James Campbell

About

I am CS PhD student at Carnegie Mellon University advised by Yiming Yang. At a high level, my research aims to understand the ingredients for general intelligence while ensuring its safety. Right now, this means thinking a lot about LLM agents and inference-time compute scaling.

As an undergrad at Cornell, I worked on LLM interpretability and truthfulness, and was a primary contributor to the papers Representation Engineering and Localizing Lying in Llama. I have also done work on semantic representations in the brain and LLM robustness.

Papers

Localizing Lying in Llama: Understanding Instructed Dishonesty on True-False Questions Through Prompting, Probing, and Patching

James Campbell*, Phillip Guo*, Richard Ren*

Accepted at NeurIPS 2023 SoLaR Workshop

Summary: We prompt Llama-2-70B-chat to lie and localize mechanisms involved using activation patching and linear probing.

ArXiv | Code | Thread

Representation Engineering: A Top-Down Approach to AI Transparency

Andy Zou, Long Phan*, Sarah Chen*, James Campbell*, Phillip Guo*, Richard Ren*, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, J. Zico Kolter, Dan Hendrycks

Summary: We introduce the field of Representation Engineering which seeks to understand and control LLM's using a top-down approach.

ArXiv | Website | Code | Coverage | Explainer

Considerations of Biological Plausibility in Deep Learning

James Campbell

Published front page of the Cornell Undergraduate Research Journal (CURJ)
Winner of the $300 James E. Rice Award

A literature review of learning algorithms and their biological plausibility

Paper

Projects

ProctorAI

Proctor is a multimodal AI companion that watches your screen and yells at you if it sees you being unproductive. Within a few days, it gained over 250 stars on Github.

Code

Theoretical Bound on the Weights of a Neural Network under Nesterov's Accelerated Gradient Flow

I solved an "open question" posed in a previous paper by deriving this bound on the weights of a neural network. This was done in summer of 2021 during an REU at Johns Hopkins under Rene Vidal.

PDF

Neural Network Linearization for Out-of-Distribution Robustness

In summer of 2022, I worked with Yi Ma on improving out-of-distribution robustness using the empirical neural tangent kernel.

Poster | Slides | Code

Data-driven clustering of neural responses to a large set of natural images

I investigated the clusterability of visual representations in the brain. This work was done in the Computational Connectomics Lab at Cornell and was presented as a poster at OHBM.

Abtract | Poster

Exploring Additive Compositionality in the Brain

I conducted experiments attempting to understand the extent to which semantic representations in the brain exhibit additive compositionality, i.e. does representation(A) + representation(B) = representation(A+B).

Slides | Code

Cornell Machine Learning Kaggle Competition

I came in first out of 155 participants in a Kaggle competition hosted by Cornell's big machine learning class.

Kaggle | Code

CereBERTo: Improving Distributional Robustness with Brain-Like Language Representations

Improving the out-of-distribution robustness of BERT and GPT-2 by pretraining them to predict brain fMRI data.