Research · Updated 16 Oct, 09:30 am IST
Paper introduces RealtimeBench and TORSI to evaluate agents under a live clock
Why it matters for readers: If you follow AI, this paper explains why being fast matters as much as being smart when the world keeps changing.
- RealtimeBench contains 180 tasks across six game types to evaluate agents while games continue running during inference.1
- Paired evaluations show GPT-6.1-sol outperforms the faster Jev when games are paused, but GPT-6.1-sol falls behind Jev when games keep running.1
- The authors propose Two-Timescale Offline Recursive Self-Improvement (TORSI): a slow teacher studies failures and writes reusable instructions for a fast student that acts during gameplay without teacher calls.1
- Applying two rounds of TORSI raises Jev's average test score from 13.51 to 24.19 on a 0–100 progress scale across the benchmark.1
- The instructions produced by the teacher also improve average scores for three other models without retuning.1
Get a brief like this every morning
Uzha reads hundreds of sources and gives you the stories that matter for your work, with every source linked. Free.
Get started