IPR-1: Interactive Physical Reasoner

arXiv 25.11 2025 51.6 method, benchmark

TLDR

IPR-1 combines world-model rollouts with a VLM policy and PhysCode action space to improve physical reasoning across 1000+ games, outperforming GPT-5.

Reasoning

The paper introduces a novel integration of world models and VLMs for interactive physical reasoning, supported by a large benchmark and strong empirical results. However, the abstract lacks discussion of limitations and the evaluation is limited to simulated games rather than real-world environments.

Read-first score

Read-first score 51.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.

Recency 6%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 29%

62.9

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=benchmark,result

Citation impact 18%

52.4

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.52442022

Reproducibility 18%

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

Citation velocity 12%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 376.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

world simulator

generative world model

video world model

Deep Analysis

Innovations

IPR (Interactive Physical Reasoner) using world-model rollouts to score and reinforce a VLM's policy
PhysCode: a physics-centric action code aligning semantic intent with dynamics to provide a shared action space for prediction and reasoning
G2U (Game-to-Unseen) benchmark of 1,000+ heterogeneous games with significant visual domain gaps

Methodology

IPR uses world-model rollouts to score and reinforce a VLM's policy, and introduces PhysCode as a shared action space for prediction and reasoning. The model is pretrained on 1,000+ games from the G2U benchmark.

Key Results

IPR performs robustly on levels from primitive intuition to goal-driven reasoning, surpasses GPT-5 overall, and shows improved performance with more training games and interaction steps, including zero-shot transfer to unseen games.

Resources

Paper