Haowei Lin (@AndyLin2001) profile picture

Undergrad and PhD student @PKU1898 LLM Agent | AI4Science | Generative Models | Continual Learning. Intern @ Tencent Hy agent team

Beijing, China
Joined May 2023
Many people sent me this news today. But I want to say: This ability is already in Claude's system card (twice: Opus 5.5 and Fable/Mythos 5.2)! 😹 When I first read this, I was confused about whether training on sequences would be useful.
Claude was trained on protein sequences, at scale. We gave 20 language models the start of little-studied bacterial proteins and asked for the next amino acids. Only the latest Claude models predict them. Not GPT, not Gemini, not seven open-weight models.
5
368
Excited to release SimpleTES: a better open-sourced AlphaEvolve! With gpt-oss, SimpleTES discovers SOTA solutions across 21 tasks: quantum compilation, LASSO speedup, scaling law discovery, kenel optimization... Project: wizardquant.com/will/simplet… Code: github.com/wq-will/SimpleTES
1
5
22
887
The metareview:
3
2
58
59,472
Winning the award for Highest Scored Rejected Paper at @iclr_conf . The AC is a harsh reviewer from NeurIPS and overrides all positive reviews and ignores every improvement made since NeurIPS. Devastated because this is genuinely my best work yet. openreview.net/forum?id=mnI8…
11
8
1
296
40,900
Spent countless hours carefully reviewing for ICLR (even stepping in as an emergency reviewer), giving both positive and negative scores. To then be hit with this kind of unreasonable criticism is infuriating and deeply disappointing. Judge my reviews for yourself. @iclr_conf
4
1
15
6,638
Impressive results from our latest run: OpenEvolve + GPT-OSS-120B with single-phase evolution reached a score of 2.6359830849. This not only tops AlphaEvolve (2.635863) but is within 1e-8 of ShinkaEvolve results—all without adding any new features. @SakanaAILabs @asankhaya
1
2
2
539
An example of SLDAgent discovered law v.s. human law: conceptually sound and better fitness. This is really a newly discovered knowledge / science.
1
189
This isn't just theoretical! We showed its practical utility in two key applications: 1. Hyperparameter Tuning: deriving near-optimal learning rates and batch sizes for pre-training. 2. Model Selection: Perfectly identifying the best pre-trained LLM for SFT across 3 datasets.
2
57
The results are stunning. Paired with GPT-5, SLDAgent beats the human-expert-derived laws on all tasks in SLDBench. The discovered laws aren't just a better fit; they are more conceptually sound, capturing the underlying dynamics in a more principled way than the human versions.
1
52
Our key innovation is SLDAgent, an evolution-based agent that mimics the scientific process. It starts with a simple guess, then iteratively mutates and refines both the mathematical formula and the method to fit it to data, constantly improving its "hypothesis."
1
50
Our key innovation is SLDAgent, an evolution-based agent that mimics the scientific process. It starts with a simple guess, then iteratively mutates and refines both the mathematical formula and the method to fit it to data, constantly improving its "hypothesis."
75
We built SLDBench, a testbed for this task, curated from over 5,000 real-world LLM experiments across 7 tasks This isn't just about rediscovering known formulas; it's an open-ended challenge where even human experts don't know the "perfect" answer. huggingface.co/datasets/pkuH…
1
1
52
For decades, finding scaling laws has been a slow, manual process for human experts. It's a major bottleneck for developing foundation models, requiring tons of trial-and-error to predict how a model will behave when you spend millions to train it.
2
132