AutoEnvScaling: Automating the Data Flywheel with Terminal Agents
AutoEnvScaling turns RL environment design itself into a terminal-agent task.
The paper says a proposer agent builds new verifiable environments from the solver’s rollouts, then uses later solver runs to drive the next round. It reports up to a 25x higher valid-environment rate than prompt-only generation, at up to 75% lower cost per valid environment. Training on these generated environments improved Terminal-Bench scores for Qwen models, including a 12.2-point gain when Qwen3.6-35B-A3B played both proposer and solver roles. HF Daily Papers' note
The paper says a proposer agent builds new verifiable environments from the solver’s rollouts, then uses later solver runs to drive the next round. It reports up to a 25x higher valid-environment rate than prompt-only generation, at up to 75% lower cost per valid environment. Training on these generated environments improved Terminal-Bench scores for Qwen models, including a 12.2-point gain when Qwen3.6-35B-A3B played both proposer and solver roles. HF Daily Papers' note
score 6