RISED: Rubrics for Agentic Multi-Env Self-Distillation

1 reports archived1 independent sourcesupdated 10/6/2026, 12:00:00 AM

Synthesis & Latest Updates

1 Sources Cross-Validated

The new RISED framework replaces scalar rewards with LLM-generated structured rubrics to address data selection and lack of intra-group contrast in multi-environment RL. By using textual feedback to guide online data filtering and policy supervision, it achieves the highest mean pass rates across diverse interactive environments.

LATEST/The new RISED framework replaces scalar rewards with LLM-generated structured rubrics to address data selection and lack of intra-group contrast in multi-environment RL. By using textual feedback to guide online data filtering and policy supervision, it achieves the highest mean pass rates across diverse interactive environments.

TIMELINECoverage timeline

Total 1 reports · Latest first

Apple Machine Learning ResearchT1·Oct 6, 12:00 AM68 pts
RISED: Rubrics for Agentic Multi-Env Self-Distillation
Original: RISED: Rubrics for Agentic Multi-Environment Selection and Self-Distillation
- Addresses limitations of scalar rewards in multi-env RL by introducing rubric-based textual feedback to handle batches with uniform outcomes.
- Repurposes rubrics not just as rewards but to guide online data selection and policy supervision via an LLM judge.
- Experiments show RISED achieves the highest mean pass rate across environments and ranks first or second individually.
View Evidence Original Source