Papers3

#attention heads

Instruction Anchors: Dissecting the Causal Dynamics of Modality Arbitration

The paper asks a simple question: when an AI sees a picture and some text but the instructions say 'only trust the picture,' how does it decide which one to follow?

#multimodal instruction following#modality arbitration#instruction tokens

Not triaged yet

Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models

Intermediate

Hengyuan Zhang, Zhihao Zhang et al.Jan 20arXiv

This survey turns model understanding into a step-by-step repair toolkit called Locate, Steer, and Improve.

#mechanistic interpretability#residual stream#attention heads

Not triaged yet

Mechanistic Interpretability of Large-Scale Counting in LLMs through a System-2 Strategy

Intermediate

Hosein Hasani, Mohammadali Banayeeanzade et al.Jan 6arXiv

Large language models (LLMs) are good at many math problems but often mess up simple counting when the list gets long.

#mechanistic interpretability#counting in LLMs#System-2 prompting

Not triaged yet