Interpretable Adaptive Sampling for LLM Test-Time Scaling
The paper proposes a fuzzy controller that decides how many answer samples an LLM should generate per prompt.
Instead of using the same test-time compute budget everywhere, it maps prompt complexity and model confidence to a per-query sampling budget. Easier or higher-confidence prompts get fewer samples; harder or less certain prompts get more. The authors report gains over several standard baselines on question-answering and math reasoning tasks while using fewer samples on average. ArXiv · AI/CL/LG's note
Instead of using the same test-time compute budget everywhere, it maps prompt complexity and model confidence to a per-query sampling budget. Easier or higher-confidence prompts get fewer samples; harder or less certain prompts get more. The authors report gains over several standard baselines on question-answering and math reasoning tasks while using fewer samples on average. ArXiv · AI/CL/LG's note
score 5