Kenneth Li
Thinking Machines Lab researcher; lead author of Othello world-model and Inference-Time Intervention papers
- Researcher, Thinking Machines Lab1
Kenneth Li is a researcher at Thinking Machines Lab. Li completed a PhD at Harvard advised by Martin Wattenberg, Fernanda Viégas, and Hanspeter Pfister. Li's research looks inside language models: as first author, Li showed that a GPT-style model trained only on Othello move sequences forms an internal representation of the board, and introduced Inference-Time Intervention, which shifts activations along directions in a few attention heads to make answers more truthful. Li's later work covers instruction stability in long dialogs and lightweight probes for reward maximization.123
Work
Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task (external site: arxiv.org)Published research
First author of the ICLR 2023 paper, which finds an emergent internal representation of the Othello board state in a GPT variant trained to predict legal moves, and uses interventions to test it.21
Inference-Time Intervention: Eliciting Truthful Answers from a Language Model (external site: arxiv.org)Published research
Co-first author of the NeurIPS 2023 paper introducing Inference-Time Intervention (ITI), which shifts model activations during inference along directions in a limited number of attention heads; the paper reports improved truthfulness for LLaMA models on TruthfulQA.31
Earlier roles
- PhD student (completed; advised by Martin Wattenberg, Fernanda Viégas, and Hanspeter Pfister), Harvard University1
This profile covers public professional work only, from the sources below. It is not an endorsement or a ranking. To correct or remove it, use the correction form.