Home▸AI & Machine Learning▸VLM Object Hallucination Lab — Evidence/Prior Phase Plane (2D)

VLM Object Hallucination Lab — Evidence/Prior Phase Plane (2D)

Watch a vision-language model's caption decoder play out in 2D evidence-vs-prior phase space: each candidate object is a dot at (visual evidence, language prior), sized by its live softmax probability of being named next, so you can see the CHAIR hallucination metric emerge directly from the decoding math.

AI & Machine Learning2DAdvanced60 FPS📱 Mobile-adapted⇄ 3D version

Vision-language models like GPT-4V, Gemini Vision, LLaVA and Florence-2 caption images by blending two signals at every generated word: real visual evidence from the image encoder, and a language prior learned from co-occurrence statistics in training captions. Rather than rendering the scene the model is looking at, this 2D build plots the decoder's own decision space directly — an evidence-vs-language-prior phase plane, with a live decoding-score field in the background and every candidate object drawn as a dot sized by its real softmax probability of being the next word. Push the language-prior weight or the occlusion/noise slider up and watch phantom objects (dots on absent items) swell and get named anyway — a hallucinated mention, tracked live by the same CHAIR_i metric researchers use to benchmark real VLMs.

⚙ Under the hood

Watch a vision-language model's caption decoder play out in 2D evidence-vs-prior phase space: each candidate object is a dot at (visual evidence, language prior), sized by its live softmax probability of being named next, so the CHAIR hallucination metric emerges directly from the decoding math instead of a rendered scene.

vision-language modelVLMhallucinationCHAIR metricsoftmax decodingphase space

2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install