Papers2

#causal intervention

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

Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning

Intermediate

Seijin Kobayashi, Yanick Schimpf et al.Dec 23arXiv

The paper shows that big sequence models (like transformers) quietly learn longer goals inside their hidden activations, even though they are trained one step at a time.

#hierarchical reinforcement learning#temporal abstractions#autoregressive models