A new training method called EnvACE lets large language model agents practice inside their own heads instead of relying on external environments. The approach, described in a paper submitted to arXiv on 6 Aug 2026, replaces real-world interaction during training with what the authors call world rehearsal. The result is a policy that internalizes how actions change the world, and it performs strongly across four benchmarks.
A New Way to Train Agents
Training LLM agents for long-horizon tool use typically depends on real or synthesized executable environments. Those environments are costly to construct and verify. External simulators offer an alternative, but they are difficult to ground in reality.
EnvACE sidesteps both problems. During training, the policy alternates between acting and rehearsal. It generates a tool call, then plays the role of the environment to produce a response, and conditions its next decision on that rehearsed response. Both roles are optimized together, end to end, using task-success rewards.
This process internalizes the relationship between actions and environment responses directly into the policy parameters. The authors call the result an agent world model. That model supports decision making without needing a live environment at every step.
Strong Results Across Benchmarks
The paper reports that EnvACE achieves strong and transferable performance on BFCL-v4, tau^2-Bench, VitaBench, and FinMCP-Bench. That is four benchmarks in total, spanning tool use, agentic tasks, and financial tool interactions.
According to the authors, EnvACE outperforms environment-scaling baselines in the overall evaluation. Controlled studies also show that world rehearsal consistently improves policy learning across model scales. The gains hold even when the underlying model size changes.
The findings establish world rehearsal as a new path toward scaling LLM agent training beyond the constraints of external environments. That claim comes directly from the paper, which frames the method as a way to break the dependence on costly environment infrastructure.
Private Rehearsal at Test Time
The internalized world model does more than speed up training. At test time, it enables private rehearsal before committed execution. The agent can simulate possible responses internally, weigh them, and only then act.
The paper reports that this yields further gains under a moderate rehearsal budget. Crucially, it requires no additional external interaction. The agent rehearses privately, then commits to a tool call when it is confident.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
That design makes the method practical for deployment. Agents can think before they act, without needing a simulator or a live environment to check their plans.
Paper Details and Authorship
The paper, with arXiv ID 2608.06197, was submitted at 15:54:36 UTC on 6 Aug 2026. Version 1 is the only version listed. The PDF is 6,881 KB in size. A DOI is pending registration at https://doi.org/10.48550/arXiv.2608.06197.
Zishan Xu is the lead author, joined by 11 co-authors for a total of 12. The full author list includes Zhiyuan Yao, Yuxin Chen, Yifu Guo, Zhengxi Lu, Yuquan Lu, Jinyang Huang, Yan Xu, Yasheng Wang, Weinan Zhang, Xingshan Zeng, and Weiwen Liu.
The paper is categorized under Artificial Intelligence, cs.AI. It includes full-text links for PDF, experimental HTML, and TeX source. Code is publicly available, though the paper text does not specify a URL. The submission history lists only the single version from 6 Aug 2026.
Broader Context
The method arrives as LLM agents are increasingly used for long-horizon tool use. That use case demands reliable interaction with complex, changing environments. EnvACE offers a way to train such agents without building or maintaining those environments at every step.
The paper does not speculate on future directions. It sticks to the method, the benchmarks, and the reported results. The authors position world rehearsal as a training paradigm, not a one-off trick.
The arXiv listing also includes standard tools for readers. References and citations link to NASA ADS, Google Scholar, and Semantic Scholar. Bibliographic tools include Bibliographic Explorer, Connected Papers, Litmaps, and scite.ai. Code and data links point to alphaXiv, CatalyzeX Code Finder, DagsHub, GotitPub, Huggingface, and ScienceCast. Demos are available through Replicate, Hugging Face Spaces, and TXYZ.AI.
The listing also notes arXivLabs, a framework for experimental projects with community collaborators, and arXiv's commitment to openness, community, excellence, and user data privacy.