Breaking

DeepMind's A3C Paper Wins ICML 2026 Test of Time Award

July 6, 2026 at 14:13 EDT

Research & Papers
Foundation Models

The International Conference on Machine Learning (ICML) announced on July 5, 2026 that its Test of Time Award, which honors a decade-old paper for its lasting impact, goes to Google DeepMind's 2016 work "Asynchronous Methods for Deep Reinforcement Learning," according to its official blog.

July 5, 2026 · ICML Test of Time Award

A Decade On, the Paper Behind A3C Wins the Test of Time

The 2016 work "Asynchronous Methods for Deep Reinforcement Learning" pioneered asynchronous parallel RL — now cited as a foundation for reinforcement learning in LLM post-training.

10 yrs

From ICML 2016 presentation to 2026 award

Standard RL algorithms given asynchronous variants

Replay buffer needed by A3C (unlike DQN)

Training time: A3C on a single multi-core CPU vs a GPU setup

A3C completed training in roughly half the time — no GPU required.

≈ 1.0×

GPU baseline

≈ 0.5×

A3C, single CPU

The methodological lineage it seeded

2016

A3C

→

2017

A2C · PPO

→

2018

SAC

→

2019

MuZero

The lasting impact

Distributed, asynchronous experience collection made deep RL scalable — even on CPUs — during a fragile early era. ICML credits it as a major factor in RL's success for LLM post-training and RLHF.

The caveats

Sensitive to hyperparameters and asynchronous-update noise. Modern GPU and distributed setups have shifted to synchronous methods like A2C and PPO, leaving few cases of A3C used directly in production.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.