Collin Burns

Career

Collin Burns did doctoral research at UC Berkeley, co-advised by Jacob Steinhardt and Dan Klein, and Steinhardt's page lists Burns among former students who went on to OpenAI. Burns has since worked at OpenAI and at Anthropic, and Burns's own site gives the current description of AI researcher.²¹

Research

Burns is co-first author, with Haotian Ye, of "Discovering Latent Knowledge in Language Models Without Supervision" (ICLR 2023), written with Dan Klein and Jacob Steinhardt. The paper answers yes-no questions using only a model's unlabeled internal activations.⁴¹

Burns is the first-listed author of "Weak-to-Strong Generalization" (ICML 2024, oral), which tests whether a weaker model's labels can bring out the abilities of a stronger pretrained model, as a stand-in for humans overseeing future systems.³¹

Earlier, Burns co-authored two evaluation benchmarks with Dan Hendrycks and collaborators: MMLU, which measures a model's accuracy across 57 subjects, and MATH, a dataset of competition mathematics problems.⁵⁶¹

Work and contributions

• Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision Published research

  First-listed author and one of the paper's primary contributors; the 2023 study tests whether weak model supervision can elicit the full capabilities of a much stronger pretrained language model.³¹

• Discovering Latent Knowledge in Language Models Without Supervision Published research

  Co-first author of the ICLR 2023 paper, which introduces an unsupervised method for answering yes-no questions from a language model's internal activations alone.⁴¹

• Measuring Massive Multitask Language Understanding (MMLU) Published research

  Co-author of the 2020 paper proposing a test of a text model's multitask accuracy across 57 tasks.⁵¹

• Measuring Mathematical Problem Solving With the MATH Dataset Published research

  Co-author, with Dan Hendrycks and others, of the 2021 paper introducing the MATH dataset of competition mathematics problems, presented in the NeurIPS 2021 Datasets and Benchmarks Track.⁶¹

Timeline

1. Sep 7, 2020Co-authored the MMLU benchmark paper, Measuring Massive Multitask Language Understanding.⁵
2. Mar 5, 2021Co-authored the paper introducing the MATH dataset of competition mathematics problems.⁶
3. Dec 7, 2022Released Discovering Latent Knowledge in Language Models Without Supervision as co-first author.⁴
4. Dec 14, 2023Released Weak-to-Strong Generalization as first-listed author.³

Documented roles

The sources do not confirm a current role; these are the roles they document.

• Worked at OpenAI, OpenAI¹²
• Worked at Anthropic, Anthropic¹
• PhD student (co-advised by Jacob Steinhardt and Dan Klein), University of California, Berkeley²

Companies in their history

• OpenAI 63 people in the wiki
• Anthropic 34 people in the wiki
• UC Berkeley 27 people in the wiki

Related people

• Diogo AlmeidaConnected to OpenAI
• Amanda AskellConnected to OpenAI
• Shengjia ZhaoConnected to OpenAI
• Shunyu YaoConnected to OpenAI
• Jacob SteinhardtConnected to OpenAI
• Margaret MitchellSafety and alignment · Evaluation and benchmarks
• Eric Michael SmithSafety and alignment · Evaluation and benchmarks
• Ethan PerezSafety and alignment · Evaluation and benchmarks
• Owain EvansSafety and alignment · Language models
• Samuel R. BowmanSafety and alignment · Evaluation and benchmarks