s/balajisRECURSIVE AI•Nov 9
10
votes
676
seen
Terminal-RSI raised a 27B terminal agent’s success rate from 20.68% to 50.58%
A terminal-task agent more than doubled its average success rate after two rounds of generating and training on its own verified work. Across four benchmarks, performance rose from 20.68% to 50.58%, showing how much leverage comes from turning successful runs into new training examples instead of relying only on manually written data.
Terminal-RSI uses a Qwen-based 27-billion-parameter agent: it keeps verified terminal runs, rewrites them into reproducible demonstrations, trains on them, then applies reinforcement learning before starting the next cycle. On Terminal-Bench 4.0, the reported score went from 0% to 19.84%.
Terminal-RSI uses a Qwen-based 27-billion-parameter agent: it keeps verified terminal runs, rewrites them into reproducible demonstrations, trains on them, then applies reinforcement learning before starting the next cycle. On Terminal-Bench 4.0, the reported score went from 0% to 19.84%.
Timeline2
Nov 8
Researchers reported Terminal-RSI results for a recursively self-improving terminal-task agent.
Nov 8
A report described the two-cycle training process and benchmark gains, including Terminal-Bench 4.0 results.
1 comment
Nov 9