Results

What is added

  1. Gain of each robust method over its base algorithm and over the standard RL pool
  2. Retention across shift severity
  3. Isolated shifts on 50 configurations
  4. Compound shifts on four tasks
  5. Unified comparison under matched training budget and checkpoint rule
  6. Best robust algorithm vs best standard algorithm

All scores are unclipped normalized scores, with task-level reference returns shared by every method and condition.