Valdi: Value Diffusion World Models

Valdi doesn't just speed up diffusion-based world models; it exposes why that speed comes at a cost. The paper reveals that single-step inference for planning requires sacrificing the multimodal future representations that make diffusion attractive in the first place, suggesting the field may have been chasing the wrong architectural path for control.

This connects directly to the Generative Model Proposal based Particle Filtering work from the same day, which also tackles learned dynamics in high dimensions but takes the opposite approach: it keeps the full probabilistic machinery of Bayesian filtering intact rather than collapsing it for speed. Where Valdi trades expressiveness for real-time planning, FPPF trades computational efficiency for statistical rigor. Together they frame an emerging choice in robotics: do you want fast deterministic control or slow probabilistic state estimation? The tension Valdi exposes (multimodality vs. performance) is precisely what FPPF tries to preserve through conditional generative proposals.

If Valdi's single-step inference approach outperforms FPPF-style methods on the same continuous control benchmarks over the next 6 months, that suggests practitioners will accept deterministic approximations. If instead papers start citing Valdi's multimodal trade-off as a reason to return to full Bayesian approaches, the field is signaling that uncertainty representation in learned dynamics requires keeping the probabilistic loop intact, even if it costs latency.