Unverified50% confidenceFactExact time
A unified video-action joint model does not automatically 'emerge' compositional generalization capability; the team observed two problems: first, vision can generalize but execution fails—the video prediction model can generate reasonable future frames when faced with unseen instructions, but after connecting the Action Head to drive the robotic arm for actual execution, the policy's success rate drops sharply on compositional out-of-distribution tasks; second, extremely high hyperparameter sensitivity.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
Unverified能力不够的模型会因为要处理Harness的优化任务而分散本身的交互能力，导致整体表现下降70% similarUnverified多智能体强化学习领域的非平稳性问题在多人世界模型中被放大，因为其他玩家的行为分布会随策略动态改变69% similarVerifiedCharacter consistency is a core challenge in AI video/comic generation, because diffusion models inherently lack the ability to remember the appearance of the same character, as each generation resamples from noise.69% similarUnverifiedCarefully edited robot demonstration videos often showcase cherry-picked results and cannot represent the true success rate and stability of a model across a large number of random tasks.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1024137API
curl https://kongchang.com/api/v1/knowledge/claims/1024137MCP
get_claim(id=1024137)