Kairos: A Native World Model Stack for Physical AI

arXiv 2026 61.5 method, system

TLDR

Kairos is a native world model stack for Physical AI with pre-training, unified architecture, and deployment-aware co-design, achieving top benchmark results.

Reasoning

The paper presents a comprehensive world model stack with novel components like cross-embodiment data curriculum and hybrid temporal attention, supported by theoretical bounds and benchmark experiments. However, the abstract lacks detailed experimental results and comparisons, and the claims about 'top level' are vague without specific metrics.

Read-first score

Read-first score 61.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 46.

Recency 6%

100

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

Citation impact 18%

95.7

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.95709928

Methodology quality 18%

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

Topical relevance 29%

65.7

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

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

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 458.

Keyword Scores

world model

world dynamics prediction

interactive world model

generative world model

video world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Native Pre-training Paradigm with Cross-Embodiment Data Curriculum that organizes heterogeneous experience (open-world videos, human data, robot interactions) into a progressive developmental pathway
Native Unified Architecture with Hybrid Linear Temporal Attention combining sliding-window, dilated sliding-window, and gated linear attention for local, mid-range, and persistent global memory, with formal theoretical bounds on error accumulation
Deployment-Aware System Co-Design enabling low-latency rollout generation on server and consumer-grade hardware for real-world observation-action-feedback loops

Methodology

Kairos employs a Native Pre-training Paradigm using a Cross-Embodiment Data Curriculum that sequences open-world videos, human behavioral data, and robot interactions into a progressive learning pathway. Its Native Unified Architecture features Hybrid Linear Temporal Attention, where sliding-window attention captures local dynamics, dilated sliding windows capture mid-range dependencies, and gated linear attention maintains persistent global memory, with formal theoretical guarantees on state propagation. The system is co-designed for deployment on server and consumer hardware to support low-latency rollout generation. Evaluation is performed on embodied world-model, long-horizon, and action-policy benchmarks.

Key Results

Kairos achieves top-level performance on embodied world-model, long-horizon, and action-policy benchmarks while offering a strong efficiency-capability trade-off, and establishes formal theoretical bounds demonstrating that its temporal factorization strictly limits error accumulation across extended horizons.

Resources

Paper