ICLR 2026 Award Winners: Two Outstanding Papers and Alec Radford’s Classic Work Honored with Test‑of‑Time Award

ICLR 2026 Overview

ICLR 2026 received roughly 19,000 valid full‑paper submissions and accepted about 28% after peer review.

Outstanding Paper Awards

Transformers are Inherently Succinct

Authors: Pascal Bergsträßer, Ryan Cotterell, Anthony Widjaja Lin

Link: https://openreview.net/pdf?id=Yxz92UuPLQ

The paper introduces a theoretical framework that measures a model’s ability to encode formal concepts succinctly. Succinctness is defined as the number of parameters required to represent formal languages such as finite automata and linear‑temporal‑logic (LTL) formulas. The authors prove that a Transformer can represent these languages with significantly fewer parameters than standard representations based on finite automata or LTL, demonstrating a stronger expressive capability compared with recurrent neural networks (RNNs). As a corollary, they show that the decision problem for Transformer properties is EXPSPACE‑complete, establishing theoretical intractability.

LLMs Get Lost In Multi‑Turn Conversation

Authors: Philippe Laban, Hiroaki Hayashi, Yingbo Zhou, Jennifer Neville

Link: https://openreview.net/pdf?id=VKGTGGcwl6

The authors identify a mismatch between training data—predominantly single‑turn text‑completion—and real‑world deployment, which often involves multi‑turn dialogues with ambiguous or incomplete instructions. They design a scalable evaluation framework for multi‑turn capability and run large‑scale experiments covering six generation tasks and more than 200 k simulated dialogues. Results show a consistent performance drop of 39 % on average when moving from single‑turn to multi‑turn settings. Analysis attributes the degradation to two factors: (1) a modest decline in the model’s intrinsic ability and (2) a substantial loss of reliability. The study also observes that LLMs frequently make premature assumptions early in a conversation, leading to a cascade of errors that are difficult to recover from.

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 Awards (ICLR 2016 papers)

Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks (DCGAN)

Authors: Alec Radford, Luke Metz, Soumith Chintala

Link: https://arxiv.org/pdf/1511.06434

DCGAN was among the first works to demonstrate that learned generative models could synthesize diverse, realistic, and complex images, establishing a foundation for modern image‑generation research. Its architecture—deep convolutional generators trained with adversarial loss—proved that unsupervised representation learning could produce high‑quality visual samples, influencing subsequent developments such as diffusion models.

Continuous Control with Deep Reinforcement Learning (DDPG)

Authors: Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, Daan Wierstra

Link: https://arxiv.org/pdf/1509.02971

Before DDPG, applying reinforcement learning to physical systems suffered from hand‑crafted state features and the curse of dimensionality caused by discretization. DDPG combines a deterministic actor‑critic architecture with stabilization techniques from DQN, enabling neural networks to map raw sensor inputs directly to precise continuous actions. This algorithm demonstrated that deep reinforcement learning could succeed in continuous‑control domains, reshaping the field and spurring extensive follow‑up research.