Test-time scaling (TTS) improves large language model reasoning by sampling multiple answers and selecting the best one. But conventional TTS uses a fixed sampling budget, which ignores the fact that some queries are easy and need few samples, while others are hard and need many. It also overlooks the changing energy cost of inference when GPUs are under contention.
The paper introduces ATEST, an adaptive resource-aware test-time scaling method. ATEST dynamically adjusts the number of samples based on query difficulty and current GPU load. This is designed to allocate compute where it is needed most and reduce energy consumption during periods of high contention.
The provided abstract is truncated, so the full evaluation details are not available. The authors state they "measure..." but the specific metrics and results are cut off. Still, the core idea is clear: making test-time scaling adaptive to both query and system conditions.