Source of reduced degradation on general reasoning performance

Determine whether the smaller degradation in MMLU-Pro performance produced by Self-Routing results from its conditional assignment of optimization paths according to the model’s own outputs, thereby reducing the influence of noisy or overly specialized training signals and limiting drift from general-purpose behavior.

Background

The paper reports that mathematical post-training generally reduces MMLU-Pro performance relative to the base model, but Self-Routing exhibits the smallest degradation among the post-training methods evaluated. The authors attribute this observation tentatively to the method’s conditional routing rather than uniform updates.

This explanation is explicitly presented as a conjecture rather than an established causal result. Confirming it would require isolating whether adaptive assignment among GRPO, OPSD, regularization, and skipping actually reduces noisy or overly specialized updates and thereby preserves broader reasoning capabilities.

References

We conjecture that this smaller degradation may come from the conditional nature of Self-Routing: instead of applying a uniform math-oriented update to all training instances, it selects different self-improvement paths according to the model’s own outputs.

— From Rollouts to Recipes: Self-Contained Post-Training for LLMs  (2609.01422 - Li et al., 1 Sep 2026) in Section 4.2, Main Results, p. 5

These are single-replicate operating characteristics on fixed model outputs. They are not measurements of the cross-seed R1 or R2 rules, whose power remains unestablished.

— Routing Probes Can Improve Without New Information: An Exact-Null Audit of Uncertainty Beyond Model Outputs  (2609.38956 - Liang et al., 30 Sep 2026) in Section 3.2, “Removing null detections does not create sensitivity” (Section sec:power); also discussed in Section 3.3 and Appendix B, Section “Independent-split linear power endpoint”