Designing XM-specific guidance methods

Develop inference-time guidance techniques tailored specifically to Explorative Modeling (XM) generative models that explicitly leverage the training-time signal of K candidate generations and the identification of the best candidate for each example, rather than reusing guidance methods such as classifier-free guidance or autoguidance designed for non-explorative models.

Background

The paper observes that standard guidance methods used in diffusion models, such as classifier-free guidance and autoguidance, do not transfer uniformly to Explorative Models (XMs). Despite stronger unguided performance with XMs, these guidance techniques yielded smaller gains than expected.

Because XM training produces a distinctive signal—multiple candidates per step and explicit knowledge of the winning (closest) candidate—the authors argue that new, XM-tailored guidance could exploit this signal to improve inference-time generation, making guidance design for XMs an important unresolved direction.

References

Guidance is known to not transfer uniformly across models—for example, vanilla CFG also fails to improve models trained on representation autoencoder latents by default—suggesting even our autoguidance results likely undershoot what XMs could achieve with guidance designed for exploration. Exploration also supplies a signal base models lack, namely K candidates and a notion of which was best, so we believe using it to design guidance tailored specifically to XMs is one of the most important open problems.

— Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation  (2607.27372 - Gladstone et al., 29 Jul 2026) in Section 7: Limitations and Conclusion