TL;DR
WeEnv manages agentic RL environments across packaging, initialization, and resource provisioning. Composable layers and on-demand fetching reduce environment startup costs, while elastic CPU and memory quotas adapt to task demand.
Why it matters
Scaling rollout generation does not remove the cost of preparing and running tool environments. WeEnv studies this environment overhead across the full lifecycle and separates startup gains from measured RL iteration gains.
Key findings
- 01
WeEnv publishes task, harness, and evaluator components as independent layer groups and composes them at launch. Tasks sharing a repository reuse its artifact, and changing a shared component requires republishing only that group.
- 02
The main evaluation uses Slime with colocated Megatron and SGLang on four eight-GPU H20 servers, plus two dedicated environment servers. Median initialization is 10.6 seconds, versus 59.7 for AgentENV, 111.0 for Docker, and 150.6 for E2B in the tested configurations.
- 03
Environment initialization accounts for 28.9–53.4% of iteration time under the baselines and 9.1% with WeEnv. The reported 5.6–14.2× factors describe median initialization, not complete RL training.
- 04
An internal ablation isolates on-demand fetching: median initialization falls from 32.5 to 10.6 seconds. Total time across the evaluated RL iterations falls from 8,092 to 5,988 seconds, a reported 1.35× speedup; the cold first iteration shows a larger 2.4× improvement.
Evidence & limits
AgentENV’s on-demand fetching was disabled because the test servers lacked ublk kernel support. Its measured result does not represent that system with all intended features enabled, although offline image conversion was tuned.
The results cover particular software-engineering workloads, hardware, cache states, and harnesses. Startup speedup factors should not be used as whole-training speedups or model-quality improvements.
Benefits shrink when task execution dominates: the C++ workload shows only a 1.2× median task-time advantage over the best baseline. The paper also evaluates a larger Qwen3-14B setting, but does not establish a universal scaling law.
The report states deployment at WeChat; the published experiments do not quantify organization-wide training cost savings or compare model convergence at a fixed total compute budget.
Scaling dimensions
Models, methods & benchmarks
- Models
- Qwen3-8B, Qwen3-14B
- Algorithms
- Layer composition, On-demand fetching, Elastic resource provisioning
- Benchmarks
- SWE-Smith
Topics