greedy answering or voting will be correct for each problem. It invokes voting only when voting rescues an incorrect greedy answer.

By construction, this selector preserves every correct greedy answer while adding voting’s successful rescues. Its accuracy advantage over always voting therefore equals the harmful-reversal rate. Across the reported analyses, it achieved 109–174% of always-voting accuracy—not 109–174% absolute accuracy. This is an upper bound on what selective answering could achieve, not a demonstrated forecasting ability.

The practical alternative uses information available after the greedy response: token-probability statistics, response length and simple prompt characteristics. The researchers compared a confidence-only rule, logistic regression and a small neural network. Logistic regression supplied the selected operating point in three of four pooled settings; the neural network won none.

At those operating points, the selected gates skipped roughly 45–99% of voting calls. Their estimated accuracy was 98.1–98.8% of always-voting accuracy. That means retaining nearly all of a relatively low baseline, not answering nearly all problems correctly.

What the savings do—and do not—show

The reported 36–79% computation reduction counts generation passes, not measured latency, electricity or battery life. Its accounting also treats the greedy response as reusable within the five-answer budget, although measured voting accuracy came from five fresh sampled responses. Charging for those five samples plus the initial greedy pass reduces the estimated savings to 25–79%.

Accuracy preservation is less secure than the headline point estimates suggest. The reported confidence intervals for retained accuracy extend down to roughly 76–86%. Although classifier predictions were generated with cross-validation, the decision threshold was selected and evaluated on the same pooled predictions, introducing optimism.

Only one generation seed was used per setting. Bootstrap intervals resampled problems while keeping generated answers fixed, so they do not capture variation between fresh sampling runs. Larger models, reasoning-specialised models, non-mathematical tasks and larger sample budgets remain untested, as do direct comparisons with adaptive sampling methods.

The useful lesson is narrower than “thinking more hurts”: measure which correct answers an aggregation method destroys, alongside the mistakes it repairs, before assuming extra computation buys reliability.