English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Which Way Did It Move? Diagnosing Directional Motion Blindness in Video-LLMs

Forum topic · 小凯 · 2026-05-23

Summary

A 2025 arXiv paper (2505.17389) by Jongseo Lee, Hyuntak Lee, and Sunghun Kim reveals that video large language models (Video-LLMs) largely fail at a basic perception task: recognizing signed image-plane motion direction. In simple videos of a single object moving left, right, up, or down, most models perform near chance, with above-chance results largely explained by prediction bias. The authors call this failure 'directional motion blindness.' Probing shows motion direction is linearly decodable in the vision encoder, projector, and LLM hidden states, but readout mechanisms fail to bind this signal to the correct linguistic answer — a 'direction binding gap.' They introduce the MoDirect dataset family and DeltaDirect, a diagnostic-driven projector-level objective predicting normalized 2D motion vectors from adjacent-frame feature differences. Fine-tuning with DeltaDirect lifts motion direction accuracy from 25.9% to 85.4% on MoDirect-SynBench and improves real-world accuracy by 21.9 points on MoDirect-RealBench without real-world fine-tuning data, while preserving standard video understanding performance.

Paper Overview

  • Field: Computer Vision (CV)
  • Authors: Jongseo Lee, Hyuntak Lee, Sunghun Kim
  • Published: 2025-05-23
  • arXiv: 2505.17389
  • Summary

    Video large language models (Video-LLMs) have advanced rapidly in temporal video understanding, yet many fail at a fundamental perceptual primitive: signed image-plane motion direction. In simple videos where a single object moves left, right, up, or down, most Video-LLMs perform near chance level, and the few cases exceeding chance are largely attributable to prediction bias rather than genuine direction understanding. The authors term this failure directional motion blindness.

    By tracing motion direction information through the Video-LLM pipeline, the paper finds that motion direction is linearly decodable in the vision encoder, the projector, and the LLM hidden states — but the readout mechanism fails to bind this signal to the correct linguistic answer option, exposing a direction binding gap.

    While synthetic motion-direction instruction tuning reduces this gap on the source domain, motion direction concept vector analysis shows that visual complexity attenuates signal magnitude, limiting cross-domain generalization.

    Contributions

  • MoDirect dataset family: for motion-direction instruction tuning and evaluation.
  • DeltaDirect: a diagnostic-driven, projector-level objective that predicts a normalized 2D motion vector from feature differences of adjacent frames.
  • Results

  • On MoDirect-SynBench, instruction tuning with DeltaDirect improves motion direction accuracy from 25.9% to 85.4%.
  • On MoDirect-RealBench, DeltaDirect improves real-world motion direction accuracy by 21.9 percentage points without any real-world fine-tuning data, while maintaining standard video understanding performance.
---

*Auto-collected on 2026-05-23*

Tags

#video-llm#computer-vision#motion-perception#instruction-tuning#arxiv#deep-learning#model-diagnostics

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620654