GPT‑Red: Unlocking Self-Improvement for Robustness

34 pointsposted 2 days ago
by alvis

1 Comments

jing09928

2 days ago

Useful direction, but the hard part seems to be measuring novelty after each fix. Are they reporting whether later red-team cases are genuinely distinct, or mostly variants of the same failure mode?