SPTA
A teacher adaptation method for knowledge distillation in which the student generates prefixes and the teacher is trained with RLVR to recover correct reasoning from these prefixes, thereby improving the reliability of token-level supervision.
Timeline (last 90 days)
The paper proposes the Student-Prefix Teacher Adaptation (SPTA) method, which lets the teacher learn in advance to provide effective guidance on the paths students may visit.
SPTA achieves better results than standard OPD and task-level RLVR-adapted teachers on math reasoning benchmarks.
SPTA's solution is divided into two stages: stage one collects student prefixes of random lengths; stage two uses RLVR to train the teacher to adapt to these prefixes.
SPTA does not change the distillation objective itself, and only improves the reliability of the supervision signal through teacher adaptation.
In stage two of SPTA, the teacher is trained to become an expert at recovering from any student-induced prefix state (including erroneous states) to the correct reasoning path.
SPTA's two-stage pipeline is relatively lightweight, requiring only one additional round of prefix collection and RLVR training before OPD, and is engineering-feasible.
The paper has not yet discussed more complex open-domain tasks or strategies for designing the prefix length distribution.
All Facts (7)
The paper proposes the Student-Prefix Teacher Adaptation (SPTA) method, which lets the teacher learn in advance to provide effective guidance on the paths students may visit.
50%UnverifiedSPTA achieves better results than standard OPD and task-level RLVR-adapted teachers on math reasoning benchmarks.
50%UnverifiedSPTA's solution is divided into two stages: stage one collects student prefixes of random lengths; stage two uses RLVR to train the teacher to adapt to these prefixes.
50%UnverifiedSPTA does not change the distillation objective itself, and only improves the reliability of the supervision signal through teacher adaptation.
50%UnverifiedIn stage two of SPTA, the teacher is trained to become an expert at recovering from any student-induced prefix state (including erroneous states) to the correct reasoning path.
50%UnverifiedSPTA's two-stage pipeline is relatively lightweight, requiring only one additional round of prefix collection and RLVR training before OPD, and is engineering-feasible.
50%UnverifiedThe paper has not yet discussed more complex open-domain tasks or strategies for designing the prefix length distribution.
50%