Google DeepMind's FunSearch model cracks the unsolved cap set problem in pure mathematics, the first time an LLM has solved a long-standing scientific puzzle
MIT Technology ReviewWill Douglas Heaven
Context & Ripple Effects
The report follows coverage of an earlier model breakthrough that could handle basic mathematics, and moves the benchmark from routine problem solving toward a long-standing research question. It gives Google DeepMind a concrete research result around FunSearch rather than a general capability claim.
Google DeepMind gains a high-visibility validation for FunSearch as a system that can contribute to pure-mathematics discovery, not only generate answers to established exercises.
Mathematicians working on related combinatorial-search problems gain a new AI-assisted route for proposing and testing candidate constructions.
Second-order effects
The result raises the bar for rival AI labs: credible math-reasoning claims increasingly need to show performance on difficult research or competition problems, not just basic arithmetic.
If such results continue to reproduce across fields, AI research tools could become part of the normal discovery workflow, with human experts shifting more effort toward framing problems and verifying machine-generated results.
Mathematical progress may become a more demanding evaluation domain for AI systems: success will depend on independently checkable advances, not merely fluent reasoning traces.
The trend: This is an early data point in the shift from general language models toward specialized AI systems designed to generate verifiable scientific and mathematical discoveries.
Excited for our latest breakthrough published in @Nature today, #FunSearch ('searching' in ‘function’ space) is the first system to make new verifiable discoveries in open problems in math & CS, we applied it to find a more efficient solution to the classic ‘bin packing’ problem
@johnjnay
John Nay on x
Mathematical Discoveries w/ LLMs -Pairing LLM w/ a systematic evaluator -LLM creatively builds upon programs, generates new ones, which are automatically evaluated -Applied to a central combinatorics problem, discovers new constructions of large cap sets https://www.nature.com/..…
@barekatainamin
Amin Barekatain on x
✨ So thrilled to share the work the team and I have been working on this year, which has been published in @Nature! #FunSearch (Search in the Function Space) 🔍 FunSearch has made the *first scientific discoveries* for established open problems using an LLM 🚀 What's more,...
@googledeepmind
@googledeepmind on x
Introducing FunSearch in @Nature: a method using large language models to search for new solutions in mathematics & computer science. 🔍 It pairs the creativity of an LLM with an automated evaluator to guard against hallucinations and incorrect ideas. 🧵 https://dpmd.ai/... [image]
@mfigurnov
Michael Figurnov on x
A super cool project from GDM's Science team! It's the ‘highest surprisal’ project for me in quite some time: I haven't thought LLMs + evolution could do such incredible stuff :) Huge congrats to the team.
@sashavnovikov
Alexander Novikov on x
Excited to share our @nature paper on #FunSearch (Search in Functional space)! tl;dr: we find new results in maths by evolving code with a pretrained (no finetuning involved) medium-size LLM acting as crossover/mutation operator https://deepmind.google/... https://storage.google…
@pengmingwang
Pengming Wang on x
AI won't “solve” math anytime soon, but no doubt human + AI will make many new discoveries
@alhusseinfawzi
Alhussein Fawzi on x
LLMs effectively compress the current state of human knowledge into a model. But can such models be used to go beyond the knowledge they were trained on, and discover new results on open problems? That's what we study in our paper appearing in @Nature. https://deepmind.google/...
@pushmeet
Pushmeet Kohli on x
Can LLMs uncover new knowledge - the scientific equivalent of AlphaGo's move 37? In a Nature paper today, we @GoogleDeepMind unveil that FunSearch, our new LLM based approach for program search, has uncovered new results in Maths and Computing. See details below.
@wtgowers
@wtgowers on x
Here's a great example of human-AI collaboration to obtain new results in mathematics. In this case, a team from DeepMind has used LLMs to obtain the best known lower bound for the cap-set problem. Of course, there's more to it than talking to ChatGPT. https://www.theguardian.com…
@matejbalog
Matej Balog on x
Can LLMs be used to discover something new? Yes! Happy to share our new paper in @Nature on #FunSearch, a system that uses LLMs for making new discoveries in mathematical sciences. Blog: https://deepmind.google/... Paper: https://storage.googleapis.com/ ... 1/n [image]
@ber24
Bernardino Romera on x
Today in @Nature we present FunSearch: an LLM-powered system that has found new discoveries in an established mathematical problem. FunSearch is versatile too: it can be applied to impactful practical problems. Blogpost: https://deepmind.google/... Paper: https://storage.googleap…
@gfodor
@gfodor on x
If DeepMind just definitively proved neural networks can generate genuinely new knowledge then it's the most important discovery since fire. 😅 https://deepmind.google/...
@emollick
Ethan Mollick on x
A highly important open question about AIs has been whether they can actually discover new things. Today, a new Nature paper has shown that LLMs can! A model “discovered new solutions for the cap set problem, a longstanding open problem in mathematics” https://www.nature.com/... …