Search papers, labs, and topics across Lattice.
5
2
5
4
It is proved that symmetric losses enable successful policy improvement even with noisy labels, as the resulting reward is rank-preserving鈥攁 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.
Centering advantages in policy gradients can drastically reduce variance and improve performance in reinforcement learning tasks.
OrderGrad transforms policy-gradient optimization by enabling precise control over distributional properties, allowing for risk-averse and exploratory learning in real-world applications.
Even with noisy human preferences, symmetric losses can guarantee rank-preserving rewards, unlocking robust policy optimization for aligning language models.
