Self-play Task Generation for the Modal-sdk with Vmax’s Campaign infrastructure

At Vmax, we believe that crafting tasks and environments for agents to learn from—just like chess, Go and Mathematics—is a domain where agents themselves will soon far outperform humans. With this in mind, we are building Campaign, a platform that generates tasks in your domain for a given solver model. As a preview of how Campaign works, we are sharing work we did earlier this summer for Modal, which used an early version of Campaign's self-play task generation algorithms to generate software engineering tasks that target the failure modes of closed model APIs when using the Modal SDK.

The tasks we generated required translating python code to code runnable on the Modal platform. The resulting tasks were complex, requiring agents to use, for example, a Modal Volume to cache model artifacts behind an OpenAI-compatible endpoint, or a GPU class that builds an inference engine's state once at container start and reuses it across requests. For a target solver model of GPT 5.5, our task generation algorithms were able to decrease solver performance as more tasks were generated. To show that the tasks provide learning signal, we train on the tasks generated and find a ~1.8x uplift in mean reward at long contexts for Qwen3.6-35B-A3B on our validation set.

We are opening up our Campaign platform and our task generation algorithms to beta-testers so they can create their own environments and tasks with self-play. Sign up to our waitlist.

GlyphBench: A Playground for Language-Model Reinforcement Learning

GlyphBench renders over 360 game tasks as Unicode grids for language-model RL; training on them beats math and code RL at reasoning transfer.

PROPEL: Breaking the Solver Bottleneck in Task-Generator RL

PROPEL replaces solver trials with an activation probe on a frozen generator's internals, doubling frontier-task generation across math, code, and SWE.

unix-ctf: Procedural Environments for Unix-Competence Reinforcement Learning

We introduce unix-ctf, a procedural generator of capture-the-flag environments for training and evaluating Unix competence in language-model shell agents.

PopuLoRA: Co-Evolving LLM Populations for Reasoning Self-⁠Play

We introduce PopuLoRA, a population-based asymmetric self-play framework for reinforcement learning with verifiable rewards (RLVR) post-training of LLMs.