Alex Gu
@minimario1729
Followers
4K
Following
4K
Media
149
Statuses
942
mit phd student (on job market!), llm for math+code / prev intern @ meta, nvidia, aws, jane street / enjoys 🎹✈️⛷️⛵
bay area
Joined March 2020
Lean model generated proofs can be optimized, especially by another model! A really cool work from @AIatMeta FAIR from @minimario1729 et al The main idea is to take a model that is first heavily finetuned on a Lean related dataset (with natural language information), and then
1
1
5
visiting seattle for a bit, let me know if you're around and want to have a chat :)
1
1
33
👥ProofOptimizer is work w/ awesome co-authors at Meta FAIR: Bartosz Piotrowski, @FabianGloeckle, @KaiyuYang4, @aramHmarkosyan We think this direction has a lot of potential. Check out our paper, reach out, and chat with us! 🏠 https://t.co/GXLQLqlV4E 📝
0
3
13
2) Simplified proofs often run faster, with 22/75 Putnam proofs achieving over 50% speedup. In our paper, we also optimize more directly for run-time instead than proof length with even better results!
1
1
8
🔮We discover two downstream effects of proof simplification: better training and faster run-time. 1) Training on simplified proofs can improve generation abilities compared to training on longer proofs
1
1
8
We tried ProofOptimizer on Seed-Prover's IMO 2025 proofs with an increased sampling budget to achieve an average proof length reduction of 49%! The proof checking time of two problems also went down, from 434 seconds -> 363 (P4) and 61 -> 34 (P5) ⚡
2
1
8
🤖Inference: we use a natural iterative shortening algorithm: take a proof, sample 64 times, take the shortest proof, and repeat. This already shows decent results! Check out our shortened proofs: https://t.co/cjuZIIgZRA
1
2
8
⛏️Data: We develop a four-stage pipeline to automatically mine data for proof simplification: 1) high-quality problem collection (Goedel-Pset) 2) proof sketching 3) therorem extraction and filtering (remove with AUTO) 4) proof generation (Goedel-Prover-V2)
1
1
12
Towards mitigating long AI-generated Lean proofs, we provide an end-to-end data, training, and inference recipe for proof simplification. While we only consider proof length in this paper, our methods generalize to other measures as well (e.g. runtime).
1
1
10
✂️Introducing ProofOptimizer: a training and inference recipe for proof shortening! 😰AI-written formal proofs can be long and unreadable: Seed-Prover's proof of IMO '25 P1 is 16x longer in Lean vs. English. Our 7B shortens proofs generated by SoTA models by over 50%! 🧵⬇️
6
35
204
🚨 Attention aspiring PhD students 🚨 Meta / FAIR is looking for candidates for a joint academic/industry PhD! Keywords: AI for Math & Code. LLMs, RL, formal and informal reasoning. You will be co-advised by prof. @Amaury_Hayat from ecole des ponts and yours truly. You'll have
24
119
899
check out cwm!! open weights and inference code, many fun details to ponder from the report, and most importantly a lot of new research directions opened :)
(🧵) Today, we release Meta Code World Model (CWM), a 32-billion-parameter dense LLM that enables novel research on improving code generation through agentic reasoning and planning with world models. https://t.co/BJSUCh2vtg
0
3
44
Thanks to MIT News for covering our vision of AI for code! A lot of progress made, but still a long way to go!
Can AI actually code for us? 🧵 MIT research reveals there’s a "long way to go" due to bottlenecks like assessment, codebase scale, & incorrect retrievals. The work reflects a vision to let humans focus on high-level design while routine work is automated:
1
3
28
postering this work on behalf of awesome coauthors at the ai for math workshop tomorrow :)
Do LLMs truly understand math proofs, or just guess? 🤔Our new study on #IneqMath dives deep into Olympiad-level inequality proofs & reveals a critical gap: LLMs are often good at finding answers, but struggle with rigorous, sound proofs. ➡️ https://t.co/h5f8Qv8Xlv To tackle
0
4
37
come to our ai for math workshop tomorrow it'll be super fun!! 🎉🎉
1
7
65
ai for math workshop papers released, it's a fun batch🚀 https://t.co/9MU4cxWzKc
1
2
19
come hang out at our poster tomorrow (tuesday) at 11am let's have some fun and savor the ai4code hype! 🎉 🥳 📌 East Exhibition Hall A-B #E-605
📢 Excited to share our new paper: Challenges and Paths Towards AI for SWE We discuss: 🛠️ 6 sub-tasks needed for SWE 🤖 9 challenges of today's AI in SWE 🔮 9 future directions to address the challenges w/ collaborators from MIT, Berkeley, Cornell, Stanford, and UPenn ⬇️ (1/n)
2
4
41