HaoKun Lin

Joint Ph.D. Student at UCAS & CityU

[HaoKun Lin @HaokunLin]

🚀 Introducing STEPQuant for recurrent state quantization in linear attention!

6-bit states → near-FP32 accuracy + >5× compression.

📄 Paper: https://t.co/26ypnLX20f
💻 Code: https://t.co/gNThpQm8Im

#LLM #Quantization

[HaoKun Lin @HaokunLin]

🎉 Update: “Efficient Diffusion Language Models: A Comprehensive Survey” has been accepted to EMNLP 2026 Findings!

[HaoKun Lin @HaokunLin]

📢 We just released “Efficient Diffusion Language Models: A Comprehensive Survey” + an accompanying Awesome list of efficient dLLM papers!

📌 PDF: https://t.co/rKd5nAvqco
🌟 Awesome list: https://t.co/FPeYf0m7NV

Feedback & PRs are welcome — let’s make dLLMs fast together! 🚀

[HaoKun Lin @HaokunLin]

STEPQuant tackles both dimensions with:
• Lifetime-aware Temporal Bit Allocation
• Key-Row-Aware Dual-axis Fitting
• FP16 pivots for high-risk units

Results:
�� Near-FP32 accuracy at 6-bit
→ >5× state compression
→ Up to 68.7% lower total serving memory

[HaoKun Lin @HaokunLin]

🚀 Introducing STEPQuant for recurrent state quantization in linear attention!

6-bit states → near-FP32 accuracy + >5× compression.

📄 Paper: https://t.co/26ypnLX20f
💻 Code: https://t.co/gNThpQm8Im

#LLM #Quantization

[HaoKun Lin @HaokunLin]

Why is recurrent state quantization hard?

Quantization errors accumulate over recurrent updates — but not equally.

⏳ Temporally, long-lived errors persist across many decoding steps.
📍 Spatially, errors in different state rows have very different impacts on model outputs.

[HaoKun Lin @HaokunLin]

Illustration of Cache Management for dLLMs.

[HaoKun Lin @HaokunLin]

Illustration of Parallel Decoding for dLLMs.

[HaoKun Lin @HaokunLin]

What’s inside:
⚡ Training efficiency (AR→dLLM, block diffusion, MoE)
⚡ Inference acceleration (parallel decoding, KV-cache, distillation, compression)
⚡ Long-context & variable-length generation
⚡ Frameworks/serving + multimodal diffusion

[DINQ @dinq_me]

You’ve done real work.
But most of it is hard to see.

DINQ brings your projects, code, and research onto one card.
No self-promotion. Just real signals.

Build your DINQ → https://t.co/IaEtN0Vkab

#DINQ

[𝚐𝔪𝟾𝚡𝚡𝟾 @gm8xx8]

TokLIP: Marry Visual Tokens to CLIP for Multimodal Comprehension and Generation

- adds CLIP semantics to VQ tokenizers via discrete-to-continuous arc (VQGAN + causal encoder, contrastive + distillation loss)
- disentangles generation and comprehension objectives
- enables end-to-end autoregressive training

Results:
- 80.0% ImageNet top-1 (384px), outperforms VILA-U / QLIP
- COCO TR@1 = 68.0, IR@1 = 52.87
- +9.1 on GQA and +2.3 on MMMU vs VILA-U
- FID↓ on ImageNet: 9.70 → 7.64 (300ep, GPT-B)

PAPER: https://t.co/7k16f4GeiS
CODE 🔜 https://t.co/UwJTj1EojP

[Ying Shan @yshan2u]

🚀 🚀 Introducing ARC-Hunyuan-Video-7B, a model that understands real-world short videos with accuracy and speed. The model excels at tasks essential for real-world comprehension—areas often overlooked by general-purpose video understanding models—while matching SOTA performance on broader benchmarks. Key strengths include:
🩷 Temporal Reasoning: Track events over time
🩷 Affective Intent: Spot emotions & sentiments
🩷 Creator Intent: Classify video motivations
🩷 Narrative Comprehension: Decode stories & plots
🩷 Humor/Meme Breakdown: Unpack jokes & references
🩷 Creative Analysis: Gauge originality & innovation
Paper, code, model, demo in comments.

[Chung Yiu Yau @yauchungyiu]

Our new quantization-aware training algorithm RoSTE can fine-tune quantized LLMs and break state-of-the-art benchmark accuracy on 4-bits Pythia, Qwen and Llama models! Thank you @quanwei0 for the great effort👏

Paper: https://t.co/ztc3jUjh7u
Code: https://t.co/UZpq9nyBUk

[DINQ @dinq_me]

The engine behind Google DeepMind’s Veo 3?

A powerhouse of engineers:
Abhishek Sharma,
Alina Kuznetsova,
Ali Razavi,
Aleksander Holynski,
Ankush Gupta,
Austin Waters,
Ben Poole,
Daniel Tanis,
Derek Gasaway
… and many more

It takes hundreds of minds — across ML, audio, and software — to build next‑generation video AI.

#DINQ #AITalents

[Jing Xiong @_June1126]

🔬 The HKU team presents ParallelComp: a training-free technique for efficient context length extrapolation in LLMs—from 8K up to 128K tokens—on a single A100 GPU, with minimal performance loss.

📄 Paper: https://t.co/HbKsGN0eqX
💻 Code: https://t.co/T2Au0WEGY1

[Yingjia Wan @Yingjia_Wan]

Is your model faithfully translating math into formal languages like Lean?
⚖ Introducing "FormalAlign"! #ICLR2025
⁉️To address the lack of scalable evaluation in autoformalization, we propose the FIRST method to evaluate semantic alignment between informal and formal languages.

[Han Wu @hahahawu2]

💡Unlocking Efficient Long-to-Short LLM Reasoning with Model Merging

We comprehensively study existing model merging methods on efficient Long-to-Short LLM reasoning tasks, and find their huge potential in the field.

[Yi Xu @_yixu]

🔥Are we ranking LLMs correctly?🔥

Large Language Models (LLMs) are widely used as automatic judges, but what if their rankings are unstable?😯Our latest study finds non-transitivity in LLM-as-a-judge evaluations—where A > B, B > C, but… C > A?! 🔄

[Zhijiang Guo✈️ICML2026 @ZhijiangG]

🚀Exciting to see how recent advancements like OpenAI’s O1/O3 & DeepSeek’s R1 are pushing the boundaries!
Check out our latest survey on Complex Reasoning with LLMs. Analyzed over 300 papers to explore the progress.
Paper: https://t.co/k1HGQTA2kN
Github: https://t.co/VpcNVcEBSg

[Chuanyang Jin @chuanyang_jin]

How to achieve human-level open-ended machine Theory of Mind?

Introducing #AutoToM: a fully automated and open-ended ToM reasoning method combining the flexibility of LLMs with the robustness of Bayesian inverse planning, achieving SOTA results across five benchmarks. 🧵[1/n]

[AK @_akhaliq]

RIFLEx

A Free Lunch for Length Extrapolation in Video Diffusion Transformers

TL;DR: Effortlessly extend your video with just one line of code: freq[k-1]=(2*np.pi)/(L*s).