MEAL: A Benchmark for Continual Multi-Agent Reinforcement Learning

A new benchmark called MEAL aims to push continual multi-agent reinforcement learning beyond the small task sequences that have dominated the field, enabling researchers to train agents on 100 sequential tasks in hours on a single GPU [1][2]. The benchmark, formally named Multi-agent Environments for Adaptive Learning, was introduced in a paper posted to the arXiv preprint server [1][2]. The authors note that despite the field’s motivation toward lifelong learning, most continual reinforcement learning papers consider only three to ten sequential tasks because CPU-bound environments make longer sequences impractical [1][2]. Cooperative multi-agent settings have remained largely unexplored in this context [2]. MEAL addresses both gaps by using JAX and GPU acceleration [2]. The paper’s submission history shows an initial version posted on June 17, 2025, followed by revisions in September 2025 and June 2026 [1]. The first submission was 1,711 KB, the second 5,382 KB, and the third 3,184 KB [1]. The corresponding author is listed as Tristan Tomilin [1]. A key finding reported in the abstract is that long task sequences reveal failure modes that do not appear at smaller scales [1][2]. The benchmark’s design allows a full training run on 100 tasks to complete in a few hours on a single GPU, a scale the authors argue is necessary to surface these previously hidden issues [2]. The paper appears on arXiv, an open-access repository of electronic preprints that has been operating since 1991 and now receives about 24,000 submissions per month [6]. The repository is not peer-reviewed but provides rapid dissemination of research findings [6]. The MEAL paper’s abstract page includes links to several community-developed tools under the arXivLabs framework, such as Bibliographic Explorer and CORE Recommender, which offer citation navigation and open-access paper recommendations [4][5]. arXivLabs is a formalized collaboration framework that allows third-party developers to build experimental features on the platform under guidelines emphasizing openness and user data privacy [4].