Baran Hashemi

1/ Are frontier LLMs the only path to AI math breakthroughs? I think not! We introduce FlowBoost, A lightweight RL+Flow-Matching framework that discovers new extremal geometric structures, beating AlphaEvolve with 100–1000× less compute with zero-shot geometry-aware & reward-guided generation. 🚀 arXiv: arxiv.org/pdf/2601.18005

@givHeisenberg Respekt 🫡

برنامه‌ی حاجی‌تون.💅🏋️‍♀️

Baran Hashemi: FlowBoost Reveals Phase Transitions and Spectral Structure in Finite Free Information Inequalities arxiv.org/abs/2604.11922 arxiv.org/pdf/2604.11922 arxiv.org/html/2604.11922

@givHeisenberg Like a machine 🦾

۴تا ست ۸ تایی ۶۰ کیلویی سوموآردی‌ال🏋️‍♀️

@DayShuai Interesting work. In our last work, we had provided a Flow-Based extremal mathematical structure discovery. It might be a good reference:

What if AI didn’t just solve math problems — but discovered entirely new mathematical structures? Introducing AutoMath from The Omega Institute. From ONE equation (x² = x + 1) and ZERO extra axioms, we derive 9 branches of math: algebra, combinatorics, topology, dynamical systems… all formally verified in Lean 4 (~2,350 theorems, 25k lines of code). Our method: Derive, Discover, Name • Derive → AI exhaustively explores every logical consequence • Discover → Patterns emerge that humans might never notice • Name → Human intuition connects them to deep math (rings, finite fields, golden-ratio p-adics…) This is the future of human-AI mathematical discovery. Read “The Method” 👉 #the-method" target="_blank" rel="nofollow noopener">the-omega-institute.github.io/automath/#the-… GitHub (star & contribute) 👉 github.com/the-omega-inst… What hidden structure will you uncover next? #AutoMath #AI4SCI #Lean4

@DaattavyaA Very interesting work. I enjoyed reading it.

Excited to share our new paper on AI-led maths research. We tried to model mathematical practice realistically, using the task of `rediscovering’ the concept of homology as a benchmark. Would love feedback from people working in a variety of areas! arxiv.org/abs/2603.04528

@pushmeet @GoogleDeepMind Interesting results. We have also provided a flow-based extremal discovery tool that, might be a good reference:

Happy to share new progress in AI for Maths @GoogleDeepMind . In extremal combinatorics, AlphaEvolve has helped establish new lower bounds for FIVE classical Ramsey numbers - a problem so challenging that even Erdős commented on its difficulty. Historically, computationally deriving these bounds required bespoke, human-designed search algorithms. For many of these bounds, the best previous results are at least a decade old. AlphaEvolve changes this by acting as a single meta-algorithm that automatically discovers the search procedures needed to find these new bounds. 📷

@givHeisenberg اینا کیه اطرافتن! گیو جون، شانست چرا اینجوریه آخه…

دختر هندیه بهم پیام داد و از اینکه جنگ به یه کشور محدود نمیمونه ابراز نگرانی کرد. خودم رو کنترل کردم که کسکش جاکشش نکنم.

الان اونجای هری پاتریم که ولدمورت سقط شده. اما مرگ‌خواران ما با قدرت به گه‌خوری ادامه میدن.

@burkov We introduced somewhat the same mechanism for mathematical structure discovery and closed-loop optimisation that is also being called continual learning.

When you train a neural network on a new task using examples (through supervised finetuning), it tends to forget what it already knew — a well-known problem called catastrophic forgetting. The standard fix in reinforcement learning is to have the model learn from its own outputs rather than from a fixed dataset (this is called "on-policy" learning), but that requires a reward function telling the model how good its outputs are, which is often hard to define. This recent MIT paper finds a way around that constraint: it uses the same model twice — once with a demonstration stuffed into its input as context (the 'teacher'), and once without (the 'student') — then has the student generate text, and updates the student's weights so that its token probability distributions get closer to the teacher's token probability distributions at each position in that generated text. The trick works because LLMs can already adapt their behavior when shown an example in context, so the teacher is essentially a better version of the model that stays close to the original, making the learning signal gentle enough to avoid wrecking existing capabilities. Across multiple experiments, this approach lets a single model sequentially learn three different skills while keeping all of them, where standard supervised training destroys earlier skills as soon as it moves to the next one. The math also turns out to be equivalent to a form of reinforcement learning with an implicit reward, which gives the method a clean theoretical grounding beyond just "it works empirically." Read this paper with an AI tutor: chapterpal.com/s/1b7dae70/sel… Read it alone: arxiv.org/pdf/2601.19897

WTH

تاریخ یک کلان‌روایت خطی و غایت‌مند نیست. فهم این نکته ما را از چارچوب‌هایی که تحت تأثیر باورهای جزم‌اندیشانه پیرامون «مسیرِ محتومِ تاریخ» شکل گرفته‌اند، رها می‌کند. علم جاده‌ای یک‌طرفه نیست که اگر مانعی چون «غزالی» نباشد، لزوماً به سرمنزلِ «گالیله» و «نیوتن» ختم شود.

@Erfan_khosravi چقدر زیبا نوشتی عرفان جان

@rronak_ You might want then to check our paper as well for mathematical structure discovery with such self-distillation mechanism as a closed-loop learning:

My favorite paper of 2026 so far 🔥 They took On-Policy Distillation (ie the Thinking Machines blog post), but then showed that the policy can be both the teacher and the student model. The idea is to condition the teacher off of a golden trajectory, and then train on the conditioned logprobs of the same model. The crazy part is, you can literally condition the teacher on anything!! This opens up an entire pandora's box of bridging prompt optimization/ICL + weight optimization that I'm very excited about for continual learning Authors: @IdanShenfeld @MehulDamani2 Jonas Hübotter @pulkitology

@ZimingLiu11 Cool Stuff! We also explored the world models of Transformers for Mathematical physics (enumerative geometry) and investigated how and what it learned. It might be a good reference: arxiv.org/abs/2408.14915

🚨Transformers don't learn Newton's laws? They learn Kepler's laws! Like us, transformers don't predict a flying ball via a differential equation, but by fitting a curve. Moreover, reducing context length steers a transformer from Keplerian to Newtonian. Compression in play.

Flow-based extremal mathematical structure discovery. ~ Gergely Bérczi, Baran Hashemi, Jonas Klüver. arxiv.org/abs/2601.18005… #AI4Math

@cloneofsimo I would like to draw your attention to our non-LLM approach to extremal mathematical structure discovery. I am curious on your thoughts.

This is impressive. LFG.

@littmath @littmath Daniel, I am curious on your thoughts on our non-LLM approach:

Really nice paper. Aside from the nice results resolving a number of open Erdos problems (or finding solutions in the literature), the paper does a really good job contextualizing significance and prior work.

@SebastienBubeck How about non-LLM approaches. I would love to hear your feedback on our recent approach:

This year we will see more and more LLM-aided scientific discoveries. If you want to get up to speed on where we are today, especially in mathematics, check out this video of my talk "Recent Advances in LLMs for Mathematics" youtu.be/MH3lG7V7SuU?si…

1/ Are frontier LLMs the only path to AI math breakthroughs? I think not! We introduce FlowBoost, A lightweight RL+Flow-Matching framework that discovers new extremal geometric structures, beating AlphaEvolve with 100–1000× less compute with zero-shot geometry-aware & reward-guided generation. 🚀 arXiv: arxiv.org/pdf/2601.18005

@__alpoge__ If you are interested in a non-LLM approach to extremal mathematical structure discovery, check out our recent work:

Yay lots of nice people on here and thus getting into the field:D arxiv.org/abs/2601.22401 (and arxiv.org/abs/2601.23245 as well!!)