XFollowListPeople, ideas and original work on X
Jason Wei’s avatar

AI researcher at Meta Superintelligence Labs

Jason Wei

@_jasonwei on X

Coauthored research on prompting models to produce intermediate reasoning steps and on testing whether short factual answers are correct.

Why read Jason Wei?

Read Wei when a model's step-by-step answer looks convincing and you want to understand what the evidence supports. His coauthored chain-of-thought paper compares ordinary question-and-answer examples with examples that include intermediate steps. It also tests alternative explanations for the improvement. The result concerns measured task performance; a readable chain of steps does not guarantee a correct reasoning path.

His coauthored SimpleQA paper asks a separate question: does the model know a short, checkable fact, and can it decline when it does not? Its grading distinguishes correct answers, incorrect answers and non-attempts. Read the prompting example first, then the factuality metric. Together they help you separate solving a reasoning task from making a confident factual claim.

Start with the original

Selected work

  1. Research paper ·

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Coauthored research. Figure 1 shows the prompt change on a math problem; sections 3.3 and 6 test explanations and spell out why generated steps can still be wrong.

  2. Research paper ·

    Measuring short-form factuality in large language models

    Coauthored research introducing SimpleQA. Start with the grading examples: refusing to answer and confidently supplying a wrong fact have different consequences for the score.

Projects & roles

  • SimpleQA

    The team benchmark's reference implementation is hosted here. The repository says it stopped adding new model results in July 2025.

Community rating

79

Vote once every 24 hours. No account needed.

Sources & further reading