spec-exec
Specification-to-Execution Loop
Translates natural-language goals into runnable ML/agent pipelines automatically.
Survey snapshot / English mirror
Treat self-evolution as a system process that still needs verification.
This is a dated survey snapshot, not a completed systematic review. It organizes 196 display papers, 686 classified GitHub repos, and 97 public discussion signals into working loops while keeping claim limits visible.
Working loops
Five review lenses, not a final field standard.
search
Search Loop
Explores architecture, prompt, code, agent, and hyperparameter spaces to find better designs.
evaluator
Evaluator Loop
Tests, benchmarks, and verifies whether a candidate change is a genuine improvement.
reflection
Reflection Loop
Converts failures into memory, generates revision candidates from feedback signals.
population
Population Loop
Maintains multiple candidates, selects, mutates, and recombines across generations.
Method families
The distribution is a display lens, not a claim of field completeness.
Prompt/Search Optimization
Largest family: LLMs rewriting their own instructions.
68 display papers 34.7%
Reward/RL/Self-Play
Self-generated rewards eliminate human annotation bottlenecks.
51 display papers 26.0%
Code/Self-Modification
Code as the mutable substrate — agents rewrite their own source.
28 display papers 14.3%
Multi-Agent Reflection/Debate
Social pressure as selection mechanism.
16 display papers 8.2%
Memory/Knowledge Evolution
Accumulated experience becomes heritable skill libraries.
16 display papers 8.2%
Web/Tool/Environment Adaptation
Real-world environments as the ultimate fitness function.
13 display papers 6.6%
Evaluation/Safety/Governance
Smallest but most critical: guarding against misevolution.
4 display papers 2.0%