Can LLMs be used to discover interpretable models of human and animal behavior?🤔
Turns out: yes!
Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior.
1/12
Turns out: yes!
Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior.
1/12
February 10, 2025 at 12:21 PM
Can LLMs be used to discover interpretable models of human and animal behavior?🤔
Turns out: yes!
Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior.
1/12
Turns out: yes!
Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior.
1/12
#Google #DeepMind FunSearch: Making new discoveries in mathematical sciences using Large Language Models deepmind.google/discover/blo...
FunSearch: Making new discoveries in mathematical sciences using Large Language Models
We introduce FunSearch, a method for searching for “functions” written in computer code, and find new solutions in mathematics and computer science. FunSearch works by pairing a pre-trained LLM,...
deepmind.google
August 1, 2025 at 4:40 PM
#Google #DeepMind FunSearch: Making new discoveries in mathematical sciences using Large Language Models deepmind.google/discover/blo...
e.g. deepmind's funsearch used an llm to find a previously unknown solution to the cap-set problem, improving a 20-year-old bound. the solution is mechanically verifiable and the previous record is published. you can check both yourself:
www.nature.com/articles/s41...
www.nature.com/articles/s41...
Mathematical discoveries from program search with large language models - Nature
FunSearch makes discoveries in established open problems using large language models by searching for programs describing how to solve a problem, rather than what the solution is.
www.nature.com
August 10, 2026 at 9:38 AM
e.g. deepmind's funsearch used an llm to find a previously unknown solution to the cap-set problem, improving a 20-year-old bound. the solution is mechanically verifiable and the previous record is published. you can check both yourself:
www.nature.com/articles/s41...
www.nature.com/articles/s41...
It's so annoying that people (even within DeepMind) keep saying that OpenAI is ahead "in AI", when DeepMind has models like AlphaFold, AlphaProof, GenCast and GraphCast, FunSearch, etc
AI for Science has such a clearer potential for helping people, but VCs and the media are obsessed by GenAI
AI for Science has such a clearer potential for helping people, but VCs and the media are obsessed by GenAI
September 21, 2024 at 7:15 PM
It's so annoying that people (even within DeepMind) keep saying that OpenAI is ahead "in AI", when DeepMind has models like AlphaFold, AlphaProof, GenCast and GraphCast, FunSearch, etc
AI for Science has such a clearer potential for helping people, but VCs and the media are obsessed by GenAI
AI for Science has such a clearer potential for helping people, but VCs and the media are obsessed by GenAI
At 11:30 I will be presenting our work on using FunSearch (now Alpha evolve) to discover interpretable models of cognition:
3/4
bsky.app/profile/pcas...
3/4
bsky.app/profile/pcas...
Can LLMs be used to discover interpretable models of human and animal behavior?🤔
Turns out: yes!
Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior.
1/12
Turns out: yes!
Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior.
1/12
July 14, 2025 at 3:17 PM
At 11:30 I will be presenting our work on using FunSearch (now Alpha evolve) to discover interpretable models of cognition:
3/4
bsky.app/profile/pcas...
3/4
bsky.app/profile/pcas...
FunSearch is an optimization process that uses LLMs to evolve Python programs. we adapted it to our use-case to evolve parameterized symbolic cognitive models using a bi-level optimization: each program's params are fit on the training data for evaluation.
4/12
4/12
February 10, 2025 at 12:21 PM
FunSearch is an optimization process that uses LLMs to evolve Python programs. we adapted it to our use-case to evolve parameterized symbolic cognitive models using a bi-level optimization: each program's params are fit on the training data for evaluation.
4/12
4/12
@garymarcus.bsky.social I enjoyed your recent debate on doom debates!
How do you feel your substack on FunSearch holds in light of the AlphaEvolve release?
How do you feel your substack on FunSearch holds in light of the AlphaEvolve release?
May 16, 2025 at 11:58 AM
@garymarcus.bsky.social I enjoyed your recent debate on doom debates!
How do you feel your substack on FunSearch holds in light of the AlphaEvolve release?
How do you feel your substack on FunSearch holds in light of the AlphaEvolve release?
Even with LLMs you *could* do verifiable output, though I agree that's not how students use ChatGPT. Something like the FunSearch paradigm for example seems promising to me – get an LLM to prove theorems, but in machine-checkable form, not just prose that you have to check yourself for correctness.
FunSearch: Making new discoveries in mathematical sciences using Large Language Models
We introduce FunSearch, a method for searching for “functions” written in computer code, and find new solutions in mathematics and computer science. FunSearch works by pairing a pre-trained LLM,...
deepmind.google
September 30, 2024 at 4:39 PM
Even with LLMs you *could* do verifiable output, though I agree that's not how students use ChatGPT. Something like the FunSearch paradigm for example seems promising to me – get an LLM to prove theorems, but in machine-checkable form, not just prose that you have to check yourself for correctness.
LLM systems can be thought of as string search engines. And I think it is becoming clearer that some of the strings it can search out include valuable answers that are better than any a human has been able to produce, e.g. FunSearch www.nature.com/articles/s41...
Mathematical discoveries from program search with large language models - Nature
FunSearch makes discoveries in established open problems using large language models by searching for programs describing how to solve a problem, rather than what the solution is.
www.nature.com
July 21, 2025 at 5:43 PM
LLM systems can be thought of as string search engines. And I think it is becoming clearer that some of the strings it can search out include valuable answers that are better than any a human has been able to produce, e.g. FunSearch www.nature.com/articles/s41...
☑️ Robust gains: EvoTune also outperforms FunSearch on test sets probing generalization at budgets up to 22K sampled algorithms, more details in the paper. 10/12
April 26, 2025 at 4:56 PM
☑️ Robust gains: EvoTune also outperforms FunSearch on test sets probing generalization at budgets up to 22K sampled algorithms, more details in the paper. 10/12
Google FunSearch Shows a Path for LLMs to Improve Math Problem Solving Skills https://synthedia.substack.com/p/google-funsearch-shows-a-path-for (more thoughts on FunSearch) #AI #math #discoveries
Google FunSearch Shows a Path for LLMs to Improve Math Problem Solving Skills
Automated code iteration leads to novel ideas
synthedia.substack.com
December 17, 2023 at 3:52 PM
Google FunSearch Shows a Path for LLMs to Improve Math Problem Solving Skills https://synthedia.substack.com/p/google-funsearch-shows-a-path-for (more thoughts on FunSearch) #AI #math #discoveries
Wohoo🥳 Thrilled to announce this paper 📢. We have shown that it is possible to significantly improve the FunSearch method with RL and achieve impressive algorithmic discoveries on challenging NP-complete combinatorial optimization tasks like TSP and bin-packing.
Excited to share our latest work on EvoTune, a novel method integrating LLM-guided evolutionary search and reinforcement learning to accelerate the discovery of algorithms! 1/12🧵
April 26, 2025 at 5:02 PM
Wohoo🥳 Thrilled to announce this paper 📢. We have shown that it is possible to significantly improve the FunSearch method with RL and achieve impressive algorithmic discoveries on challenging NP-complete combinatorial optimization tasks like TSP and bin-packing.
🔎 The challenge: Crafting algorithms traditionally demands extensive human expertise and time. LLM-guided evolutionary search methods like FunSearch have shown impressive results in problems ranging from mathematical discovery to robotics and competitive programming. 2/12
April 26, 2025 at 4:56 PM
🔎 The challenge: Crafting algorithms traditionally demands extensive human expertise and time. LLM-guided evolutionary search methods like FunSearch have shown impressive results in problems ranging from mathematical discovery to robotics and competitive programming. 2/12
FunSearch: Making new discoveries in mathematical sciences using Large Language Models https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/ #AI #math #discoveries
December 17, 2023 at 12:00 AM
FunSearch: Making new discoveries in mathematical sciences using Large Language Models https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/ #AI #math #discoveries
We evaluate EvoTune on three combinatorial optimization problems:
📦 Bin packing
🗺️ Traveling salesman
🧩 Flatpack
📈 Key result: Across all problems and three LLMs, EvoTune discovers programs with higher rewards faster compared to the Funsearch (no finetune) baseline 7/12
📦 Bin packing
🗺️ Traveling salesman
🧩 Flatpack
📈 Key result: Across all problems and three LLMs, EvoTune discovers programs with higher rewards faster compared to the Funsearch (no finetune) baseline 7/12
April 26, 2025 at 4:56 PM
We evaluate EvoTune on three combinatorial optimization problems:
📦 Bin packing
🗺️ Traveling salesman
🧩 Flatpack
📈 Key result: Across all problems and three LLMs, EvoTune discovers programs with higher rewards faster compared to the Funsearch (no finetune) baseline 7/12
📦 Bin packing
🗺️ Traveling salesman
🧩 Flatpack
📈 Key result: Across all problems and three LLMs, EvoTune discovers programs with higher rewards faster compared to the Funsearch (no finetune) baseline 7/12
#google #deepmind FunSearch: Making new discoveries in mathematical sciences using Large Language Models https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/
August 1, 2025 at 4:41 PM
#google #deepmind FunSearch: Making new discoveries in mathematical sciences using Large Language Models https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/
This release inspired me! I went and read FunSearch, what an awesome and simple idea.
I implemented something inspired by it to optimise agents within my jax "swarm" environment. It works with only 3x1080 Ti running Qwen3-14B, amazing!
Repo github.com/JakeForsey/s...
Best github.com/JakeForsey/s...
I implemented something inspired by it to optimise agents within my jax "swarm" environment. It works with only 3x1080 Ti running Qwen3-14B, amazing!
Repo github.com/JakeForsey/s...
Best github.com/JakeForsey/s...
May 21, 2025 at 8:29 AM
This release inspired me! I went and read FunSearch, what an awesome and simple idea.
I implemented something inspired by it to optimise agents within my jax "swarm" environment. It works with only 3x1080 Ti running Qwen3-14B, amazing!
Repo github.com/JakeForsey/s...
Best github.com/JakeForsey/s...
I implemented something inspired by it to optimise agents within my jax "swarm" environment. It works with only 3x1080 Ti running Qwen3-14B, amazing!
Repo github.com/JakeForsey/s...
Best github.com/JakeForsey/s...
We review the literature to show which aspects of a compositional framework are missing in major existing approaches including evolutionary algorithms, autocurricula (POET, XLand), & LLM guided search (Funsearch, AlphaEvolve), interactive LLM skills (Voyager); as well as extended related work.
4/n
4/n
June 17, 2026 at 3:24 PM
We review the literature to show which aspects of a compositional framework are missing in major existing approaches including evolutionary algorithms, autocurricula (POET, XLand), & LLM guided search (Funsearch, AlphaEvolve), interactive LLM skills (Voyager); as well as extended related work.
4/n
4/n
Podcast CB SyR 443: Betelgeuse y (319) Leona, galaxia M33, Pequeña Nube de Magallanes, contaminación lumínica, Inteligencia artificial en lingüística y macrodatos, y FunSearch de Deepmind en matemáticas francis.naukas.com/2023/12/23/p...
December 23, 2023 at 9:11 PM
Podcast CB SyR 443: Betelgeuse y (319) Leona, galaxia M33, Pequeña Nube de Magallanes, contaminación lumínica, Inteligencia artificial en lingüística y macrodatos, y FunSearch de Deepmind en matemáticas francis.naukas.com/2023/12/23/p...
Every tech designer I know who in generative works in hope of coming up with something that a human cannot; this, while rare, emerges into the generally probable when you combine our curiosity with its speed.
We shouldn't let speculation drive these things, on this, we are agreed. Yet it do.
We shouldn't let speculation drive these things, on this, we are agreed. Yet it do.
FunSearch: Making new discoveries in mathematical sciences using Large Language Models
We introduce FunSearch, a method for searching for “functions” written in computer code, and find new solutions in mathematics and computer science. FunSearch works by pairing a pre-trained LLM,...
deepmind.google
July 2, 2025 at 5:27 PM
Every tech designer I know who in generative works in hope of coming up with something that a human cannot; this, while rare, emerges into the generally probable when you combine our curiosity with its speed.
We shouldn't let speculation drive these things, on this, we are agreed. Yet it do.
We shouldn't let speculation drive these things, on this, we are agreed. Yet it do.
Our approach builds upon FunSearch, a recently proposed evolutionary framework utilizing large language models to propose iterative improvements to model hypotheses.
February 7, 2025 at 6:54 AM
Our approach builds upon FunSearch, a recently proposed evolutionary framework utilizing large language models to propose iterative improvements to model hypotheses.
In recent years, Google DeepMind has given us more than chess games and Go.
They’ve given us AlphaFold, GNoME, RLAS, FunSearch and AlphaGeometry.
They’ve given us AlphaFold, GNoME, RLAS, FunSearch and AlphaGeometry.
November 30, 2024 at 4:30 AM
In recent years, Google DeepMind has given us more than chess games and Go.
They’ve given us AlphaFold, GNoME, RLAS, FunSearch and AlphaGeometry.
They’ve given us AlphaFold, GNoME, RLAS, FunSearch and AlphaGeometry.