Excessive LLM Sampling Can Worsen Answer Quality and Waste Compute

Large Language Models (LLMs) often employ test-time scaling, where a question is sampled multiple times to increase the chance of generating a correct answer. While this strategy can improve the *coverage* (the likelihood of at least one correct answer appearing), this research demonstrates that it doesn't necessarily improve the *selection* of the best answer. In fact, excessive sampling can be detrimental. The study introduces the concepts of the "modal ceiling" and "correlation ceiling." The modal ceiling indicates that for selecting a single answer, the consensus (vote) often settles within a few dozen samples. Beyond this point, additional samples primarily increase the model's confidence in a potentially incorrect answer, rather than improving accuracy. The correlation ceiling applies to benchmark scoring, where the effective number of samples is even lower. This phenomenon, termed the "identifiability gap," highlights that LLMs can generate correct answers they cannot reliably pick. The key takeaway is that the bottleneck isn't generation, but recognition. Over-sampling wastes computational resources and can lead to worse outcomes by reinforcing confident mistakes, suggesting that optimal sampling is far less extensive than commonly assumed.