Tencent Boosts Qwen3.6 Shell AI Scores 18 Points With Evolving Training Tasks

Training capable terminal agents has hit a wall that has little to do with the model itself. Once a frontier LLM can already solve most of the shell tasks you throw at it, those tasks stop teaching it anything. Reinforcement learning needs environments that sit right at the edge of what the policy can do, and once that edge moves, the training data has to move with it.

A team from Tencent Hunyuan is proposing a different way to keep that edge alive. Their method, called environment evolution, grows lineages of terminal tasks that get harder generation by generation, without ever needing to roll out the model being trained. Applied to Qwen3.6-27B and Qwen3.6-35B-A3B, it lifts Terminal-Bench 2.1 scores by 14.4 and 18.0 percentage points respectively.

Why synthesized environments go stale

Terminal agents are LLMs that drive a shell, chain command-line tools, and iterate on feedback to finish long-horizon jobs. The community has built pipelines that turn GitHub repos, skills, and webpages into executable sandboxes with an instruction and a verifier. The problem is what happens next: these tasks are often too easy for current frontier models, which consistently solve them across repeated rollouts. When every rollout in a group succeeds, GRPO and its cousins get no signal to learn from, so the environment is discarded.

The obvious fix has been agent-environment co-evolution: roll out the current policy, find where it fails, and synthesize new tasks around those failures. It works for a while, but the authors argue it has a structural ceiling. The resulting tasks are constrained by the rollout model and the initial task distribution, which limits generalization and dries up the learning signal as training saturates and model failures become sparse. You end up chasing a moving target with a tool that only sees where the target used to be.