ATI-OPD
A teacher intervention method for long-sequence knowledge distillation that locates and replaces early divergence segments between student and teacher and excludes ineffective suffixes under persistent divergence, thereby stabilizing training and improving the student's independent reasoning ability.
Timeline (last 90 days)
A paper from arXiv proposes Adaptive Teacher Intervention for On-Policy Distillation (ATI-OPD), which improves teacher guidance by correcting the student's reasoning prefix.
ATI-OPD works in two steps: first, it locates the earliest segment where the student and teacher diverge and replaces it with brief teacher reasoning, letting the student continue reasoning autonomously from the corrected text; second, if the student diverges again, it excludes the persistently deviating suffix from the distillation loss while retaining the student's subsequent useful content.
ATI-OPD's dynamic intervention is fundamentally different from static truncation: it preserves the continuity of long-sequence reasoning while avoiding the contamination of subsequent learning by erroneous prefixes.
The paper's training diagnostics show that ATI-OPD brings three observable improvements: fewer divergences between teacher and student, a decrease in the number of gradient spikes, and a reduction in the number of extremely long responses.
By correcting deviations at the prefix level and excluding persistently diverging suffixes from the loss, ATI-OPD reduces the conditions that produce anomalous gradients themselves, so the decrease in gradient spikes can serve as direct evidence of improved training stability.
The paper evaluates ATI-OPD on five mathematical reasoning benchmarks and two teacher-student model pairs, and the results show it improves the student's average accuracy in standalone reasoning compared with standard OPD, with gains covering multiple benchmarks and model pairs.
ATI-OPD has direct reference value for distillation tasks requiring long-chain reasoning (such as mathematics, code generation, and multi-step planning).
All Facts (7)
The paper's training diagnostics show that ATI-OPD brings three observable improvements: fewer divergences between teacher and student, a decrease in the number of gradient spikes, and a reduction in the number of extremely long responses.
50%UnverifiedATI-OPD works in two steps: first, it locates the earliest segment where the student and teacher diverge and replaces it with brief teacher reasoning, letting the student continue reasoning autonomously from the corrected text; second, if the student diverges again, it excludes the persistently deviating suffix from the distillation loss while retaining the student's subsequent useful content.
50%UnverifiedATI-OPD's dynamic intervention is fundamentally different from static truncation: it preserves the continuity of long-sequence reasoning while avoiding the contamination of subsequent learning by erroneous prefixes.
50%UnverifiedATI-OPD has direct reference value for distillation tasks requiring long-chain reasoning (such as mathematics, code generation, and multi-step planning).
50%UnverifiedBy correcting deviations at the prefix level and excluding persistently diverging suffixes from the loss, ATI-OPD reduces the conditions that produce anomalous gradients themselves, so the decrease in gradient spikes can serve as direct evidence of improved training stability.
50%UnverifiedThe paper evaluates ATI-OPD on five mathematical reasoning benchmarks and two teacher-student model pairs, and the results show it improves the student's average accuracy in standalone reasoning compared with standard OPD, with gains covering multiple benchmarks and model pairs.
50%UnverifiedA paper from arXiv proposes Adaptive Teacher Intervention for On-Policy Distillation (ATI-OPD), which improves teacher guidance by correcting the student's reasoning prefix.
50%