A preprint reports that GRIP matched SLERP accuracy with 14.5% fewer tokens and outscored Qwen3-Thinking on average benchmarks.