2026 · PKU-YuanGroup · arXiv 2606.07962

ChronoPhyBench

Do MLLMs Truly Understand the World or Merely Exploit Language Priors?

A comprehensive benchmark that evaluates whether video-based large language models genuinely reason about physical processes — or merely rely on superficial textual shortcuts. ChronoPhyBench exposes the prevalence of shortcut learning and modality bias in MLLMs via next-state prediction and chronological sorting.

10,000+ annotated videos
3 evaluation tasks
Real-world physical processes
Multi-modal cross-modal synthesis
ChronoPhyBench Figure

Paper Abstract

Exposing shortcut learning in multimodal LLMs

While multimodal large language models (MLLMs) have demonstrated remarkable performance on video understanding benchmarks, it remains unclear whether they truly grasp the underlying physical dynamics of the world or simply exploit statistical correlations in their training data.

We identify and systematically expose the prevalence of shortcut learning and modality bias in MLLMs, demonstrating how textual priors can induce hallucinations of physical reasoning capabilities. To address this, we introduce ChronoPhyBench, a unified benchmark that employs next-state prediction and chronological sorting to explicitly evaluate cross-modal synthesis and penalize reliance on a single modality.

We release ChronoPhy, a large-scale dataset of over 10,000 annotated videos, providing the community with a rigorous framework for stress-testing multimodal robustness and advancing the development of Physical AI toward genuine Artificial General Intelligence.

Our benchmark targets the models' grasp of fundamental physical laws — gravity, collision, conservation, friction, deformation — and their ability to reason about causal chains in dynamic, real-world scenes.

ChronoPhyBench Logo

Benchmark Design

Three tasks targeting cross-modal physical reasoning

ChronoPhyBench evaluates Video-LLMs across three complementary task types, each designed to probe a different aspect of physical understanding. Tasks require synthesis of visual dynamics and physical priors — single-modality shortcuts are explicitly penalized.

Multiple-Choice QA

Watch a video and answer physics questions. Includes both standard QA and hallucination-robust QA variants to test reasoning vs. guessing.

Metric: Accuracy

Temporal Frame Selecting

Given a historical video, select the single physically-plausible next frame from multiple semantically similar but physically incorrect candidates.

Metric: Accuracy

Temporal Frame Sorting

Given an initial background and shuffled future-state images, rearrange them into the correct temporal and physical order of evolution.

Metric: Exact Match

Resources

Open-source and publicly available

The ChronoPhy dataset, evaluation code, and baseline results are fully open-sourced. All materials are available on GitHub and HuggingFace for the research community.