Orca: The World is in Your Mind

arXiv 2026 47.6 method, benchmark

TLDR

Orca is a general world foundation model learning unified world latent space via Next-State-Prediction from multimodal data, enabling text, image, and action generation.

Reasoning

Strengths include a novel unified next-state prediction paradigm, large-scale pre-training data (125K hours video, 160M events), and strong performance across modalities. Weaknesses are limited detail on conscious learning and evaluation metrics, and unclear generalization beyond seen domains.

Read-first score

Read-first score 47.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 54.

Recency 6%

100

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

Topical relevance 29%

77.1

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=baseline,experiment,result

Reproducibility 18%

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

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 515.

Keyword Scores

world model

world dynamics prediction

video world model

generative world model

interactive world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Unified world latent space learned from multimodal world signals with multimodal readout interfaces
Next-State-Prediction modeling as a unified state-transition approach replacing isolated next-token, next-frame, or next-action prediction
Two complementary learning paradigms: unconscious learning (dense natural state transitions from continuous videos) and conscious learning (sparse meaningful state transitions via language-described events and VQA supervision)
Large-scale world-learning inventory: 125K hours of video data and 160M event annotations for pre-training
Frozen backbone with lightweight modality-specific decoders for downstream tasks (text, image, action)

Methodology

Orca is a general world foundation model that learns a unified world latent space from multimodal signals using Next-State-Prediction. It employs two complementary pre-training paradigms: unconscious learning captures dense state transitions from continuous videos, and conscious learning models sparse meaningful transitions from language-described events and VQA supervision. Pre-training uses 125K hours of video and 160M event annotations; after pre-training, the backbone is frozen and only lightweight modality-specific decoders are trained for downstream tasks (text generation, image prediction, embodied action generation).

Key Results

Orca outperforms similar-sized specialized baselines on three downstream readouts, demonstrating that stronger world latent representations enable stronger downstream performance and that the proposed paradigm scales effectively.

Resources

Paper