Loading
Life sciences · Preprint
arXiv · September 9, 2026
Raises a question worth testing. It does not answer one.
CompassOPD is a proposed modification to on-policy distillation for cross-family language model training that decomposes and reweights gradient signals to isolate within-family capability shifts from baseline offsets. The method shows numerical improvements in reasoning accuracy on benchmark tasks, but the work is methodological, not peer reviewed, and does not evaluate clinical or real-world outcomes.
Computational method development with empirical validation across model families. Large language models from three student model families and multiple teacher model families evaluated on reasoning tasks; not applicable to human subjects.. Intervention: CompassOPD: modified on-policy distillation that transfers within-family log-likelihood shifts while anchoring to frozen student reference. Compared with: Standard on-policy distillation (OPD).
CompassOPD improves average reasoning accuracy by up to 5.50 points over standard cross-family OPD MoE teacher variant achieves 3.43-point gain over OPD without requiring a separate reference checkpoint Standard OPD effectiveness degrades in cross-family settings even after tokenizer alignment
Safety was not reported in the material analysed. Check the source before drawing any conclusion about harm.
The source did not state who this applies to in practice.
This is a preprint describing a machine learning method proposal with experimental validation on reasoning tasks, but it addresses a technical algorithmic problem rather than a clinical or health outcome, and has not undergone peer review.
As stated by the source record.
Quoted from the source exactly as published.
Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.
On-policy distillation (OPD) provides dense token-level supervision on student-generated trajectories. Although OPD performs strongly when teacher and student belong to the same model family, we find that its effectiveness degrades in cross-family settings even after tokenizer alignment, with substantially stronger external teachers offering little additional improvement. To understand this disconnect, we decompose the cross-family OPD signal into two components: an offset between a low-capability teacher-family reference and the student, and the within-family log-likelihood shift from that reference to the strong teacher. Standard OPD transfers both components together, allowing the offset to dominate the update direction and obscure the changes associated with teacher capability improvements. We propose CompassOPD, which removes this offset and transfers the within-family shift, while a frozen student reference anchors updates to the student's initial policy. Thus, both teacher-side and student-side changes are measured within their respective model families. Experiments across three student families and multiple teacher families show that CompassOPD consistently outperforms standard cross-family OPD, improving average reasoning accuracy by up to 5.50 points. For an MoE teacher, we further construct the reference directly from the teacher checkpoint by reducing expert activation, eliminating the need for a separate reference checkpoint while retaining a 3.43-point gain over OPD.
Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.