Those who use unlicensed data for their own training should not be able to prohibit others from distilling their own models—a regulatory gap with competitive…
Direct-OPD transfers RL-induced policy shifts from weaker to stronger models by leveraging the implicit reward signal from the log-ratio of the RL-shifted and original policy.
Anthropic accuses Alibaba of using Claude outputs to train its own models and asks the US government for support against such terms-of-service violations.
Anthropic accuses Alibaba of systematically copying Claude through distillation and calls on the US government to impose stricter regulation of Chinese AI companies and export…
Hidden-state alignment reduces sampling variance, closes the student-teacher gap more effectively, and trains with less memory and computational time than output-only distillation.
