Asking a small AI model the same hard science question 64 times and picking the most common answer often makes it wrong more often, not less.
A new arXiv paper tested self-consistency - the common trick of sampling multiple answers and taking a majority vote - on GPQA Diamond, a benchmark of hard science questions. For Qwen2.5-7B, majority voting hurt accuracy on 56.6% of the problems it was tested on; for Llama-3-8B, it hurt 65.7%. The researchers then tried the obvious fix: a confidence gate that skips voting when the model's token-level uncertainty looks low. That gate failed too, because the entropy measurement tracks how verbose or hedgy the model's writing is, not how sure it actually is about the answer - on Qwen2.5-7B-Instruct-Turbo, samples that contradicted their own problem's majority answer still looked confident most of the time, with 75.7% scoring above a 10-nat confidence threshold, a finding that held up when re-tested on a separate, pre-registered set of 69 problems.
That matters for anyone building a system that routes between a cheap single answer and an expensive multi-sample vote based on a confidence score. The signal discriminates fine across different questions but is close to useless for the within-question call a router actually needs to make, and a second natural fix - checking whether the top two answers agree - also failed, with no explanation the authors could pin down. It's a useful check on the assumption that tricks proven on huge frontier models transfer cleanly down to the 7B-8B models many teams actually run.
The paper's honesty extends to its own limits: a planned replication on a second model couldn't be evaluated, and three reasoning-focused open models each failed to evaluate on hosted serverless inference at a small budget, for three separate reasons the paper details - all three are downloadable, so the bottleneck was the metered API, not the models themselves.