Mirage Hallucinations Linearly Decodable from Model Activations
ziv_ravid · x · 2026-07-04
Research shows that a model's Mirage behavior (fabricating images/details) can be linearly decoded from its internal activations, even when the images actually exist. Pure text baselines fail to recover this signal. The study also suggests that activation signals can help distinguish between different types of Mirage behaviors.
More from Research
- Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them — heathercmiller · 2026-10-27
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27