Video Model Predictions Ignored by Action Head
chris_j_paxton · x · 2026-07-11
The post summarizes a rollout observation:
- For ID tasks, misleading predictions from the video model are sometimes ignored by the action head, allowing the task to be completed successfully anyway.
- For OOD tasks, even if the video model makes correct predictions, the action head might ignore them, ultimately leading to failure.
This indicates that downstream control modules don't always utilize video model predictions consistently, and these discrepancies are amplified on out-of-distribution tasks.
More from Multimodal
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- Full animation short made with Google Gemini tools, 10,564 frames — AI_Andrew · 2026-10-26
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26
- Community checkpoint 'Narrative Drift' brings art-directed visuals to Krea 2 Turbo — MoonbearAIArt · 2026-10-26
- Same prompt, split results: Claude Opus 5.5 vs GPT-6 Astra on a 3D C-section explainer — Sourcecode12 · 2026-10-26
- Slopcore: AI-generated meme films turn low-fi slop aesthetics into a subculture — Yamapama · 2026-10-26