2024 SAILAB LLM Summer Study

LLM Study conducted in SAILAB

Materials

THe scope of this study consists of two parts:

Internal Mechanism + Transformer : technical study on the mechanism.
LLM Insight : Insights on LLMs such as jailbreak, reliability, safety, or urgent research. You can select a paper describing novel approaches.

Date / Presenter	Material
07/03 (8:00 pm) Bumjin	Tech : Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet, Anthropic 2024. [link] Insight : Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews, ICML 2024. [link]
07/10 (08:00 pm)	Tech : Identifying Linear Relational Concepts in Large Language Models
07/10 (08:00 pm)
07/17 (08:00 pm)
07/24 (08:00 pm)
07/31 (08:00 pm)
August
08/01 (08:00 pm)
08/08 (08:00 pm)
08/15 (08:00 pm)
08/22 (08:00 pm)
08/29 (08:00 pm)
End

Anthropic :The list of selected papers in Anthropic
- Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet, Anthropic 2024. [link]
- TBD
ICML 2024 :The list of selected papers in openreview.
- Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews, ICML 2024. [link]
The Remarkable Robustness of LLMs: Stages of Inference?
A Systematic Comparison of Sparse Autoencoder Variants for Model Steering on the IOI Task
Investigating How Large Language Models Leverage Internal Knowledge to Perform Complex Reasoning
Memory Augmented Language Models through Mixture ofWord Experts
LLM Critics Help Catch LLM Bugs (OpenAI)