Training a single LLM agent jointly across diverse interactive environments is an increasingly attractive route to generalist agents. Existing curriculum and data-selection strategies often allocate training at the environment level or prioritize local reward-based signals, so they do not explicitly consider relationships between current rollouts across environments for prompt-group selection. A second challenge appears when environments are learned at different rates: a batch can contain rollout groups that all fail or all succeed, and scalar rewards provide no within-group contrast for those groups. The paper argues that these limitations motivate richer textual feedback, specifically rubrics describing rollout behavior, rather than relying solely on scalar rewards.

RISED defines a rubric vocabulary shared across environments. An LLM judge tags each rollout against that vocabulary, producing rubric profiles. Those profiles have three roles. During online data selection, they guide the choice of data that aligns with the overall behavioral composition of the mixed-environment batch while limiting overlap with already selected data. Available positive rubrics, which describe desired behaviors, provide privileged context for an on-policy self-distillation teacher, adding token-level supervision. Negative rubrics, which describe undesired behaviors, guide subsequent rollout generation away from recurring failure modes. Rubrics are therefore not used only as rewards; they are repurposed for data selection and policy supervision.

Across model backbones, RISED achieves the highest mean pass rate across environments and ranks first or second in every individual environment. Rubric-based analysis of RISED further characterizes the behavioral changes accompanying these gains, connecting the improvement to specific shifts in rollout behavior rather than only aggregate performance.

The contribution targets multi-environment RL for LLM agents, where scalar rewards supply limited information about cross-environment relationships and no within-group reward contrast when rewards are identical. Using rubric profiles as a shared behavioral signal gives the training procedure a way to compare and supervise rollouts across heterogeneous environments.