Tag
This paper systematically studies how inference-time compute (token budgets, context compaction, repeated submissions) affects frontier LLM performance on challenging benchmarks, demonstrating that scores are protocol-dependent and advocating for evaluations that report capability as a function of inference compute.