paper-buf

• LLM safety
  • Generalizing trust: Weak-to-strong trustworthiness in language models
  • InfAlign: Inference-aware language model alignment
  • Towards effective discrimination testing for generative ai
    • automatic testing of llm misbehaviors
• LLM
  • Why transformers are obviously good models of language, by Felix Hill
  • NOTES
  • Improving factuality with explicit working memory
  • Iteration head: A mechanistic study of chain-of-thought, NeurIPS 2024
  • NOTES
  • PromptBoosting: Black-box text classification with ten forward passes, ICML 2024
  • NOTES
  • Specifications: The missing link to making the development of llm systems an engineering discipline
  • NOTES
  • Do llms learn a true syntactic universal?, EMNLP 2024
  • NOTES
• Others
  • The unbearable slowness of being: Why do we live at 10 bits/s?

memorizing Felix

• Bye, Felix by Kyunghyun Cho
• Felix Hill is a continual inspiration and role model during my early PhD life. He is very thoughtful, especially on his theory of language understanding from a cognitive science viewpoint. Sadly, he died in Dec. 2024 due to his mental illness maybe coming from the stress of his work and the chaos of AI research with the unprecedented AI wave. His legacy [1] [2] [3] is worth reading together with his early influential research [4] [5] [6] [7], to know more about his inner mind before and during his struggling with his disease. Hope everyone in our field of AI has a sound mental health while under great pressure.

ai math tutor

• https://info.mathos.ai/

pyg-ssl

• https://github.com/iDEA-iSAIL-Lab-UIUC/pyg-ssl: A github repo for semi-supervised learning on graphs, based on pyg

Alec Radford’s talk at ICML 2021

• Some thoughts on generalization, robustness and their application with CLIP