Choose a skill-bank perturbation distribution without sacrificing peak performance

Determine how to choose and sample a distribution of skill-bank perturbations for training the policy without sacrificing peak return on the delivered skill bank.

Background

The paper models the training objective as optimizing policy performance on a single delivered skill bank, represented by a point-mass distribution. Equation (6) shows that averaging performance over a distribution of perturbed banks would expose average degradation that the point-mass objective ignores. The authors therefore identify the selection and sampling of such a perturbation distribution, while preserving peak performance, as unresolved.

References

How to choose and sample from $P$ without sacrificing peak return remains open.

— Evolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving Agents  (2610.20845 - Feng et al., 14 Oct 2026) in Section 2.3, following Proposition 2

The natural direction is the distributional objective of Eq. (6)—training against a distribution of bank perturbations rather than a single fixed bank—but turning it into a practical method, and keeping the net-benefit gate satisfied, is future work.

— Evolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving Agents  (2610.20845 - Feng et al., 14 Oct 2026) in Appendix A, Limitations

Whether broader exposure—more benchmarks, larger banks, more data—reduces sensitivity to wording or merely raises the peak from which the worst case falls away is beyond our design, and is the most interesting open question beyond the fix itself.

— Evolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving Agents  (2610.20845 - Feng et al., 14 Oct 2026) in Appendix A, Limitations