Scaling AI with Test-Time Compute: New Insights Ahead
In a highlight lecture for the Richard M. Karp Distinguished Series, experts discuss the theory of test-time scaling in large language models (LLMs). This talk will introduce novel ideas indicating that increasing both parameters and compute at inference time can boost AI performance, impacting speech recognition and language translation systems. Test-time compute models present new avenues for self-improvement and efficient task solving, perhaps allowing future systems to utilize excess compute resources for more refined outputs.