dai-2023-icl-paper-provenance-arxiv-code

IN premise — summaries/2026/08/24/dai-2023-icl-gradient-descent-s0-abstract.md

Created 2026-08-24T17:10:53+00:00

The ICL-as-gradient-descent paper (arXiv 2212.10559v3, ACL 2023) was authored across Peking University, Tsinghua University, and Microsoft Research, with code at https://aka.ms/icl.

Summary

This records where the "in-context learning as gradient descent" paper came from and where its implementation lives, so any downstream work that cites or builds on that result has a concrete, verifiable source to point to. It matters because it pins the claim to specific institutions and a public code repository, making it checkable rather than anonymous or unverifiable.