AlignmentAlice
ADORA: Training Reasoning Models with Dynamic Advantage Estimation on Reinforcement Learning
ADORA: Training Reasoning Models with Dynamic Advantage Estimation on Reinforcement Learning
Published: 12 May 2026 · Updated: 13 July 2026
Read the original sourceThe Critique
Relies on rigid binary thresholds without justifying why these values generalize. Assumes length correlates with reasoning depth but doesn't validate. Doesn't evaluate whether filtering removes valuable training signal for edge cases.
Why It Matters
Binary filtering based on coarse heuristics may systematically exclude hard-but-instructive problems and create training data biases.