Thanawat Lodkaew

It is proved that symmetric losses enable successful policy improvement even with noisy labels, as the resulting reward is rank-preserving—a property that is identified as sufficient for policy improvement.

Capped evaluation reveals that many high scores from coding agents are just clever shortcuts, not true problem-solving.

Even with noisy human preferences, symmetric losses can guarantee rank-preserving rewards, unlocking robust policy optimization for aligning language models.