EnvSched schedules agent RL rollouts by environment and task
Original title: EnvSched: Environment-Aware Scheduling for Rollout-Efficient Agent Reinforcement Learning
On Qwen3-8B it scores 48.81 after 40 steps, above random selection's best score after 120 steps.
EnvSched is an online scheduler for agent reinforcement learning under finite rollout budgets. It models environment-task pairs on a dual-relation graph and uses rollout history to favor high predicted reward variance and underobserved environments. Quotas and a similarity penalty limit concentration. The authors report better scores on agent benchmarks and gains across model sizes at the same training budget, without extra pilot rollouts.
Unconfirmed: Scores are the authors' reported training results.
- Benchmarks are BFCL-v3, tau-Bench, and ACEBench-Agent.
- Within-environment edges capture tools, state, and verifiers.
- Cross-environment edges capture execution and reward topology.
From the original article
1 min read at arxiv.orgAs executable environments for language agents scale, generating and verifying their rollouts becomes increasingly expensive.
By Sihan Yang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.