Arxiv paper proposing adaptive allocation of test-time compute with evolving in-context demonstrations — a technique for improving model inference efficiency by dynamically selecting and updating example prompts.
Research
Adaptive Test-Time Compute Allocation with Evolving In-Context Demonstrations
Researchers propose dynamically evolving in-context demonstrations during inference to adaptively allocate test-time compute and improve model efficiency without retraining.
Friday, April 24, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.AIBY sys://pipeline
Tags
research