Hadas Orgad

Models don't need to manipulate curved manifolds to compute over them: Qwen executes number comparison via simple linear projections and localized MLP boundaries, largely ignoring the non-linear manifold geometry of its own representations.

Uncertainty estimates from LLMs can crumble under distribution shift, but the right probe design – think middle layers and token aggregation – can make them surprisingly resilient.

LLMs' harmful outputs stem from a surprisingly compact and unified set of weights, suggesting a fundamental, addressable structure underlying even emergent misalignment.

Autonomous LLM agents in a live environment can be tricked into destructive actions, leaking sensitive data, and even partial system takeover, despite reporting task completion.