The julien31/Soar-qwen-14b is a 14.8 billion parameter Qwen-based language model developed by Julien Pourcel, Cédric Colas, and Pierre-Yves Oudeyer. It is fine-tuned using the SOAR (Self-improving Operators for Automated program Refinements) framework, specializing in program synthesis for the Abstraction and Reasoning Corpus (ARC) tasks. This model excels at generating Python programs to solve complex reasoning problems, outperforming larger closed-source LLMs on ARC-AGI-1 by learning from its own evolutionary search and refinement processes.
Loading preview...
Soar-qwen-14b: Self-Improving Program Synthesis for ARC-AGI
Soar-qwen-14b is a 14.8 billion parameter model based on the Qwen architecture, developed by Julien Pourcel, Cédric Colas, and Pierre-Yves Oudeyer. It is specifically fine-tuned using the SOAR (Self-improving Operators for Automated program Refinements) framework to solve tasks from the challenging Abstraction and Reasoning Corpus (ARC) by synthesizing Python programs. The SOAR framework enables the model to learn from its own experience through a "virtuous cycle" of evolutionary search and learning, significantly outperforming much larger closed-source LLMs on ARC-AGI-1.
Key Capabilities
- Evolutionary Program Synthesis: Utilizes an LLM to generate and intelligently refine thousands of candidate Python programs for ARC tasks.
- Self-Improvement via Hindsight Learning: Learns from both successful and failed program synthesis attempts by "relabeling" failures as correct solutions for synthetic tasks, creating a diverse training dataset.
- Specialized for ARC-AGI: Designed to tackle complex reasoning problems that require discovering solutions from scratch, without relying on human-engineered domain-specific languages.
- Dataset Generation: Contributed to the creation of a 5 million entry dataset of ARC solutions, available as soar_arc_train_5M.
Good For
- Automated Program Generation: Generating Python functions to solve abstract reasoning challenges.
- Research in Self-Improving AI: Exploring frameworks where models bootstrap their own capabilities through iterative learning.
- Benchmarking on ARC: Providing a strong baseline for the Abstraction and Reasoning Corpus, demonstrating performance competitive with or superior to larger models.