Hi, I'm Meg! 

I currently work on research infrastructure at Anthropic.

My previous research has been on:

◼ jailbreaking

• Constitutional classifiers: Defending against universal jailbreaks across thousands of hours of red teaming

◼ risk forecasting

• Forecasting Rare Language Model Behaviors

◼ evaluations

• • The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A"

LLMs fail to represent facts in an order-invariant way

• • Taken out of context: On measuring situational awareness in LLMs

LLMs can successfully pass tests in-context having only seen descriptions of the tests in their training data

• • Towards Understanding Sycophancy in Language Models

LLMs trained with RLHF show sycophantic behavior which may be caused by human preferences

I've also contributed to research on:

◼ safety

• Many-shot Jailbreaking

• Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

◼ steering

• Evaluating feature steering: A case study in mitigating social biases

• Steering Llama 2 via Contrastive Activation Addition