Determine the uncensored stopping lengths of GSM8K and MBPP generations

Determine the true stopping lengths of the parent model’s GSM8K and MBPP generations beyond the 256-token generation cap.

Background

The paper uses the parent model’s generated length as a proxy for output length when analyzing retention. GSM8K and MBPP generations reach the 256-token cap in every parent sample, so the observed lengths are censored rather than actual stopping lengths.

The authors did not rerun these tasks with a larger generation limit. Consequently, the true stopping lengths remain unknown, limiting interpretation of the association between output length and retention for these tasks.

References

GSM8K and MBPP hit the 256-token generation cap in every parent sample, so their length is censored at 256: the observed generated length is censored at the cap. We did not rerun these tasks with a higher cap, so the axis is bounded at 256 for them and their true stopping length is unknown.

— Context-Tower Conversion Preserves Generation While Freezing Retains Knowledge: Low-Budget AR-to-Diffusion Conversion of MoE LLMs  (2610.02657 - Lu et al., 2 Oct 2026) in Appendix, Section “Output length and retention” (Section \ref{app:length})