Siyu Chen

[Siyu Chen @siyuc3141]

How do we get LLMs to self-improve using unstructured documents without relying on human curation or naive "difficulty" heuristics?

Excited to share INFUSER, our latest work that turns self-evolution into influence-guided curriculum learning! 🚀 https://t.co/weDJ9Fjrbw

[Siyu Chen @siyuc3141]

📄 Paper: https://t.co/4vPIXEUdDJ
🤗 HF: https://t.co/EAOvltuMGu
🌐 Blog: https://t.co/GNVLfttZCv
💻 Code: https://t.co/xqLEJZ224i

🫶 Thanks to my amazing collaborators: @XuyuanXiong (co-first), @Hansshao, @Selen7005717917, @JinWeiyang18434, Ping Luo, @zhuoran_yang, @XihuiLiu

[Siyu Chen @siyuc3141]

Looped transformers often save compute by truncating backpropagation to the last few loops.

Is it worth it? We trained ~600 models to find out 👀

Short answer: only if the loop count is fixed. Otherwise, run fewer loops and train them all!

📄 https://t.co/4vPIXEUdDJ
🧵👇

[Siyu Chen @siyuc3141]

🪶 Short on GPU memory? Two loops to the rescue.

No looping is compute-optimal, but 2 fully trained loops achieve the same performance of it with 27–32% fewer non-embedding params and only 8% more training compute!

[Fengzhuo Zhang @FengzhuoZhang]

Beyond faster training, does Muon learn better features than Adam?

🚀 Ans: Yes. Muon learns features that are more robust to input corruptions and transfer better to downstream tasks.

This advantage is reflected in hidden states:

1⃣larger logit margins → stronger robustness

2⃣higher effective rank → richer, more transferable representations

Paper Link: https://t.co/lNWJ80h3a8

A thread 🧵

[Siyu Chen @siyuc3141]

@_Suresh2 Thanks! To clarify: we aren't "reference-free." We use a target domain dev-set (verified and challenging questions) to compute the influence score. Using influence score as rewards "align" the generated questions with the high quality dev-set. Check our generator quality analysis

[Siyu Chen @siyuc3141]

How do we get LLMs to self-improve using unstructured documents without relying on human curation or naive "difficulty" heuristics?

Excited to share INFUSER, our latest work that turns self-evolution into influence-guided curriculum learning! 🚀 https://t.co/weDJ9Fjrbw

[Siyu Chen @siyuc3141]

6/6: Huge thanks to my amazing collaborators: @MiaoLu231342 @beining_wu @HeejuneSheen @FengzhuoZhang @ShuangningLi @zhiyuanli_ @Jose_Blanchet @0920wth @zhuoran_yang

Check out our blog and open-source code!
🔗 Blog: https://t.co/cVJM6YtT4h
💻 Code: https://t.co/Qrf8XMpM5S

[Siyu Chen @siyuc3141]

5/6: The tech (DuGRPO)

Influence rewards are continuous and noisy, so standard GRPO is a poor fit.

We introduce DuGRPO, a dual-normalized GRPO variant that stabilizes generator training by normalizing both within-document and batch-level reward variation.

[Zhuoran Yang @zhuoran_yang]

Followup Paper -- "Neural Networks Provably Learn Group Representations: From Cyclic to Non-Abelian"

Our prior work (https://t.co/uCU4eNr7tH) showed that NNs trained on modular addition learn **sinusoidal features** with precise **phase alignment**. But why sinusoidal? Why those phases?

Because sine and cosine are the **irreducible representations** of the cyclic group Z_p. The network is discovering the representation theory of the group it computes on.

This is not a coincidence. In our new paper, we show that all of these findings can be generalized to Abelian groups, and there is some surprise for non-Abelian groups.

In particular, we study how NNs learn the group composition task: given two group elements g1 and g2, predict g1*g2. The key insight is that NN learns to leverage **group fourier features** to solve this task.

arxiv: https://t.co/WdNIuC9nPj
blog: https://t.co/MUfWOnSICj
code: https://t.co/XUp37Fhjqc

@JLiangHe @LedaW77625 @FengzhuoZhang @siyuc3141

[Fengzhuo Zhang @FengzhuoZhang]

Why do DeepSeek and Kimi use Muon instead of Adam?

🚀 Reasons from a curvature perspective:

1⃣ Under a second-order approx., Muon incurs a much smaller curvature penalty than Adam while maintaining the same first-order decrease.

2⃣ This advantage does not come from a smaller update norm. Instead, it comes from Muon having lower Normalized Directional Sharpness (NDS).

3⃣ Muon’s NDS advantage becomes larger when the training data is more imbalanced.

Paper Link: https://t.co/nGy2FeF538

A thread 🧵

[Zhuoran Yang @zhuoran_yang]

New Paper -- "On the Mechanism and Dynamics of Modular Addition: Fourier Features, Lottery Ticket, and Grokking"

We give a complete mechanistic and dynamic picture of how neural networks learn modular addition f(x,y) = (x+y) mod p. We answer three questions:
(1) What does the trained network compute?
(2) How do Fourier features emerge during training?
(3) Why does grokking happen?
Each answer comes with a mathematical characterization backed by theory and experiments.

Paper: https://t.co/uCU4eNr7tH
Blog: https://t.co/Ok9dDQcYvw
Demo: https://t.co/SX4Q6IsLHB
Code: https://t.co/03Adj6E3Bt

[Fengzhuo Zhang @FengzhuoZhang]

Large Language Models (LLMs) exhibit “slash patterns” in attention maps — a key mechanism behind prefilling acceleration.

We take a first step toward understanding why they emerge.

Main findings:
▶️ Slash patterns are OOD-generalizable
▶️ Queries and keys on these heads are near rank-one and carry little contextual information.
▶️RoPE is the primary source of the slash pattern.

Blog link:
https://t.co/uhE3y7i5xW

A thread 🧵