Emotion-Conditioned Short-Horizon Human Pose Forecasting with a Lightweight Predictive World Model

arXiv 2026 47.1 method, application

TLDR

A lightweight autoregressive world model uses facial emotion embeddings to improve short-term human pose forecasting via gated fusion.

Reasoning

The paper introduces a novel approach by conditioning pose prediction on emotion signals, with a lightweight LSTM-based model and counterfactual analysis. However, it relies on small-scale datasets and lacks comparisons to strong baselines, limiting generalizability.

Read-first score

Read-first score 47.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 17.

Recency 6%

100

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

Methodology quality 18%

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

Citation impact 18%

64.9

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.6491417

Reproducibility 18%

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

Topical relevance 29%

24.3

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

Citation velocity 12%

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

Field roles

FrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 363.

Keyword Scores

world model

world dynamics prediction

generative world model

video world model

interactive world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Introducing facial expression-derived emotion embeddings as auxiliary conditional signals for short-term human pose prediction, moving beyond geometric motion cues alone.
Proposing a lightweight autoregressive predictive world model with a learnable gating mechanism for multimodal fusion of pose keypoints and emotion embeddings.
Using counterfactual perturbation experiments to demonstrate that predicted trajectories are measurably sensitive to changes in multimodal input, confirming emotion embeddings act as conditional signals rather than redundant features.

Methodology

The paper proposes a lightweight autoregressive predictive world model based on a two-layer LSTM architecture. It combines pose keypoints with facial expression-derived emotion embeddings through a learnable gating mechanism, performing 15-step rolling pose prediction in an autoregressive unfolding manner. The model is evaluated on two small-scale pose-emotion video datasets: one with controlled motion sequences and minimal facial expression changes, and another with natural emotion-driven motion sequences and considerable facial expression changes.

Key Results

Normalized gating fusion significantly enhances prediction accuracy on emotion-driven motion sequences, while simple multimodal fusion does not consistently improve performance. Counterfactual perturbation experiments show that the predicted trajectory exhibits measurable sensitivity to changes in multimodal input, indicating that facial expression embeddings serve as auxiliary conditional signals.

Limitations

The experiments are conducted on small-scale pose-emotion video datasets, which may limit generalizability.
The model is limited to short-horizon (15-step) prediction, and its performance on longer horizons is not evaluated.
Simple multimodal fusion without gating does not consistently improve prediction accuracy, indicating that naive integration of emotion embeddings is not always beneficial.
The study only considers facial expression-derived emotion embeddings; other modalities or emotion sources are not explored.

Resources

Paper