Researchers say they've found a more efficient way to make large language models think harder without spending a fortune on compute.
Self-consistency is a standard trick for improving LLM accuracy on reasoning tasks: generate a bunch of responses to the same question, then pick the most common answer. It works, but it's expensive — running dozens of samples per question across a large dataset adds up fast. A new paper from arXiv introduces Blend-ASC, a variant that dynamically decides how many samples each question actually needs rather than applying a fixed number to everything. The result is a claimed 4.8x reduction in samples used on average, while still matching or beating the accuracy of conventional self-consistency approaches. Crucially, Blend-ASC requires no hyperparameter tuning and supports batching, so it can slot into existing pipelines without much friction.
The efficiency gap matters because test-time compute has become one of the main levers AI labs pull to squeeze more performance out of existing models — OpenAI, Google, and others have all leaned on it as a way to boost benchmark scores without retraining. A principled, theoretically grounded approach to sampling efficiency could meaningfully reduce inference costs, which remain a real constraint for anyone running reasoning workloads at scale.
The paper also offers what the authors call the first formal power-law analysis of how self-consistency scales with sample count — useful context, though published benchmarks have a way of looking less impressive once they meet production traffic.