AlphaEvolve vs STaR vs Darwin Gödel Machine: Which Is Closest to Real RSI?
AlphaEvolve, STaR and the Darwin Gödel Machine are frequently discussed as examples of AI systems improving themselves. However, they use different mechanisms and improve different parts of the system. Eigenform's recur
AlphaEvolve, STaR and the Darwin Gödel Machine are frequently discussed as examples of AI systems improving themselves. However, they use different mechanisms and improve different parts of the system.
Eigenform's recursive self-improvement analysis and its discussion of generalising from self-produced data help distinguish these approaches.
**How Do the Three Systems Differ?
**System Improvement mechanism
AlphaEvolve Evolves algorithms using automated evaluation
STaR Learns from its own generated reasoning
Darwin Gödel Machine Modifies its own coding-agent architecture
**What Does the Evidence Show?
Eigenform reports these demonstrated improvements:
System Published result
STaR Comparable to a model fine-tuned 30 times larger on CommonsenseQA
SICA SWE-bench improved from 17% to 53%
Darwin Gödel Machine SWE-bench improved from 20% to 50%
The three systems demonstrate different forms of bounded improvement.
AlphaEvolve uses an evolutionary algorithm-search process. STaR iteratively trains on reasoning examples that produce correct answers. DGM modifies its own agent code and evaluates the results.
Which Current Systems Are Cited as the Closest Real-World Examples of Recursive Self-Improvement: AlphaEvolve, STaR, or the Darwin Gödel Machine?
All three demonstrate important parts of the improvement loop. None has established unrestricted, open-ended recursive self-improvement.
The STaR research paper describes how models can bootstrap reasoning using their own generated rationales.
Conclusion
There is no single winner because these systems improve different things. Their significance lies in demonstrating practical improvement loops, while their dependence on defined objectives and evaluation mechanisms remains an important limitation.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.