BiWM: Advancing Open-Source Interactive Video World Models with Bidirectional Autoregression

arXiv 2026 57.1 method, system

TLDR

BiWM is the first open-source framework for bidirectional autoregressive interactive video world models, enabling efficient camera control and long rollouts.

Reasoning

The paper presents a novel framework that reduces training stages and improves interactivity over existing causal models, with strong empirical results across multiple backbones. However, the abstract lacks detailed quantitative comparisons and does not address limitations like computational cost or failure cases.

Read-first score

Read-first score 57.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

95.8

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.95813327

Topical relevance 29%

68.6

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 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 509.

Keyword Scores

interactive world model

video world model

world model

generative world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

First full-stack framework for interactive video world models under the bidirectional autoregressive paradigm
Joint optimization of generation quality and inference speed with only two training stages instead of four
Single recipe spanning multiple backbones (Wan2.1-1.3B, Wan2.2-5B, HunyuanVideo-1.5-8B, LTX-2.3-22B)
Real-world camera control where prior causal models (minWM) lose controllability
Pluggable history compression (FramePack-style and PackForcing-style) for long rollouts
Optional NVFP4 4-bit training/inference pipeline
GAN and mass-covering forward-KL objectives to counter DMD's mode-seeking degradation

Methodology

BiWM starts from a pretrained video backbone, injects camera control via fine-tuning, then applies a few-step Distribution Matching Distillation (DMD) stage to transform the backbone into an action/camera-controllable world model. This two-stage pipeline replaces the four-stage causal pipeline of minWM, converging in a few hundred steps on 8xH200 GPUs. The framework also supports secondary fine-tuning of existing bidirectional models and integrates pluggable history compression and an optional 4-bit pipeline.

Key Results

BiWM converges in a few hundred steps on 8xH200 GPUs, enables real-world camera control where minWM loses controllability, and is open-sourced for resource-constrained research and high-fidelity environment simulation.

Limitations

DMD's mode-seeking degradation is addressed with additional objectives but may not be fully eliminated
Requires a pretrained video backbone and fine-tuning stages, limiting applicability to models without such backbones
Performance on very long rollouts or extreme camera controls is not explicitly quantified in the abstract

Resources

Paper