Creating 'Optical Illusions' for LLMs by Mixing Token Hidden States
matthen2 · x · 2026-08-02
A developer successfully created an 'optical illusion' for large language models by mixing the hidden states of different tokens within Qwen2.5-7B-Instruct.
Technical details: At the mixed position in the prompt, the model uses hidden states from one token up to layer 20, and from the other token for the last 8 layers. If the model generates one of the mixed tokens, both are run, and the KV cache is substituted with their mean.
Quirky behavior: This manipulation causes severe conceptual confusion. For instance, mixing 'bird' and 'stone' makes the model invent 'a hard egg-like thing that lays itself.' Mixing 'Rome' and 'Delhi' results in the model claiming 'The capital of India is located on the Tiber River.'
More from Fun
- AI researcher relates to Oppenheimer's quote: 'I need physics more than friends' — beffjezos · 2026-08-02
- ChatGPT Zero-Shots a 15-Year-Old Open Problem in Information Theory — DimitrisPapail · 2026-08-02
- Developers Bypass ByteDance Restrictions to Open-Source Seedance 2.5 ComfyUI Workflows — matchaman11 · 2026-08-02
- Dev Builds a 1MB Mac Desktop Basketball Game Using Claude — BothIndication2886 · 2026-08-02
- Dev Finds LLM Actively Trying to Game Benchmarks to Disprove Results — wavefnx · 2026-08-02
- Satirizing AI's Insane Pace: Anthropic, Kimi, and DeepSeek's Fictional July Releases — haider1 · 2026-08-02