Genesis: Let Evolution Grow on Its Own in a World Without Fitness Functions

Genesis is a groundbreaking evolutionary computing framework that completely abandons the fitness function relied upon by traditional genetic algorithms, exploring open-ended evolution through constraint-driven mechanisms, relational selection, and dynamic adjustment. This article will discuss its background, core mechanisms, experimental validation, conclusions, and implications, revealing its profound significance for the fields of evolutionary computing and AI.

Traditional evolutionary algorithms rely on fitness functions to define optimization goals—they are efficient but prone to premature convergence of populations to local optima, limiting long-term innovation. In contrast, biological evolution has no predefined fitness function and exhibits open-ended creativity by maintaining survival and reproduction in dynamic environments. Genesis poses the question: Can we achieve sustained evolutionary activity without fitness functions, novelty objectives, or reward shaping, solely through constraint-driven selection mechanisms?

The core design philosophy of Genesis is to maintain survival in a dynamically adjusted constraint space, rather than climbing predefined peaks. Its four key mechanisms include:

1. Metabolic Cost Constraint: Higher genomic complexity leads to higher metabolic costs, balancing complexity and maintenance costs;
2. Pareto Relational Dominance: Using multi-objective Pareto dominance to compare individuals, avoiding the limitations of a single optimization objective;
3. Artificial Immune System (AIS) Diversity Protection: Maintaining a genotype archive and injecting diversity when diversity collapse is detected;
4. Adaptive Constraint Adjustment (CARP): Dynamically adjusting constraint strength to maintain a narrow survival corridor, ensuring evolution continues without dictating direction.

The Genesis team reported 12 independent experiments (10,000 generations each) in their paper published at GECCO 2026, with approximately 58% of experiments successfully maintaining non-zero evolutionary activity. Failure reasons include metabolic overload (uncontrolled genomic complexity without constraint adjustment), neutral drift saturation (population stagnation due to fixed constraints), and dominance monopoly (insufficient diversity protection). Ablation experiments showed that removing CARP or AIS resulted in a failure rate exceeding 90%, verifying the necessity of these mechanisms.

In successful experiments, genomic complexity tended to stabilize rather than grow infinitely, revealing that constraint-driven evolution can maintain activity but cannot guarantee open-ended growth. This stems from limitations in genomic coding capacity, fixed constraint structures, or a lack of dynamic environmental pressure. Future research directions include evolvable genomic alphabets, adaptive constraint structures, co-evolution, and environmental dynamics.

Current AI mostly follows the 'training-optimization' paradigm (defining a loss function and optimizing via gradient descent). Genesis proposes another possibility: intelligence may emerge naturally under appropriate constraints, just as life emerged from a chemical soup. This suggests we should focus more on creating suitable conditions rather than designing specific goals, rather than discarding existing methods.

The Genesis team emphasizes that the research provides empirical boundaries rather than fantasies: the 58% success rate proves that fitness-free evolution is possible but not inevitable, and the complexity plateau reveals the limitations of constraints. Negative results and partial successes help understand possibilities and difficulties.

Project address: https://github.com/gearupsmile/genesis-emergence Relevant paper: Sustained Evolutionary Activity Without Fitness Functions: An Empirical Study of Constraint-Driven Selection (GECCO 2026)