A Red Heart Can Steer Code: DeepMind-Stanford Paper Probes Visual Meaning in VLMs
rohanpaul_ai · x · 2026-11-04
A Google DeepMind and Stanford paper, A Heart Means Recursion, reveals surprising gaps between how vision-language models (VLMs) interpret visual symbols and human conventions.
- Simple shapes like a red heart in a prompt can strongly steer model behavior on unrelated tasks, e.g., pushing models toward recursive solutions on simple coding problems
- The steering is largely implicit: models' direct interpretations and thinking traces diverge substantially from their observed behavior
- Model-to-model visual communication shows a generation–interpretation gap: models send shape primitives they find intuitive, but these cues often fail to steer other models
- In communication games with abstract concepts and stories, models creatively use visuals to convey meaning, though some messages remain non-intuitive to humans
The authors conclude VLMs are capable of rich, creative visual communication, but their production and interpretation of visual meaning diverges from human conventions in unexpected ways.
Related event: A Heart in the Prompt Makes VLMs Write Recursion(2 posts)→
More from Models
- Anthropic faces questions over Claude's rules on AI-generated video game clones — emmanuelvivier · 2026-11-04
- Musk cuts Grok Bot usage limits, citing 'unsustainably expensive' compute costs — emmanuelvivier · 2026-11-04
- Simon Willison replicates GPT-6 Astra weight-interpretation feat with Qwen 3.8 27B in 45 minutes — cephaloform · 2026-11-04
- Qwen, GPT and Gemini Flash all fail the same way on this service — matdryhurst · 2026-11-04
- Sub-1B open OCR models benchmarked: Surya 2.5 tops OmniParseBench at 85.9 — VikParuchuri · 2026-11-04
- Gemini 4 Carbon tested on Arena: near Opus on 3D and web design, still behind — altryne · 2026-11-04