
Seed-Prover, Lean-Workbook, RRHF, RFT and MATH-Qwen. Prev @Alibaba_Qwen, Phd at @Tsinghua_Uni
Joined August 2013
- Tweets618
- Following840
- Followers1.7K
- Likes2K
Pinned Tweet
Excited to announce Seed-Prover 1.5 which is trained via large-scale agentic RL with Lean. It proved 580/660 Putnam problems and proved 11/12 in Putnam 2025 within 9 hours. Check details at github.com/ByteDance-Seed/Se…. We will work on autoformalize towards contributing to real math!
Zheng Yuan retweeted
We formalized the proof of the 3D Kakeya conjecture in Lean from the Sticky Kakeya theorem.
Together with the unconditional Sticky Kakeya formalization by Nankai University × ByteDance Seed AI4Math, this gives a complete formal proof with no sorry and no project-specific mathematical axiom.
A great example of how AI can help turn frontier mathematics into machine-verified proofs.
github.com/project-numina/ka…
github.com/M32026/3d-sticky-…
projectnumina.ai/article?slu…
@ProjectNumina @GanjinZero
Introducing EdgeBench, a benchmark designed to study how agents learn from environments over at least 12~72-hour runs. We find that performance follows a log-sigmoid function of environment interaction time with high precision.
EdgeBench is built with three ingredients:
- 🌍 Real & Diverse: 134 real-world tasks across 6 task categories, spanning scientific problems, professional knowledge work, software engineering, optimization, formal math, and games.
- ⏳ Ultra-Long-Horizon: Each task supports 12–72 hours of agent work. Recorded human effort averages 57.2 hours.
- 🔁 Informative Feedback: Agents receive real-world feedback for continuous improvement.
After 38,000 hours of agent runs on EdgeBench, a scaling law for learning from environments emerges:
- 📈 As agents interact with task environments over time, their aggregate performance is precisely fit by a log-sigmoid function.
- 🧠 This phenomenon can be explained by an elegant theory of graph exploration.
We are releasing an initial 51 of the 134 tasks, together with the full evaluation framework, to help advance long-horizon agent research. Check our blog & paper for more findings!
Blog edge-bench.org/
Paper edge-bench.org/paper.pdf
GitHub github.com/ByteDance-Seed/Ed…
Dataset huggingface.co/datasets/Byte…
Details below 👇🧵
Congrats!
Announcing Numina-Lean-Agent: an open-source framework achieving SOTA in formal theorem proving using generic models.
Two major breakthroughs: 🏆 12/12 on Putnam 2025 📜 Formalized the "Effective Brascamp-Lieb inequalities" research paper (>90% AI-generated)
Code: github.com/project-numina/nu…
Demo: demo.projectnumina.ai/
Paper: github.com/project-numina/nu…
8
2,046
I proof riemann hypothesis using native_decide 🤡: import Mathlib
def foo : Bool := withPtrEq True False (fun () => false) fun eq => nomatch cast eq ⟨⟩
theorem f : False := nomatch show foo = true by native_decide
theorem anything : RiemannHypothesis := by
exfalso
exact f
This is the arxiv: arxiv.org/abs/2512.17260
There are more players focus on formal math reasoning. Excited to see what axiom will build
Today, I am launching @axiommathai
At Axiom, we are building a self-improving superintelligent reasoner, starting with an AI mathematician.
3
1,204
Zheng Yuan retweeted
🚀 LongCat-Flash-Chat Launches!
▫️ 560B Total Params | 18.6B-31.3B Dynamic Activation
▫️ Trained on 20T Tokens | 100+ tokens/sec Inference
▫️ High Performance: TerminalBench 39.5 | τ²-Bench 67.7
🔗 Model: huggingface.co/meituan-longc…
💻 Try Now: longcat.ai
The speed of AI improvements will mostly depend on how fast it can be verified by the real world. So MATH could be the easiest subject to be self-improved. Verifying a math research equivalents to compiling its formal version.
We don’t have AI self-improves yet, and when we do it will be a game-changer. With more wisdom now compared to the GPT-4 days, it's obvious that it will not be a “fast takeoff”, but rather extremely gradual across many years, probably a decade.
The first thing to know is that self-improvement, i.e., models training themselves, is not binary. Consider the scenario of GPT-5 training GPT-6, which would be incredible. Would GPT-5 suddenly go from not being able to train GPT-6 at all to training it extremely proficiently? Definitely not. The first GPT-6 training runs would probably be extremely inefficient in time and compute compared to human researchers. And only after many trials, would GPT-5 actually be able to train GPT-6 better than humans.
Second, even if a model could train itself, it would not suddenly get better at all domains. There is a gradient of difficulty in how hard it is to improve oneself in various domains. For example, maybe self-improvement only works at first on domains that we already know how to easily fix in post-training, like basic hallucinations or style. Next would be math and coding, which takes more work but has established methods for improving models. And then at the extreme, you can imagine that there are some tasks that are very hard for self-improvement. For example, the ability to speak Tlingit, a native american language spoken by ~500 people. It will be very hard for the model to self-improve on speaking Tlingit as we don’t have ways of solving low resource languages like this yet except collecting more data which would take time. So because of the gradient of difficulty-of-self-improvement, it will not all happen at once.
Finally, maybe this is controversial but ultimately progress in science is bottlenecked by real-world experiments. Some may believe that reading all biology papers would tell us the cure for cancer, or that reading all ML papers and mastering all of math would allow you to train GPT-10 perfectly. If this were the case, then the people who read the most papers and studied the most theory would be the best AI researchers. But what really happened is that AI (and many other fields) became dominated by ruthlessly empirical researchers, which reflects how much progress is based on real-world experiments rather than raw intelligence. So my point is, although a super smart agent might design 2x or even 5x better experiments than our best human researchers, at the end of the day they still have to wait for experiments to run, which would be an acceleration but not a fast takeoff.
In summary there are many bottlenecks for progress, not just raw intelligence or a self-improvement system. AI will solve many domains but each domain has its own rate of progress. And even the highest intelligence will still require experiments in the real world. So it will be an acceleration and not a fast takeoff, thank you for reading my rant
Maybe solving an open question by LLM is not far.
🔥 @GoogleDeepMind just dropped their "formal conjectures" project - formalizing statements of math's biggest unsolved mysteries in #LeanLang and #Mathlib!
This Google-backed project is a HUGE step toward developing "a much richer dataset of formalized conjectures", valuable for benchmarks and growing the Lean ecosystem.
The project was open sourced today! And you can be part of it! Check it out: github.com/google-deepmind/f…
#LeanProver #FormalMath #AIResearch #GoogleDeepMind
Cool direction!
New HealthBench eval! Very excited we (@OpenAI) are investing in AI for health, a defining use case for AGI. Favorite plot is how the performance-cost frontier has improved over time.
Congrats @rahularoradfs @thekaransinghal & team! Follow them for more exciting work to come
3
712
Single problem proving is not enough
We are excited to share our latest work, "APE-Bench I: Towards File-level Automated Proof Engineering of Formal Math Libraries." In this paper, we introduce a novel paradigm of Automated Proof Engineering (APE), which goes beyond the traditional focus on isolated theorem proving by applying large language models to real-world development and maintenance tasks within Mathlib4. Under this new paradigm, models are expected to generate syntactically correct and semantically aligned code patches based on natural language instructions and pre-edit Lean files.
Inspired by the design of SWE-Bench, we present APE-Bench I, the first benchmark targeting file-level structural edits in formal mathematics. We also develop a scalable verification infrastructure tailored for Lean, and combine compilation with LLM-based semantic evaluation. Finally, our work includes a systematic evaluation of current state-of-the-art models on this challenging new task. The paper has been uploaded (xinhuajian.wordpress.com/wp-…), and all datasets and code will be open-sourced in the near future.
As the starting point of the APE-Bench series, this work lays the groundwork for automated proof engineering. Looking ahead, we plan to extend the benchmark towards multi-file coordination scenarios (APE-Bench II) and autonomous agent systems with task planning and feedback-driven repair capabilities (APE-Bench III), paving the way toward practical, scalable, and engineering-ready formal mathematics powered by LLMs.
5
710
Will be at @ICLR2025, welcome any chat about formal math reasoning
1
10
1,001























