![[2602.15725] Recursive Concept Evolution for Compositional Reasoning in Large Language Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.15725] Recursive Concept Evolution for Compositional Reasoning in Large Language Models
Summary
The paper presents Recursive Concept Evolution (RCE), a novel framework that enhances compositional reasoning in large language models by dynamically modifying their internal representation during inference.
Why It Matters
As large language models struggle with compositional reasoning tasks, RCE offers a significant advancement by allowing models to create new abstractions on-the-fly, potentially improving their performance on complex benchmarks. This innovation could lead to more robust AI applications in various fields, including natural language processing and machine learning.
Key Takeaways
- RCE enables models to modify their internal representation geometry during inference.
- The framework dynamically generates low-rank concept subspaces to address representational inadequacies.
- RCE shows significant performance improvements on various compositional reasoning benchmarks.
- The method preserves stability while allowing for the construction of new abstractions.
- Integration with Mistral-7B demonstrates practical applicability and effectiveness.
Computer Science > Artificial Intelligence arXiv:2602.15725 (cs) [Submitted on 17 Feb 2026] Title:Recursive Concept Evolution for Compositional Reasoning in Large Language Models Authors:Sarim Chaudhry View a PDF of the paper titled Recursive Concept Evolution for Compositional Reasoning in Large Language Models, by Sarim Chaudhry View PDF HTML (experimental) Abstract:Large language models achieve strong performance on many complex reasoning tasks, yet their accuracy degrades sharply on benchmarks that require compositional reasoning, including ARC-AGI-2, GPQA, MATH, BBH, and HLE. Existing methods improve reasoning by expanding token-level search through chain-of-thought prompting, self-consistency, or reinforcement learning, but they leave the model's latent representation space fixed. When the required abstraction is not already encoded in this space, performance collapses. We propose Recursive Concept Evolution (RCE), a framework that enables pretrained language models to modify their internal representation geometry during inference. RCE introduces dynamically generated low-rank concept subspaces that are spawned when representational inadequacy is detected, selected through a minimum description length criterion, merged when synergistic, and consolidated via constrained optimization to preserve stability. This process allows the model to construct new abstractions rather than recombining existing ones. We integrate RCE with Mistral-7B and evaluate it across compositio...
