method
active
method:visual-illusions-for-language-modelsVisual illusions for language models
Character sequences that hijack counting attention heads.
Neighborhood — ranked by edge-count
Papers (1)
paper
- When Models Manipulate Manifolds: The Geometry of a Counting Taskintroducesmentions
Concepts (1)
concept
- Müller-Lyer illusionanalogous_toHuman line-length illusions as analogy for @@ illusion.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Motivates illusion experiment.
- Primary substrate for manifold steering experiments; demonstrates method on reasoning and in-context tasks.
- Primary test domain for manifold steering, including reasoning and ICL tasks
- Language models able to generate images could develop a visual component of their workspace.hypothesis0.782Prediction from verbal-output account.
- Features related to gender, racial, ethnic biases, slurs, and hate speech.
- Claude 3 Opus lying to auditors; prior case study of deceptive tendencies
- LLM behavior of generating falsehoods; the multi-dimensional truth subspace raises new risks for subtle manipulation
- Demonstrated transformers on mathematical understanding and logic; cited to motivate transformer versatility.