Qingyang Wu
AI Researcher · Together AI
I am an AI Researcher at Together AI, where I work on large language models — post-training, reinforcement learning, and efficient inference.
I received my Ph.D. in Computer Science from Columbia University, advised by Prof. Zhou Yu, working on LLMs, dialog systems, and multimodal learning.
My research interests are broad, spanning coding agents (DeepSWE, DeepCoder, CoderForge), reinforcement learning and post-training, efficient inference, and multimodal models (LLaVA, GLIGEN). My goal is to build simple, efficient, and scalable systems for LLMs.
highlights
- XoRL — a Tinker-compatible RL system for high-performance distributed LLM training.
- CoderForge-Preview — the largest open test-verified coding agent dataset: 258K trajectories across 51K tasks.
- DeepSWE — state-of-the-art open-source coding agent trained from scratch with RL.
- DeepCoder-14B — fully open-source 14B coder model at o3-mini level.
- LLaVA: Visual Instruction Tuning — the original open-source visual instruction-tuned assistant. NeurIPS 2023 Oral, 15,000+ citations.
- GLIGEN: Grounded Text-to-Image Generation — open-set grounded image generation with spatial control. CVPR 2023, 1,300+ citations.
open-source projects
- TokenSpeed — a speed-of-light LLM inference engine designed for agentic workloads.
- TorchSpec — a torch-native framework for speculative decoding training at scale.
- Aurora — an RL framework that makes speculative decoding self-improving via a serve-to-train loop.
- XoRL — a Tinker-compatible RL system: high-performance distributed training for LLMs (RL, SFT, MoE, and beyond).
talks
| Jun 2026 | Open-Source Inference Engineering for the Agentic Era — workshop on TokenSpeed at the AI Engineer World's Fair 2026, San Francisco. With Zain Hasan, Yubo Wang, and Jue Wang. |
publications
Also on Google Scholar.
Taylor-Calibrate: Principled Initialization for Hybrid Linear Attention Distillation
arXiv preprint, 2026
Squeeze Evolve: Unified Multi-Model Orchestration for Verifier-Free Evolution
arXiv preprint, 2026
Opportunistic Expert Activation: Batch-Aware Expert Routing for Faster Decode Without Retraining
ICML 2026
Understanding and Steering the Cognitive Behaviors of Reasoning Models at Test-Time
arXiv preprint, 2025
AdapTive-LeArning Speculator System (ATLAS): A New Paradigm in LLM Inference via Runtime-Learning Accelerators
Together AI, 2025
Data Diversification Methods in Alignment Enhance Math Performance in LLMs
NeurIPS 2025 Workshop on Efficient Reasoning
How Well Can General Vision-Language Models Learn Medicine By Watching Public Educational Videos?
arXiv preprint, 2025
Think Deep, Think Fast: Investigating Efficiency of Verifier-free Inference-time-scaling Methods
arXiv preprint, 2025
DECOR: Improving Coherence in L2 English Writing with a Novel Benchmark for Incoherence Detection, Reasoning, and Rewriting
EMNLP 2024
kNN-ICL: Compositional Task-Oriented Parsing Generalization with Nearest Neighbor In-Context Learning
NAACL 2024
Towards Better Language Models: Algorithms, Architectures, and Applications
Ph.D. Thesis, Columbia University, 2024
DiactTOD: Learning Generalizable Latent Dialogue Acts for Controllable Task-Oriented Dialogue Systems
SIGDIAL 2023
Quantifying Intrinsic Uncertainty in Classification via Deep Dirichlet Mixture Networks
arXiv preprint, 2019
Deep Learning Based RF Fingerprinting for Device Identification and Wireless Security
Electronics Letters, 54(24), 2018