RELIC: Interactive Video World Model with Long-Horizon Memory

arXiv 25.12 2025 62.2 method, system

TLDR

RELIC is a 14B-parameter interactive video world model with long-horizon memory, enabling real-time, memory-aware exploration from a single image.

Reasoning

Strengths: addresses three key challenges simultaneously, uses novel memory-efficient self-forcing paradigm, achieves real-time 16 FPS. Weaknesses: trained on synthetic Unreal Engine data, not real-world; limited to 5-second training horizon extended via teacher model; no evaluation on real-world benchmarks mentioned.

Read-first score

Read-first score 62.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 52.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%

74.3

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,dataset

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 304.

Keyword Scores

world model

interactive world model

video world model

generative world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Unified framework addressing real-time long-horizon streaming, consistent spatial memory, and precise user control simultaneously
Compact camera-aware memory structure using compressed historical latent tokens with relative actions and absolute camera poses in the KV cache for implicit 3D-consistent content retrieval
Memory-efficient self-forcing paradigm that enables full-context distillation over long-duration teacher and long student self-rollouts, extending generation beyond original 5-second training horizon

Methodology

RELIC is built upon autoregressive video-diffusion distillation techniques. It represents long-horizon memory using highly compressed historical latent tokens encoded with both relative actions and absolute camera poses within the KV cache. A bidirectional teacher video model is fine-tuned to generate sequences beyond its original 5-second horizon, then transformed into a causal student generator using a memory-efficient self-forcing paradigm for full-context distillation. The model is a 14B-parameter transformer trained on a curated Unreal Engine-rendered dataset.

Key Results

RELIC achieves real-time generation at 16 FPS while demonstrating more accurate action following, more stable long-horizon streaming, and more robust spatial-memory retrieval compared with prior work.

Resources

Paper