Ari Holtzman

Mapping Overlaps in Benchmarks through Perplexity in the Wild

Dynamic Weight Grafting: Localizing Finetuned Factual Knowledge in Transformers

LLM Probability Concentration: How Alignment Shrinks the Generative Horizon

Linearly Decoding Refused Knowledge in Aligned Language Models

Know Thyself? On the Incapability and Implications of AI Self-Recognition

Forking Paths in Neural Text Generation

MUSE: Machine Unlearning Six-Way Evaluation for Language Models

Leveraging Set Assumption for Membership Inference in Language Models