Microsoft's AutoEnvScaling lets agents build their own RL environments, lifting Qwen3 Terminal-Bench from 41.1 to 53.3

omarsar0 · x · 2026-10-15

Microsoft and collaborators published AutoEnvScaling, which treats RL environment design as a terminal task: a proposer agent watches the solver's rollouts and builds new training environments from them, with the solver's next rollouts guiding the next round.

Key results:

Qwen3.6-35B-A3B jumps from 41.1 to 53.3 on Terminal-Bench 2.1 (+12.2), acting as both proposer and solver for co-evolution.
Building environments in a terminal raises the valid-environment rate up to 25× over prompt-only generation, at up to 75% lower cost per valid environment.
With GPT-5.6-sol or Claude Opus 5 as proposer, Qwen3.5-35B-A3B scores 4.3 and 3.8 points above matched static environment pools; gains shrink when the proposer can't see solver feedback.
The pipeline generalizes across software engineering, computer use, professional work and GPU kernel programming — a path toward recursive self-improvement with less human reliance.

Original post →

Microsoft's AutoEnvScaling lets agents build their own RL environments, lifting Qwen3 Terminal-Bench from 41.1 to 53.3

More from coding & agent