Topic
Two-Timescale Offline Recursive Self-Improvement (TORSI)
Latest news
- 16 OctPaper introduces RealtimeBench and TORSI to evaluate agents under a live clock
Researchers introduce RealtimeBench, a 180-task benchmark that measures LLM agents while environments keep running during inference, and show latency can make otherwise correct actions fail. They propose TORSI, a two-timescale teacher–student method that boosts a slower agent's instruction set and raises a fast student agent's scores in live settings.
Research · 1 source
Articles
No articles about Two-Timescale Offline Recursive Self-Improvement (TORSI) yet.