Concept推理时计算优化 / inference-time optimization / 测试时方法
Inference-Time Optimization
A method that improves output quality after model training by increasing computation during the decoding stage, such as generating more candidates, multi-step reasoning, and self-evaluation and selection, without fine-tuning or additional training data.
Timeline (last 90 days)
Oct 21
The method significantly increases inference cost, and practical deployment requires weighing latency against quality.
Unverified50%