1

Probing whether vision-language models answer from visual evidence, language priors, or spurious visual representations.

Connecting JEPA representations to speech compression through density-adaptive attention and compact audio tokens.

A research system that combines retrieval and specialized agents for structured policy debate.

Compressing an LLM for the task it actually needs to perform, allocating precision to the layers that matter.

Using chess and Chess960, we investigate how strategic concepts appear across model layers and how stronger play can diverge from human conceptual understanding.

Detecting and reducing repetitive language patterns through sampling and targeted fine-tuning.

Connecting attention sinks to representation compression to explain how information changes across LLM layers.

Reassessing hallucination detection with evaluation metrics that better reflect meaning.

Finding useful embeddings inside language models. We study why intermediate layers can outperform the final layer and how to identify them using information, geometry, and downstream evaluation.

Studying which layers support mathematical reasoning and how their roles persist after post-training.