Mechanism
Token-level negative-feedback objective that penalizes toxic, repetitive, or invalid tokens by contrasting their logit directly against a top- plausible positive token sampled from the model's own distribution at position , avoiding sequence-level unlikelihood collapse.