Mμ Lab

Muhan’s Research Lab @PKU1898 @muhanzhang93
Revealing the order behind the noise.
Lab Website:

[Mμ Lab @mulabpku]

🎉Welcome to Muhan's (@muhanzhang93) Research Lab!

Mμ Lab is dedicated to pursuing principled and transformative research in artificial intelligence and machine learning.
While our current focus spans graph learning and large language models, our long-term mission is broader: to accelerate the development of artificial general intelligence (AGI) and deepen the scientific understanding of intelligence itself.

[Mμ Lab @mulabpku]

Check out our project page for the full position paper: 🔗 https://t.co/YS0hTKMCYU

[Mμ Lab @mulabpku]

Can we build true lifelong AI simply by relying on longer context, external docs, and complex harnesses?
In our latest position paper, we argue no. ❌
Future AI demands more than in-context learning (ICL)—it requires continuous in-parameter learning (IPL). 🧵👇

[Mμ Lab @mulabpku]

As a concrete step, SHINE demonstrates the potential of IPL: an in-context hypernetwork that maps any context into a LoRA module in a single forward pass—instant weight generation in seconds!

Paper: https://t.co/St9RV8ye8S

[alphaXiv @askalphaxiv]

"HISA: Efficient Hierarchical Indexing for Fine-Grained Sparse Attention"

Sparse attention can still be slow.

And the slow part is often not the attention step itself, but the search step that scans the whole context to find useful tokens.

This paper's HISA makes that search cheaper. It first finds the best blocks, then finds the best tokens inside those blocks.

This keeps token-level precision, needs no retraining, works with the same downstream attention, and gives up to 3.75x speedup while staying close to the original quality.

[DailyPapers @HuggingPapers]

DeepSeek Sparse Attention gets a hierarchical upgrade

HISA replaces the flat token scan with a two-stage block-then-token filtering pipeline, eliminating the indexing bottleneck at 64K context without any additional training.

[The AI Timeline @TheAITimeline]

🚨This week's top AI/ML research papers:

- HISA
- Embarrassingly Simple Self-Distillation Improves Code Generation
- FIPO
- SKILL0
- Reasoning over mathematical objects
- Screening Is Enough
- Path-Constrained Mixture-of-Experts

read this in thread mode for the best experience

[Mμ Lab @mulabpku]

Thanks for sharing our work!

[DailyPapers @HuggingPapers]

Paper: https://t.co/qKuBreq4kF

On Needle-in-a-Haystack and LongBench, HISA closely matches the original DeepSeek Sparse Attention in quality while substantially outperforming block-sparse baselines.

[Cai Zhou @ ICML2026 @zhuci19]

4/4 Takeaway

Canonicalization is a principled “gauge fixing” for symmetry: collapse each orbit to a unique representative, learn with expressive generic backbones and less variance, and preserve invariance by construction via random symmetry transforms at sampling.

Paper: arXiv 2602.15022

[Cai Zhou @ ICML2026 @zhuci19]

Stop paying the equivariance tax — canonicalize instead!

We propose canonical diffusion: map each molecule to a canonical pose/order, train a plain (non-equivariant) diffusion/flow on the canonical slice, then Haar-randomize at sampling to restore invariance.

CanonFlow reaches SOTA on GEOM-DRUG, with a large edge in few-step generation.

arXiv: 2602.15022
https://t.co/4tYpqjiFSm

[Yi Hu @AheadOFpotato]

🔥Our survey on LRM Mechanisms is now available at: https://t.co/IHXOlgh1JH

[Mμ Lab @mulabpku]

Heating Continual Learning/Knowledge Adaptation with Reasoning Skills! 🔥
Great work led by @tngpngzh111992 @blancokdb!

[Yiding Wang @ACL 2026 @blancokdb]

SFT gives the model "knowledge," but RL gives it the "skill" to use it. In Continual Learning, how do we get both efficiently?

Introducing PaST: A method to inject pre-learned reasoning skills into SFT models via simple vector arithmetic. 🧠⚡️

Blog: 👉 https://t.co/BDoKBd2HGt

[Yi Hu @AheadOFpotato]

Unveiling the "Black Box" of Reasoning Models (o1/R1)! 🧠 New Survey by PKU, Tsinghua & NTU on LRM Mechanisms. We dissect:
1️⃣ Training Dynamics (SFT, RL)
2️⃣ Inference Behaviors
3️⃣ Failures (Overthinking/Faithfulness/Hallucination/Safety)
🌟 https://t.co/KhxtF1NDDo

[Mμ Lab @mulabpku]

Do LLMs/LRMs really excel at subtoken reasoning?
⭐️ Check out our SubTokenTest!
Great job by @Shuyoung_Rain and @AheadOFpotato!

[Yi Hu @AheadOFpotato]

https://t.co/Z0WPcCftlx
Beyond "strawberry has how many 'r's": Many real-world tasks like parsing tables and reading maps depend on sub-token understanding.
We introduce SubTokenTest, a practical benchmark that evaluates this often-overlooked capability in real-world scenarios.