Yang Chen

I am a research scientist at NVIDIA, where I work on post-training for large language models. I received my PhD in computer science from Georgia Tech, where I was fortunate to be advised by Alan Ritter and Wei Xu. Previously, I interned at Google DeepMind (formerly Google Brain) and TTIC.

[Google Scholar] [CV] [X]

Projects

Unified Audio Intelligence Without Regressing on Text Intelligence
Core contributor: multi-domain on-policy distillation (MOPD).
Technical Report, 2026. 300+ likes on Hugging Face.
[model] [collection] [demo] [abstract]

Nemotron-Labs-Audex-30B-A3B, a unified audio-text LLM built on Nemotron-Cascade-2-30B-A3B, delivering state-of-the-art audio understanding, speech recognition and translation, text-to-speech, audio generation, and speech-to-speech generation with marginal or no regression on text intelligence.
Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning
Core contributor: math proof intelligence and MOPD.
Technical Report, 2026. 350+ likes on Hugging Face.
[model] [abstract]

A 550B total / 55B active parameter MoE hybrid Mamba-Attention model with 1M-token context, post-trained with SFT, RL, and multi-teacher on-policy distillation, reaching up to ~6x higher inference throughput than comparable open LLMs at on-par accuracy.
Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation
Core contributor: math post-training, IMO 2025 gold-medal performance, MOPD.
Technical Report, 2026. 600+ likes on Hugging Face.
[model] [data] [collection] [abstract]

We scale Cascade RL and multi-domain on-policy distillation to train Nemotron-Cascade-2-30B-A3B, achieving gold-medal-level results on IMO, IOI, and ICPC World Finals, along with strong performance on ArenaHard and instruction following.
Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models
Core contributor: math reinforcement learning.
In Conference on Neural Information Processing Systems (NeurIPS), 2026. Oral (top 0.3%). 300+ likes on Hugging Face.
[models & data] [abstract]

We scale Cascade RL to train general-purpose reasoning LLMs spanning RLHF, instruction following, math, code, and SWE. Our 14B Thinking model achieves SOTA on LiveCodeBench, outperforming Gemini-2.5-Pro, o4-mini, Qwen3-235B and DeepSeek-R1-671B, and reaches silver-medal performance at IOI 2025.
AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy
Zihan Liu, Zhuolin Yang, Yang Chen, Chankyu Lee, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping.
In International Conference on Learning Representations (ICLR), 2026. 100+ likes on Hugging Face.
[models] [data] [abstract]

We further scale SFT and study its interplay with RL, releasing the SOTA 7B model, AceReason-Nemotron-1.1-7B.
AceReason-Nemotron: Advancing Math and Code Reasoning through Reinforcement Learning
Yang Chen*, Zhuolin Yang*, Zihan Liu, Chankyu Lee, Peng Xu, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. (*equal contribution)
In Conference on Neural Information Processing Systems (NeurIPS), 2025. 100+ likes on Hugging Face.
[models] [data] [abstract]

We scale RL to both math and code domains, releasing the SOTA medium-sized model, AceReason-Nemotron-14B.
AceMath-RL-Nemotron: Training Math Reasoning Model with Reinforcement Learning
Yang Chen, Zihan Liu, Chankyu Lee, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping.
Research Blog, 2025.
[model] [abstract]

Our pilot study on scaling RL for competitive math reasoning LLMs, released as AceMath-RL-Nemotron-7B.
AceMath: Advancing Frontier Math Reasoning with Post-Training and Reward Modeling
Zihan Liu*, Yang Chen*, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. (*equal contribution)
In Findings of the Association for Computational Linguistics (ACL Findings), 2025. 200+ likes on Hugging Face.
[models & data] [abstract]

Math reasoning and reward models: AceMath-1.5B / 7B / 72B, plus AceMath-RM-7B / 72B, which surpassed GPT-4o and Qwen2.5-Math at the time.

PhD Research

Translation and Fusion Improves Cross-lingual Information Extraction
Yang Chen, Vedaant Shah, and Alan Ritter.
In Annual Meeting of the Association for Computational Linguistics (ACL), 2025.
Granular Privacy Control for Geolocation with Vision Language Models
Ethan Mendes, Yang Chen, James Hays, Sauvik Das, Wei Xu, and Alan Ritter.
In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2024. Oral.
UniIR: Training and Benchmarking Universal Multimodal Information Retrievers
Cong Wei, Yang Chen, Haonan Chen, Hexiang Hu, Ge Zhang, Jie Fu, Alan Ritter, and Wenhu Chen.
In European Conference on Computer Vision (ECCV), 2024. Oral.
Constrained Decoding for Cross-lingual Label Projection
Duong Minh Le, Yang Chen, Alan Ritter, and Wei Xu.
In International Conference on Learning Representations (ICLR), 2024.
Can Pre-trained Vision and Language Models Answer Visual Information-Seeking Questions?
Yang Chen, Hexiang Hu, Yi Luan, Haitian Sun, Soravit Changpinyo, Alan Ritter, and Ming-Wei Chang.
In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023.
Open-domain Visual Entity Recognition: Towards Recognizing Millions of Wikipedia Entities
Hexiang Hu, Yi Luan, Yang Chen, Urvashi Khandelwal, Mandar Joshi, Kenton Lee, Kristina Toutanova, and Ming-Wei Chang.
In International Conference on Computer Vision (ICCV), 2023. Oral.
Frustratingly Easy Label Projection for Cross-lingual Transfer
Yang Chen, Chao Jiang, Alan Ritter, and Wei Xu.
In Annual Meeting of the Association for Computational Linguistics (ACL), 2023.
Model Selection for Cross-Lingual Transfer
Yang Chen and Alan Ritter.
In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021.
GigaBERT: Zero-shot Transfer Learning from English to Arabic
Wuwei Lan, Yang Chen, Wei Xu, and Alan Ritter.
In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020.