# Jason Wei > ~1995– · AI Researcher > > **Recorded contribution:** Chain-of-thought prompting; instruction tuning; emergent abilities of LLMs ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation AI researcher Jason Wei led or co-authored influential work on chain-of-thought prompting, instruction tuning, and analyses of apparent emergent abilities in large language models. His 2022 chain-of-thought paper showed that examples containing intermediate natural-language steps could improve performance on some reasoning benchmarks at sufficient scale. Public personal chronology is sparse, so the registry's approximate birth year remains unverified. Chain-of-thought prompting changes the demonstrations rather than the model weights: examples include intermediate natural-language steps before the answer, steering a sufficiently capable model toward a longer computation path. Instruction tuning similarly broadens supervised task mixtures so new instructions are more likely to evoke useful behavior. ## 2. The problem inherited Language models often produced wrong answers on multi-step tasks even when they could imitate relevant operations, and ordinary few-shot examples exposed only final labels. ## 3. The central contribution Wei and collaborators established chain-of-thought exemplars as a prominent prompting method and systematically studied instruction tuning and scale-dependent capability measurements. ## 4. Reconstruct the mechanism 1. Place worked examples containing question, intermediate text, and answer in the prompt. 2. Ask the model to continue the demonstrated pattern for a new problem. 3. Parse or score the final answer separately from the intermediate generated text. 4. Compare against answer-only prompting across model scales, tasks, perturbations, and alternative rationales. ## 5. What changed downstream - Chain-of-thought prompting became a standard tool and research object for eliciting multi-step behavior. - The work accelerated study of model scale, prompting, reasoning evaluation, faithfulness, and hidden computation. - These results reshaped prompting and post-training practice and stimulated research into reasoning traces, verifiers, test-time computation, and whether abrupt benchmark gains genuinely constitute emergence. ## 6. Attribution, limits, and uncertainty - The papers are multi-author works and build on scratchpads, rationales, prompting, and earlier intermediate-computation methods. - Generated reasoning can be unfaithful or post hoc; benchmark gains can reflect format imitation or contamination. Personal birth details remain unresolved. - A written rationale can be wrong, post hoc, or unfaithful to the mechanism producing the answer; benchmark thresholds and metric scales can also make smooth changes appear discontinuous. ## 7. Reconstruction lab Test answer-only and worked-rationale prompts on twenty arithmetic problems. Insert a plausible but wrong rationale and measure whether accuracy and explanation faithfulness diverge. Compare answer accuracy with step validity under correct, misleading, and hidden rationales, then plot raw continuous metrics before claiming an emergent transition. ## 8. Evidence trail - [Chain-of-Thought Prompting Elicits Reasoning in Large Language Models](https://arxiv.org/abs/2201.11903) — NeurIPS / arXiv - [Finetuned Language Models Are Zero-Shot Learners](https://arxiv.org/abs/2109.01652) — ICLR / arXiv - [Emergent Abilities of Large Language Models](https://arxiv.org/abs/2206.07682) — Transactions on Machine Learning Research --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*