DeepMind-Stanford paper: a red heart in your prompt flips code from 0% to 100% recursive
rohanpaul_ai · x · 2026-11-04
A Google DeepMind–Stanford paper, "A Heart Means Recursion," shows that irrelevant colored shapes in a prompt can steer VLM coding behavior — often invisibly.
- Method: add a plain shape (e.g. a red heart) to coding prompts across 288 shapes, tracking how often Claude Opus 4.7, Gemini 3.1 Pro, and GPT-5.4 chose recursive solutions
- A red heart moved Claude Opus 4.7 and GPT-5.4 from 0% to 100% recursive solutions on a binary tree problem; when Gemini 3.1 Pro was steered, 71.5% of its thinking traces called the image irrelevant
- The paper also finds a generation–interpretation gap: models send shape primitives they find intuitive, but these cues often fail to steer other models; models can creatively communicate abstract concepts visually, sometimes in ways non-intuitive to humans
- Practical takeaway: if your agent sees images, test decisions with and without them, and don't trust reasoning traces to flag visual influence
Related event: A Heart in the Prompt Makes VLMs Write Recursion(2 posts)→
More from Models
- Mat Dryhurst: flash models optimized for a service all fail, same models fine outside it — matdryhurst · 2026-11-04
- 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