ICLR 2026 Award Winners: Outstanding Papers and Alec Radford’s Test‑of‑Time Honor

ICLR 2026 was held from April 23 to 27 in Rio de Janeiro and received roughly 19,000 submissions, resulting in an overall acceptance rate of about 28% for peer‑reviewed full papers.

Outstanding Paper Award 1: "Transformers are Inherently Succinct" (authors: Pascal Bergsträßer, Ryan Cotterell, Anthony Widjaja Lin) proposes a new perspective on why Transformers are powerful: they can encode certain concepts far more succinctly than alternatives such as RNNs. The paper proves that Transformers can represent formal languages (e.g., finite automata and linear‑temporal‑logic formulas) with significantly fewer resources and shows that verifying Transformer properties is EXPSPACE‑complete.

Outstanding Paper Award 2: "LLMs Get Lost In Multi‑Turn Conversation" (authors: Philippe Laban, Hiroaki Hayashi, Yingbo Zhou, Jennifer Neville) points out a mismatch between training data (mostly single‑turn text completion) and real‑world deployment (multi‑turn dialogue). The authors design a scalable evaluation method and, through large‑scale simulated experiments, find that LLM performance drops 39% on average in multi‑turn settings, with both capability and reliability decreasing. They also observe that early wrong assumptions cause models to drift further off track.

Honorable Mention : "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.

Test‑of‑Time Award 1 : "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks" (DCGAN) by Alec Radford, Luke Metz, Soumith Chintala. This early GAN work demonstrated that learned generative models could synthesize diverse, realistic images, laying the foundation for modern image‑generation research.

Test‑of‑Time Award 2 : "Continuous control with deep reinforcement learning" (DDPG) by Timothy P. Lillicrap et al. Introduced the deep deterministic policy gradient algorithm, combining a deterministic actor‑critic architecture with DQN‑style stabilization, enabling direct translation of raw sensor inputs into precise continuous actions and sparking a revolution in reinforcement learning for control tasks.

Links to the papers: https://openreview.net/pdf?id=Yxz92UuPLQ, https://openreview.net/pdf?id=VKGTGGcwl6, https://arxiv.org/pdf/1511.06434, https://arxiv.org/pdf/1509.02971.