HERA framework enhances physical prediction in latent world models

Researchers have developed HERA, a framework designed to improve physical prediction in latent world models by selectively routing historical evidence. This approach uses a lightweight adapter called Register-Routed Patch Memory (RRPM), which includes a Structured Memory Bank, Memory Registers, and Workspace Registers. When tested with V-JEPA 2-G on the IntPhys2 Main split, HERA demonstrated significant improvements in accuracy, particularly in scenarios involving fixed-camera continuity and immutability, suggesting historical evidence routing is an effective adaptation strategy for these models. AI

IMPACT Improves physical prediction in latent world models by enabling selective retrieval of historical evidence.

RANK_REASON The cluster contains a research paper detailing a new framework and adapter for latent world models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• HERA
• IntPhys2 Main
• Memory Registers
• Register-Routed Patch Memory
• Structured Memory Bank
• V-JEPA 2-G
• Workspace Registers

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →