Reproduce “STAR: A Benchmark for Situated Reasoning in Real-World Videos” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, Chuang Gan

STAR introduces a benchmark for situated reasoning in real-world videos, designed to evaluate AI models' ability to abstract dynamic situations and perform logical inferences. The benchmark utilizes human activity videos and procedurally generated questions, revealing that situation abstraction is the primary bottleneck for current systems, with state-of-the-art models achieving only 36.79% accuracy.

Key insights

  • Situated reasoning requires a two-step process: 'situation abstraction' (converting raw visual input into a structured, symbolic representation) and 'logic-grounded question answering' (performing explicit reasoning over that representation).
  • Procedurally generating questions with explicit functional programs for answers enables precise diagnosis of model failures, preventing reliance on superficial cues or statistical biases.
  • Leveraging real-world videos, unlike many diagnostic benchmarks that use synthetic environments, introduces essential complexity, pushing models to grapple with realistic visual ambiguity and context.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.