iiinova — where your digital self keeps evolving

Julien Pourcel

@jul-p.bsky.social

PhD student at INRIA (FLOWERS team) working on LLM4code | Prev. (MVA) ENS ParisSaclay

Julien Pourcel @jul-p.bsky.social · 10/07/2025

🤗 This project wouldn't have been possible without my incredible co-author team, @ccolas.bsky.social & @pyoudeyer.bsky.social #LLM #AI #ProgramSynthesis #ICML2025

030

Julien Pourcel @jul-p.bsky.social · 10/07/2025

I’ll be at ICML next week—let’s chat if you’re interested in self-improving LLMs, program synthesis, ARC, or other related subjects.

110

Julien Pourcel @jul-p.bsky.social · 10/07/2025

Want to learn more? We've made everything public: 📗 Blog Post: julienp.netlify.app/posts/soar/ 🤗 Models (7/14/32/72/123b) & Data: huggingface.co/collections/... 💻 Code: github.com/flowersteam/... 📄 Paper: icml.cc/virtual/2025...

julienp.netlify.app

Self-Improving Language Models for Evolutionary Program Synthesis: A Case Study on ARC-AGI

Article about SOAR paper

110

Julien Pourcel @jul-p.bsky.social · 10/07/2025

🚀 **Broader Impact**: This isn't just about ARC puzzles. SOAR's framework could enhance program synthesis tasks where search-based LLM methods are limited by static model capabilities (FunSearch, AlphaEvolve, … )

110

Julien Pourcel @jul-p.bsky.social · 10/07/2025

🌟 **Test-Time Learning**: Even on new problems, SOAR continues improving by focusing on solutions that work well on the given examples. This enables real-time adaptation to novel challenges.

100

Julien Pourcel @jul-p.bsky.social · 10/07/2025

📈 **Results**: - Qwen-7B model: 6% → 36% accuracy - Qwen-32B model: 13% → 45% accuracy - Mistral-Large-2: 20% -> 46% accuracy - Combined ensemble: 52% on ARC-AGI test set - Outperforms much larger models like o3-mini and Claude-4-Sonnet

110

Julien Pourcel @jul-p.bsky.social · 10/07/2025

🎯 Key Insight: Failed programs aren't useless! Through "hindsight relabeling," SOAR treats each failed program as the *correct* solution to a different (synthetic) problem. This massively expands the training data diversity.

100

Julien Pourcel @jul-p.bsky.social · 10/07/2025

🧠 **The Learning Process**: The system learns TWO skills simultaneously: - **Sampling**: Generate better initial solutions - **Refinement**: Enhance initial solutions We also find that learning both together works better than specializing!

110

Julien Pourcel @jul-p.bsky.social · 10/07/2025

🔄 SOAR doesn't just search harder — it gets SMARTER. It alternates between: - Evolutionary search: LLM samples and refines candidate programs. - Hindsight learning: The model learns from all its search attempts, successes and failures, to fine-tune its skills for the next round.

100

Julien Pourcel @jul-p.bsky.social · 10/07/2025

🔬 Why This Matters? Most coding tasks are too hard for even the best language models to solve in one shot. Traditional search methods help, but they hit a wall because the model’s abilities are fixed. SOAR breaks through this barrier by letting the model improve itself over time

100

Julien Pourcel @jul-p.bsky.social · 10/07/2025

Introducing SOAR 🚀, a self-improving framework for prog synth that alternates between search and learning (accepted to #ICML!) It brings LLMs from just a few percent on ARC-AGI-1 up to 52% We’re releasing the finetuned LLMs, a dataset of 5M generated programs and the code. 🧵

210