README.md

metadata

title: CoT Spatial Reasoning Degradation
emoji: 🧠
colorFrom: red
colorTo: yellow
sdk: gradio
sdk_version: 4.36.0
app_file: app.py
pinned: false

CoT Spatial Reasoning Degradation

Interactive demo showing that Chain-of-Thought degrades visual spatial reasoning.

Paper: "Chain-of-Thought Degrades Visual Spatial Reasoning Capabilities of Multimodal LLMs" (arXiv:2604.16060)

Hypothesis

CoT prompting causes shortcut learning from textual priors, degrading spatial reasoning performance.

Key Findings

• CoT degrades performance on 13 spatial benchmarks
• Models hallucinate visual details from text alone
• Need vision-centric reasoning paradigms

Features

• Spatial grid puzzles
• Mental rotation tasks
• Pattern completion tests
• Compare CoT vs No-CoT performance