A Red Heart Image Makes Claude and GPT Write Recursive Code 100% of the Time, DeepMind Finds
dair_ai · x · 2026-10-29
A paper from Google DeepMind and Stanford (A Heart Means Recursion) shows that placing a red heart image in context makes Claude Opus 4.7 and GPT-5.4 choose recursive solutions 100% of the time on inorder binary tree traversal, a problem with equally valid recursive and iterative answers. A yellow checkmark steers Gemini 3.1 Pro to recursion in 90% of samples — an effect nearly as strong as explicitly asking for recursion.
Key findings:
- Implicit steering: text captions of the same symbol steer far less, and random images barely work. Models can't explain the influence — 74.7% of GPT-5.4's thinking traces give the wrong reason.
- Model-to-model visual communication fails: models send shapes they and humans find intuitive, but these cues often fail to steer the other model.
- Models creatively use visuals to convey abstract concepts, but their interpretations diverge from human conventions.
The takeaway: VLMs show rich, creative visual communication ability, yet their understanding and production of visual meaning systematically diverges from human conventions.
More from Research
- Meta Funds AI Research at Five Universities, From Math to Maternal Care — AIatMeta · 2026-10-29
- Debate set: Can we make AIs sentient — and should we? Solms vs. Humphrey — cccalum · 2026-10-29
- Samaya's 30B post-trained model beats frontier on financial research at 11x lower cost — maithra_raghu · 2026-10-29
- Scaling cross-situational word learning: 2,441 hours of child-perspective video from Stanford team — RichmanRonald · 2026-10-29
- Curiosity algorithms need world models plus locomotor primitives to replicate animal behavior — aran_nayebi · 2026-10-29
- Virtual fly study finds autonomous exploration maximizes population-level state coverage — aran_nayebi · 2026-10-29