InfluenceField: A Differentiable Field with Interventionally Identifiable Causal Structure for Multimodal World Modeling

arXiv 2026 34.6 method, theory

TLDR

Introduces InfluenceField, an intervention-aware differentiable latent field for multimodal world models that predicts local intervention effects and recovers causal structure, improving CausalVQA accuracy by 13.1 points.

Reasoning

The paper provides strong theoretical identifiability guarantees and demonstrates substantial accuracy gains on CausalVQA with structural controls. However, the abstract omits detailed ablation results and broader real-world deployment evidence, limiting assessment of generalizability.

Read-first score

Read-first score 34.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 35.

Recency 5%

100

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

Methodology quality 15%

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

Topical relevance 25%

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 15%

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

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

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

Citation velocity 10%

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: 339.

Keyword Scores

world model

interactive world model

world dynamics prediction

world simulator

generative world model

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduces InfluenceField, an intervention-aware latent field placed between the visual encoder and language decoder that lifts patch features into a continuous spatial representation, propagates directed influence over multiple steps, and predicts local intervention effects via a shared transition operator.
Defines a joint training objective combining language modeling, cross-environment invariance, counterfactual rollout supervision, and structural regularization.
Provides theoretical identifiability results for a nonlinear finite-basis population model: target-aligned interventional supervision plus a one-step separation condition restricts representations to within-location reparameterizations and recovers the directed dependency graph exactly.
Provides a linear specialization with an exact partial-coverage characterization and finite-loss stability bound, plus field analysis of spatial coefficient intervention profiles and a shared-channel calibration result.
Demonstrates large empirical gains on CausalVQA, with capacity-matched baselines and structural controls attributing improvements to causal objectives rather than added capacity.

Methodology

InfluenceField is inserted between the visual encoder and language decoder, lifting patch features into a continuous spatial latent field where a shared transition operator propagates directed influence over multiple steps and predicts local intervention effects. Training jointly optimizes language modeling, cross-environment invariance, counterfactual rollout supervision, and structural regularization. The approach is evaluated on CausalVQA against the backbone, capacity-matched baselines, and structural controls, while theoretical properties are analyzed in nonlinear finite-basis population and linear settings.

Key Results

On CausalVQA, InfluenceField improves overall accuracy over its backbone by 13.1 percentage points, with the largest gains on planning and hypothetical categories. Capacity-matched baselines and structural controls attribute robustness and factual-counterfactual consistency gains to the causal objectives rather than added capacity; identifiability analysis shows exact directed dependency graph recovery under the stated conditions.

Resources

Paper