FigureAsia 35 Under 35 · Science
Yi Wu
Age 33 · Reinforcement learning systems · China
Principal investigator behind AReaL, a fully asynchronous system for reinforcement learning of large language models.
- Approximate age at the edition eligibility date
- 33
- Field
- Computer science
- Country or region
- China
- FigureAsia U35 Assessment
- 88.0 / 100
Profile
Career and documented record
Reinforcement learning for large models couples two uneven workloads: generating experience and updating the model. Yi Wu's AReaL system, published at NeurIPS 2025, fully decouples the two so rollout and training can proceed asynchronously on shared accelerator resources.
The team reported as much as a 2.77-fold training speed-up on the tested hardware and released the system openly. The contribution is more than scheduling. It supplies controls for stale experience and policy consistency—the scientific problems created when learning no longer proceeds in lockstep.
AReaL has already moved beyond a paper implementation into use by model-development teams. Its consequence is practical and methodological: faster experiments, better accelerator utilisation and a clearer systems foundation for studying reinforcement learning at scale.
FigureAsia selection
Why Yi Wu is on the list
Wu is selected for making an expensive research loop measurably more efficient without hiding the algorithmic costs of asynchrony. AReaL combines a strong systems idea, open implementation and early external use.
Verified work
The 2025–26 record
AReaL
Led a fully asynchronous system that decouples model rollout from reinforcement-learning updates.
2.77× measured speed-up
Reported up to 2.77-fold acceleration on the study's shared-GPU configurations.
Open and adopted system
Released the stack and documented use beyond the originating laboratory.
Field context
The work in its field
At scale, reinforcement learning is limited by orchestration as much as by algorithms. Asynchrony improves utilisation only if policy lag and data freshness remain controlled.
FigureAsia U35 Assessment
Assessment breakdown
88.0out of 100
Substantive 2025–2026 contribution
16.7 / 20
Led a fully asynchronous system that decouples model rollout from reinforcement-learning updates.
Verified scientific impact
13 / 15
AReaL reports substantial system gains and has moved into external model-development workflows.
Originality and distinction
8.8 / 10
The distinction lies in fully asynchronous rollout and learning with explicit mechanisms for the statistical costs of policy lag.
Field influence
8.9 / 10
Within computer science, the work matters because it shifts a live question in reinforcement learning systems rather than merely attracting attention.
Individual agency
8.8 / 10
Wu is the project principal investigator and senior research leader; implementation credit remains distributed across the authors.
Durability and trajectory
4.6 / 5
As Assistant Professor at Tsinghua University Institute for Interdisciplinary Information Sciences, Wu has a platform to carry the work into its next stage.
Asian significance and global relevance
4.6 / 5
Chinese computer scientist leading open reinforcement-learning systems research at Tsinghua University.
Evidential validity and reproducibility
7.1 / 8
Code and benchmarks are public, with the peak gain tied to measured configurations rather than universalised.
Advance in scientific knowledge
6.4 / 7
The work clarifies how asynchronous systems choices interact with reinforcement-learning data and optimisation.
Translational or methodological utility
4.5 / 5
It can cut accelerator idle time and shorten experimental cycles for large-model reinforcement learning.
Responsible research stewardship
4.6 / 5
The profile separates training efficiency from the separate questions of model behaviour and responsible deployment.