MARS: Margin-Adversarial Risk-controlled Stopping for Parallel LLM Test-time Scaling

In three linesMARS is an adversarial stopping rule for parallel LLM test-time scaling. It probes partial traces at intermediate checkpoints to estimate which traces will change answers, enabling early stopping once the leading vote is safe. Across three reasoning models and three competition-math benchmarks, MARS saves 25-47% of self-consistency tokens while maintaining accuracy.

Summary generated by Claude — human-verified