Abstract
Machine intelligence marks the ultimate dream of making machines' intelligence comparable to human beings. While recent progress in Large Language Models (LLMs) shows substantial specific skills for a wide array of downstream tasks, they often fall short in general intelligence. Following the correlation between intelligence and System 2 reasoning (slow thinking), we propose the Evolutionary Reasoning Optimization (ERO) framework. ERO performs "survival of the fittest" over a population of LLMs to search for individuals with strong reasoning ability. Our experiments on representative test suites demonstrate two key findings: i) frontier LLMs like GPT-5 still exhibit limited System 2 reasoning; ii) through the ERO evolution loop, a relatively weak model (Qwen-7B) can be enhanced to emerge powerful reasoning capabilities, outperforming much larger models.
Introduction
The quest for Artificial General Intelligence (AGI) has been revitalized by the success of Large Language Models. However, a critical gap remains between "System 1" thinking—fast, pattern-based recognition—and "System 2" thinking—slow, deliberate, multi-step logical reasoning. Current LLMs are typically trained on static datasets, lacking the open-ended environmental pressures that drove the biological evolution of human intelligence.
This paper investigates whether machine intelligence can be evolved to acquire abstract reasoning abilities similar to humans. We introduce ERO to bridge this gap by treating LLM parameters as a genotype subject to evolutionary selection.
Methodology: Evolutionary Reasoning Optimization (ERO)
The ERO framework treats the entire set of an LLM's neural network parameters as a "genotype" and optimizes them through an evolutionary strategy.
Evolutionary Strategy
We employ a $(\mu + \lambda)$ Evolutionary Strategy (ES). A population of LLMs is initialized, and in each generation, the best-performing "elites" are selected. These elites "reproduce" by sampling new individuals from a Gaussian distribution centered around their parameters.
Technical Optimizations
To manage the massive parameter space of modern LLMs, ERO incorporates several innovations:
- Layer-wise Variance Scaling: Variance is calculated at the layer level rather than the parameter level to stabilize the search space.
- Island Architecture: The population is distributed across "islands" (GPUs) to maintain diversity and prevent premature convergence.
- Distributed Acceleration: Utilizing the Ray library and custom caching, the framework swaps model weights on-the-fly, enabling large-scale evolution on constrained hardware.
Fitness Evaluation
Models are scored based on the Levenshtein distance between their generated output and the ground truth for a given reasoning task, providing a scalar fitness value for selection.
Experimental Setup
The primary benchmark used is the Abstraction and Reasoning Corpus (ARC), designed to measure general intelligence through visual-logical grid puzzles. We selected 15 representative tasks (T1–T15) that test core human priors such as object cohesion and basic geometry.
Results
The empirical results provide strong evidence for the efficacy of evolutionary optimization in unlocking reasoning.
Performance Emergence
The evolved Qwen-7B model demonstrated a significant performance trajectory, with its average reasoning score reaching 0.862 over 12 generations. This represents a dramatic improvement from its initial baseline.
Comparative Analysis
In head-to-head comparisons on the 15 ARC tasks, the ERO-enhanced Qwen-7B outperformed several frontier models:
- Qwen-7B (Evolved): 0.862
- GPT-5: 0.719
- GPT-4o: 0.672
- GPT-4o-mini: 0.565
Qualitatively, the evolved model transitioned from producing nonsensical outputs to "crystal correct" logical solutions in tasks where it previously failed (e.g., Tasks T6, T8, T11).
Discussion
Knowledge Redundancy and Latent Reasoning
A key insight from this work is that reasoning ability may already be latent within pre-trained LLM parameters. ERO does not necessarily "teach" new information but rather reconfigures the parameter space to activate logical pathways.
Scaling Law Defiance
Interestingly, the 7B parameter model often outperformed the 32B version of the same architecture on these tasks. This suggests that for complex System 2 reasoning, evolutionary adaptation can be more critical than raw parameter scaling.
Limitations
Currently, ERO adapts a model to a specific reasoning task. The next frontier is "meta-evolution" — evolving models across a broad distribution of tasks to achieve generalized System 2 reasoning that transfers across domains.
Conclusion
This paper provides an empirical proof-of-concept for Evolutionary Reasoning Optimization. By applying evolutionary principles to LLM parameters, we demonstrated that even smaller models can surpass the reasoning capabilities of the largest frontier models. This suggests a promising path toward closing the System 2 gap in artificial intelligence.
Summary updated