Sandbagging Evaluation in LLMs
An Inspect-based evaluation testing whether frontier LLMs sandbag after discovering a hidden threshold under different framings (human-inspected, instructed, uninstructed).
Diffing for Model Spec Midtraining
A mechanistic study of whether model diffing recovers a value installed during Model Spec Midtraining in Llama 3.1 8B.
Inverse Scaling in Chain-of-Thought Faithfulness
A study of whether CoT reasoning faithfulness decreases as model scale increases, tested across 11 open-weight Llama and Qwen models.
nanoKimi
A custom implementation of the Kimi-K2 MoE architecture with a Muon optimizer and latent attention.
MCRAG
A multi-LLM system, inspired by GAN's design, that uses a generator and two critics to generate and iteratively improve written code based on critics' feedback.