arXivAI Research
Overcoming Test-Time Scaling Bottlenecks with Adaptive Looped Transformers: Introducing TaH2
自適應循環 Transformer 突破測試時縮放瓶頸：清華團隊提出 TaH2 架構
Researchers developed TaH2, an adaptive looped Transformer that uses an iteration decider and lookahead depth supervision to dynamically allocate computation per token, unlocking highly efficient test-time scaling.
2 min read