Latent Dynamics Reasoning (LDR) World Model Lab

● Direct Pixel Auto-Regression

Divergence Compounding

Error Drift: 0.00 px

◆ Latent Dynamics Reasoning (LDR)

Conservation Preserved

Latent Drift: 0.00 px

Horizon: 0/60

Phase Space & Energy Conservation Drift (60-Step Rollout) Cyan: LDR | Red: Pixel

Extrapolation Parameters (OOD Test)

Test Initial Velocity 28.0 px/f

Training Velocity Cap (ID Limit) 12.5 px/f

Latent Reasoning Capacity (Latent Dim) 128-d (Structured)

Live Telemetry & Error Metrics

Drift Reduction

78.4%

LDR Conservation

0.96

Pixel Drift Step

Step 14

OOD Velocity Ratio

2.24x

Architectural Insight

Traditional pixel auto-regressive models fit visual appearance in-distribution. When velocity exceeds the training cap (12.5 px/f), direct pixel predictions blur boundaries, create ghost masses, and compound positional drift.

Latent Dynamics Reasoning (LDR) maps past frames to structured latent coordinates, evolving physical invariants through structured state rollouts.