Apply to SPAR Spring 2025!

Preferred Pronouns

Data Center Inspections - Challenges, Benefits, and Precedents

Various projects on mechanistic interpretability and adversarial robustness

Benchmarking Language Agent Collusion in Bargaining Tasks

Extending upon Control Evaluations - Setting 1: APPS Backdoor Setting 

Extending upon Control Evaluations - Setting 2: Bash Control Setting

Extending upon Control Evaluations - Setting 3: White Box Control Setting

Extending upon Control Evaluations - Setting 4: METR Task Set Setting

Extending upon Control Evaluations - Setting 5: Side-Task Setting

Various mechanistic interpretability projects

Soft Nationalization - US Policy Analysis on Gov. Control of AI Labs

Overview of Systemic Economic Risks from TAI Systems

Robust Evaluation Framework for LLM Unlearning

GPU side-channels for weight theft and covert communication

Understanding and verifying the autoregressive conditioning hypothesis

Compact Proofs of Model Performance via Mechanistic Interpretability

Developmental Interpretability of Unlearning

Funding priorities for AI safety & governance

Value Alignment in Collective Intelligence Systems

Can you perform trusted computations on untrusted compute hardware?

Understanding Neural Networks with Sparse Autoencoders

Empirical Investigations to Assist LLM Evaluators

Evaluating causes and mitigations of hidden serial reasoning in foundation model agents

More Interpretable Models by Architecture Design

Science of Frameworks: Procedurally-generated eval framework to probe LLM agent tendencies and failure modes

SKATE - a peer-challenge approach to LLM ELO-ranking evaluation

Near Zero-knowledge Detection of Undesired Behavior

Finding model differences after applying post-training interventions

Toy model of computation in superposition

Developing and Applying Supervised and Unsupervised Probing Methods for Concept-Based Interpretability

Hybrid VAE-SAE Architectures Exploration: Disentangling Structured Features for LMs