Standard pretraining scaling laws treat training data as a single scalar: the number of tokens. A new arXiv paper argues that agentic reinforcement learning inherits this abstraction without justification, even though modern pipelines do not simply collect tokens—they synthesize tasks. The authors suggest that task-level properties, not just token counts, should drive scaling analysis.
The paper frames agentic RL data as a structured quantity, where the way tasks are generated and distributed is a key variable. It explores synthesis strategies as a controllable axis of scaling, implying that how you create tasks can be as influential as how many you create. This moves beyond the usual compute-and-data curves.
Because the abstract is truncated, the full results are not available in the source. Still, the core claim is clear: treating agentic RL data as a flat token count obscures the role of task design. The paper calls for a more nuanced scaling framework tailored to agentic settings.