The ICLR 2026 outstanding papers
The conference named two outstanding papers: a theoretical one on the succinctness of Transformers and an empirical one showing models lose the thread in multi-turn conversation.
Why it matters
One of the two top papers of the year at a leading conference is a negative result about the very mode in which the models are used every day.
The announcement is dated 23 April 2026: two Outstanding Papers and one Honorable Mention. The first is Transformers are Inherently Succinct by Pascal Bergstrasser, Ryan Cotterell and Anthony Widjaja Lin, a theoretical account of the architecture's power through how succinctly it encodes concepts compared with alternatives such as recurrent networks. The second is LLMs Get Lost In Multi-Turn Conversation by Philippe Laban, Hiroaki Hayashi, Yingbo Zhou and Jennifer Neville, which finds a marked drop in capability in multi-turn settings. The Honorable Mention is The Polar Express: Optimal Matrix Sign Methods and their Application to the Muon Algorithm by Noah Amsel, David Persson, Christopher Musco and Robert M. Gower. The committee describes the selection as rigorous and involving external expert consultation.