Establish multi-seed training stability

Establish the stability of MASP’s reported performance across multiple independent training random seeds, rather than only across bootstrap resamples of held-out evaluation sessions.

Background

The paper reports paired bootstrap confidence intervals over 99 hard-user P4G sessions, which quantify uncertainty over held-out evaluation sessions but do not measure variation caused by different training runs. The authors explicitly distinguish this limitation from training stability and leave the latter unresolved.

References

These intervals measure uncertainty over held-out sessions rather than training variance across random seeds; multi-seed training stability remains to be established.

— Stop Training Against Soft Users: Mentalized Adversarial Self-Play for Proactive Dialogue  (2610.21156 - Liu et al., 15 Oct 2026) in Appendix, Section “Paired Bootstrap Confidence Intervals”