rlhffactbullishRL helps models generalize better than SFT, with theory supporting this claimNathan Lambert29 Jul 2026https://x.com/natolambert/status/2082188013162144066