The day in conversation for 2026-11-04

discussion

Transcript

The discussion: Rosa: The 90.8 PDMS from Adaptive-WAM is impressive for planning, but I wonder if relying on that quality-aware depth adjustment really solves the fundamental problem of real-time constraint satisfaction when the environment changes rapidly? Dev: I think it’s less about solving the constraint and more about reducing computational overhead; by terminating inference early, you bypass those costly iterative denoising loops mentioned in earlier work, which is critical if we are pushing deployment onto smaller onboard processors. Have we considered how that termination threshold might interact with the uncertainty estimates derived from a standard planner? Taro: From my perspective on the Large Multimodal Models piece, I see Adaptive-WAM as a strong local optimization tool, but it doesn't address the overarching decision framework challenge—that merging LMMs with deep RL for continuous learning feels like the more significant hurdle because that dictates *what* quality is being sought in the first place. Rosa: True. If LMMs are

More episodes

← Home