LEAF-OPD
A knowledge unlearning method for large models based on privileged context and on-policy distillation, in which a frozen teacher model provides token-level supervision over student-generated trajectories, guiding the student to learn to respond correctly on both forget and retain data.
Timeline (last 90 days)
LEAF-OPD adopts a teacher-student architecture in which a frozen teacher model provides token-level supervision on responses generated by the student itself, following the on-policy distillation approach.
LEAF-OPD gives the teacher model an additional instruction (privileged context) on forget data to guide the student to learn to refuse, while on retain data the teacher does not use privileged context; the student always runs without privileged context.
The paper introduces the Supervision Signal (SS) metric, where SS1 evaluates the supervision strength over the entire trajectory and SS0.5 evaluates the latter half; the larger the gap between the two, the more supervision is concentrated at the beginning and the worse the later persistence.
When supervision is weak, lightweight SFT can reinforce the teacher's context-based refusal behavior, significantly improving both the SS1 and SS0.5 scores.
LEAF-OPD achieved the best results on the TOFU benchmark in Forget Quality, Model Utility and forget ROUGE-L, while maintaining competitive retain performance.
LEAF-OPD's forget quality on the RWKU benchmark is 34% higher than the second-place method, while retain quality remains competitive.
LEAF-OPD was evaluated on the MUSE-News benchmark, but the paper abstract does not give specific numerical values.
Traditional unlearning methods often use gradient ascent or preference optimization to directly suppress the probability of target knowledge; the problem is that they only tear down without building up—the model learns not to say it, but does not learn what it should say.
LEAF-OPD relies on a high-quality teacher model, and the teacher's capability ceiling may constrain the student's unlearning effectiveness.
The paper abstract does not disclose LEAF-OPD's computational cost, sensitivity to teacher model scale, or performance in multi-turn dialogue scenarios.
All Facts (10)
LEAF-OPD adopts a teacher-student architecture in which a frozen teacher model provides token-level supervision on responses generated by the student itself, following the on-policy distillation approach.
50%UnverifiedLEAF-OPD gives the teacher model an additional instruction (privileged context) on forget data to guide the student to learn to refuse, while on retain data the teacher does not use privileged context; the student always runs without privileged context.
50%UnverifiedThe paper introduces the Supervision Signal (SS) metric, where SS1 evaluates the supervision strength over the entire trajectory and SS0.5 evaluates the latter half; the larger the gap between the two, the more supervision is concentrated at the beginning and the worse the later persistence.
50%UnverifiedWhen supervision is weak, lightweight SFT can reinforce the teacher's context-based refusal behavior, significantly improving both the SS1 and SS0.5 scores.
50%UnverifiedLEAF-OPD achieved the best results on the TOFU benchmark in Forget Quality, Model Utility and forget ROUGE-L, while maintaining competitive retain performance.
50%UnverifiedLEAF-OPD's forget quality on the RWKU benchmark is 34% higher than the second-place method, while retain quality remains competitive.
50%UnverifiedLEAF-OPD was evaluated on the MUSE-News benchmark, but the paper abstract does not give specific numerical values.
50%UnverifiedTraditional unlearning methods often use gradient ascent or preference optimization to directly suppress the probability of target knowledge; the problem is that they only tear down without building up—the model learns not to say it, but does not learn what it should say.
50%UnverifiedLEAF-OPD relies on a high-quality teacher model, and the teacher's capability ceiling may constrain the student's unlearning effectiveness.
50%UnverifiedThe paper abstract does not disclose LEAF-OPD's computational cost, sensitivity to teacher model scale, or performance in multi-turn dialogue scenarios.
50%