From the source
From the source
Research proposes rubric-based method for multi-environment agent training
Apple researchers propose RISED, a method for training a single LLM agent across diverse interactive environments.
An LLM judge tags rollouts using a predefined rubric vocabulary, guiding data selection and providing token-level supervision via self-distillation.
RISED achieves the highest mean pass rate across environments and ranks first or second in every individual environment.
From the source
An LLM judge tags each rollout using a predefined rubric vocabulary shared across environments. The resulting profiles guide the selection of data that aligns with the overall behavioural composition of the mixed-environment batch while limiting overlap with already-selected data.
machinelearning.apple.com