Comment by programjames
Comment by programjames 18 hours ago
> if you purely use policy optimization, RLHF will be biased towards short horizons
> most RL has some adversarial loss (how do you train your preference network?), which makes the loss landscape fractal which SGD smooths incorrectly