Yu-Yang Qian

PhD student @ LAMDA group, @NJU1902. Visiting PhD at @haoailab.

Interested in Efficient ML, Agentic RL, and Diffusion LLM.

[Hao AI Lab @haoailab]

(1/7) Last week FastVideo released FastH3. Today we bring it local: @NVIDIA DGX Spark 💚 and @Apple Silicon 🍎through MLX. 8x speedup!

FastH3 now supported on NVIDIA DGX Spark and Apple Silicon! over base @MiniMax_AI H3 for local generation!

This 15 seconds 1344×768 clip was generated on two NVIDIA DGX Sparks. Fully local. @NVIDIARTXSpark @NVIDIAAI

[Hao AI Lab @haoailab]

https://t.co/626ZrjAFHo

[Yu-Yang Qian @YuYangQian_ai]

@norpadon Congrats on the release! Really nice work, especially the factorized drafter + tree verification with GDN layers. BTW, you might also find our JetSpec interesting: it tackles the same draft-budget scaling ceiling with causal parallel tree drafting! https://t.co/RnuvfFBjUv

[Hao AI Lab @haoailab]

Thrilled to share our poster sessions at #ICML2026 in Seoul! ☕️🇰🇷

Stop by to chat with our lab members and collaborators about parallel decoding, diffusion LLMs, speculative decoding, video sparse attention, quantization, and more. We’re excited to connect!

1️⃣ Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections
📜 TL;DR: This work introduces MADQA, a benchmark with 2,250 human-authored questions grounded in 800 heterogeneous PDF documents, to test whether multimodal agents truly reason strategically or rely on brute-force search. The study finds that even the best agents can match human searchers in raw accuracy but still fail to close a nearly 20% gap to oracle performance.
📍Time and Location: Tue, Jul 7, 2026 • 10:30 AM – 10:45 AM KST, HALL B2

2️⃣ d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation
📜 TL;DR: d3LLM improves the accuracy–parallelism trade-off in diffusion LLMs through pseudo-trajectory distillation with multi-block decoding, achieving up to 10x speedup over dLLMs and 5x speedup over AR models with little accuracy drop.
📍Time and Location: Tue, Jul 7, 2026 • 2:00 PM – 3:45 PM KST, HALL A #2505

3️⃣ Fast and Accurate Causal Parallel Decoding using Jacobi Forcing
📜 TL;DR: Jacobi Forcing turns pretrained AR models into causal parallel decoders while preserving AR-level quality, achieving up to 4.0x wall-clock speedup on coding and math benchmarks, by leveraging higher-quality drafts as block sizes scale.
📍Time and Location: Wed, Jul 8, 2026 • 2:30 PM – 4:15 PM KST, HALL A #2207

4️⃣ When Drafts Evolve: Speculative Decoding Meets Online Learning
📜 TL;DR: OnlineSpec uses the verification feedback already produced by speculative decoding to continuously adapt draft models online. Grounded in dynamic regret minimization, it improves speculative decoding acceleration and achieves up to 24% speedup across 7 benchmarks and 3 foundation models.
📍Time and Location: Wed, Jul 8, 2026 • 2:30 PM – 4:15 PM KST, HALL A #1600

5️⃣ Attn-QAT: 4-Bit Attention With Quantization-Aware Training
📜 TL;DR: Attn-QAT makes 4-bit attention practical by matching low-precision attention-score recomputation during training. Across diffusion and language models, it recovers FP4 attention quality drops and delivers up to 1.5× speedup on an RTX 5090.
📍Time and Location: Thu, Jul 9, 2026 • 10:30 AM – 12:15 PM KST, HALL A #2800

[Hao AI Lab @haoailab]

https://t.co/lwO6dSHm9m

[Lanxiang Hu @Lanxiang_Hu]

https://t.co/M8JMUDIdKh

[Kaichao You @KaichaoYou]

wow glad to see the vLLM support with JetSpec!

[Yu-Yang Qian @YuYangQian_ai]

Very excited to be part of this work with @Lanxiang_Hu and @haoailab ! 🚀
Our JetSpec adds a lightweight, causality-preserving draft head, achieving up to 9.6× end-to-end speedup, fully plug-and-play. Paper, code, checkpoints, and vLLM engine are all open-source. Give it a try!

[Hao AI Lab @haoailab]

Introducing JetSpec: we find speculative decoding can push LLM generation latency to extreme by co-optimizing drafting cost and drafting quality with causal parallel tree drafting.

JetSpec reaches up to 9.64x end-to-end speedup on MATH-500 and 4.58x on open-ended chat while keeping lossless. With CUDA graph and kernel optimizations, JetSpec further translates to around 1000 TPS on a single B200. ⚡️

Check out our project page for demos and a blog post on how we built it 👇
https://t.co/M4T8jOBWQ8
https://t.co/h9uipDbTuh

[Hao AI Lab @haoailab]

Introducing JetSpec: we find speculative decoding can push LLM generation latency to extreme by co-optimizing drafting cost and drafting quality with causal parallel tree drafting.

JetSpec reaches up to 9.64x end-to-end speedup on MATH-500 and 4.58x on open-ended chat while keeping lossless. With CUDA graph and kernel optimizations, JetSpec further translates to around 1000 TPS on a single B200. ⚡️

Check out our project page for demos and a blog post on how we built it 👇
https://t.co/M4T8jOBWQ8
https://t.co/h9uipDbTuh

[Lanxiang Hu @Lanxiang_Hu]

We introduce JetSpec, which pushes speculative decoding acceptance length and TPS to a new frontier. We find that (1) preserving causality and (2) employing lightweight causal parallel decoding heads are key to scaling speculative decoding with trees, thereby converting sparse FLOPs more effectively into real end-to-end speedup. With serving-engine support, JetSpec achieves up to 1000 TPS on B200 GPUs for both Qwen3-8B and Qwen3-30B-A3B.

Check out our project page and blog for more details!
Project page: https://t.co/9ebBhoMWBI
Blog: https://t.co/N0wDKsUBHT

[Yu-Yang Qian @YuYangQian_ai]

OnlineSPEC 🚀 (https://t.co/cjdrOd3Szb): We observe that speculative decoding naturally forms an "online learning" loop. OnlineSPEC exploits this feedback to continuously evolve draft models, achieving up to 24% speedup over previous SOTAs (e.g., EAGLE-3).
See you in Seoul! (3/3)

[Yu-Yang Qian @YuYangQian_ai]

Very excited to have two papers accepted by ICML'26: d3LLM (https://t.co/1Dt3geBWpY) and OnlineSPEC (https://t.co/cjdrOd3Szb) 🎉🎉🎉 Both are about efficient LLM and model distillation. Many thanks to Prof. Hao Zhang @haozhangml and Prof. Peng Zhao @ZhaoPeng_NJU ! (1/3)

[Hao AI Lab @haoailab]

Excited to share our recent work accepted to ICML 2026! These projects span efficient causal parallel decoders, diffusion LLMs, video sparse attention, video QAT, online speculative decoding, and agentic document reasoning.

Huge thanks to all collaborators and co-authors across these efforts. Looking forward to seeing everyone in Seoul this summer! 🇰🇷

[Yu-Yang Qian @YuYangQian_ai]

d3LLM 🫧 (https://t.co/1Dt3geBWpY): We identify an accuracy-parallelism trade-off in diffusion LLMs. d3LLM pushes this Pareto frontier via pseudo-trajectory distillation and multi-block decoding, achieving ~10× speedup over LLaDA/Dream with minimal accuracy loss. (2/3)

[Hao AI Lab @haoailab]

Excited to share our recent work accepted to ICML 2026! These projects span efficient causal parallel decoders, diffusion LLMs, video sparse attention, video QAT, online speculative decoding, and agentic document reasoning.

Huge thanks to all collaborators and co-authors across these efforts. Looking forward to seeing everyone in Seoul this summer! 🇰🇷

[Peng Zhao @ZhaoPeng_NJU]

Nice work led by @YuYangQian_ai !
An ultra fast dLLM is developed, and we also present a new measure for diffusion model (AUP, Accuracy Under Parallelism)

read this blog for a quick overview: https://t.co/bHhzDMulQP
more details are in this arxiv paper https://t.co/YWfE2wuHM1

[Hao AI Lab @haoailab]

Introducing our new paper, d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation.

- 🚀 5x sppedup over AR models, 10× over vanilla LLaDA/Dream, with negligible accuracy degration
- 📊 We also introduce a new metric: AUP, to measure the accuracy-parallelism trade-off in dLLMs

📄 Our Paper is out! See arXiv:
https://t.co/UdGyGTKEKH

[Yu-Yang Qian @YuYangQian_ai]

Excited to share our new work, in which we reveal the best essential trade-off in dLLMs and design a new metric to better capture this trade-off. Furthermore, try our ultra-fast d3LLM!🚀 It may be the most efficient open-source dLLM ever!

[Hao AI Lab @haoailab]

🔥 New blog: AUP: when Accuracy Meets Parallelism in Diffusion Language Models.

🔗https://t.co/vdNerbNu8d

Diffusion LLMs promise parallel decoding, error correction, and random-order generation. But if you look at both speed and accuracy: Are dLLMs actually better than AR + speculative decoding?

Our study: not yet… Here’s why, and how we design our ultra-fast dLLM framework d3LLM 🚀 to close the gap!

[Hao Zhang @haozhangml]

Check out latest blog describing:

1. how to more appropriately characterizing the speed-accuracy trade-off of dLLMs (applies to any parallel decoding methods, too, such as LLMs + speculative decoding) and

2. Our ultra-fast d3llm which gives both strong speedup (5x over AR LLM and 10 over vanilla dLLM) and strong accuracy!

[Hao AI Lab @haoailab]

🔥 New blog: AUP: when Accuracy Meets Parallelism in Diffusion Language Models.

🔗https://t.co/vdNerbNu8d

Diffusion LLMs promise parallel decoding, error correction, and random-order generation. But if you look at both speed and accuracy: Are dLLMs actually better than AR + speculative decoding?

Our study: not yet… Here’s why, and how we design our ultra-fast dLLM framework d3LLM 🚀 to close the gap!

[Hao AI Lab @haoailab]

🚀 🚀 Dynasor is featured in @NVIDIA TensorRT-LLM new inference-time compute framework Scaffolding! Dynasor help cuts token usage by up to 29% with no accuracy loss!
🔍 NV Blog: https://t.co/S06S7dr4T4
Dynasor also Just accepted at #NeurIPS2025!