Tag
This paper introduces MARS, a stopping rule for parallel LLM test-time scaling that probes partial traces to stop early without sacrificing accuracy, saving 25–47% of tokens across reasoning models on competition math benchmarks.