Wen-Ding Li

Timeline

1. 2018

   Frobenius additive FFT

   In 2018, they are first author of "Frobenius Additive Fast Fourier Transform", written with Ming-Shing Chen, Po-Chun Kuo, Chen-Mou Cheng and Bo-Yin Yang. The paper extends the Frobenius FFT to additive FFTs and multiplies two binary polynomials with a product of degree below 256 in 29,005 bit operations, down from the previous best of 33,397.

2. 2022–2023

   Code generation in data notebooks

   They co-author "Natural Language to Code Generation in Interactive Data Science Notebooks" with Pengcheng Yin, Charles Sutton and others, posted in December 2022 and published at ACL 2023. The paper builds ARCADE, a benchmark of 1082 pandas code-generation problems, and PaChiNCo, a 62B-parameter code language model for Python notebooks. Their industry work at Google and X is on large language models for code.

3. 2023

   LambdaBeam program search

   In 2023, they co-author "LambdaBeam: Neural Program Search with Higher-Order Functions and Lambdas" with Kensen Shi, Hanjun Dai, Kevin Ellis and Charles Sutton. The search algorithm constructs lambda functions that compose operations within a DSL, and a neural policy network chooses which lambdas to build.

4. 2024

   StarCoder 2 and LiveCodeBench

   In 2024, they are among the authors of StarCoder 2 and The Stack v2 from the BigCode project, which trains code models with 3B, 7B and 15B parameters. They also co-author LiveCodeBench, a contamination-free benchmark that collects new problems from LeetCode, AtCoder and CodeForces.

5. 2024

   Programming by example with LLMs

   With Kevin Ellis, they write "Is Programming by Example solved by LLMs?", published at NeurIPS 2024. The paper finds pretrained models weak at programming by example and shows that fine-tuning lifts performance much higher, provided the test problems are in-distribution.

6. 2024

   An ARC Prize paper award

   As a doctoral student in computer science at Cornell, in a team led by Kevin Ellis, they and Keya Hu are lead authors of "Combining Induction and Transduction for Abstract Reasoning". The paper wins the first-place paper award in the 2024 ARC Prize competition, a $50,000 prize shared by the two. Their model searches for a program that explains each puzzle, Hu's model predicts the answer directly, and the two models together solve about 56% of the problems.

7. June 2025

   Magistral reasoning model

   In June 2025, they are listed among the authors of Magistral, which introduces Mistral's first reasoning model together with its own scalable reinforcement learning pipeline. Magistral Medium is trained on top of Mistral Medium 3 with reinforcement learning alone, and Magistral Small is open-sourced under Apache 2.0.

8. 2025–2026

   Devstral and Ministral 3

   They are listed among the authors of the Devstral report in 2025, which presents Devstral-Small, a 24B model for code agents. They are also listed on the Ministral 3 report in January 2026, which presents dense language models with 3B, 8B and 14B parameters.