Alphabell.
Project · Architecture and optimizer search

ShinkaEvolve

ShinkaEvolve is an open-source framework that uses an ensemble of LLMs as mutation operators in a sample-efficient evolutionary search over code.

The loop

An ensemble of LLMs, chosen per step by a bandit, rewrites candidate programs that a user-defined evaluator scores, and the best programs become parents for the next generation, with novelty rejection sampling to avoid near-duplicates. Applied to AI itself, it evolved a mixture-of-experts load-balancing loss for LLM training and an agent scaffold that improves LLM performance on AIME math problems.

The loop
ShinkaEvolve
Project · GitHub
  1. LLM ensemble rewrites candidate programs
  2. Evaluator scores and selects best programs
  3. Evolved mixture-of-experts load-balancing loss
  4. Loss improves LLM training
  1. LLM ensemble rewrites candidate programs
  2. Evaluator scores and selects best programs
  3. Evolved mixture-of-experts load-balancing loss
  4. Loss improves LLM training
↻ The improved system does the next round, and the loop turns again.

Why it is a road to recursion

The same LLM-driven evolution that improved an LLM training loss and an LLM agent scaffold can be aimed at the models and scaffolds that run the search.

Evidence

Sakana reports the evolved MoE load-balancing loss improved on state-of-the-art losses with 5.81% less inefficient token routing and 1.73% higher task performance on average, and generalized to MoEs with 5 times more active parameters.

evolutionary-searchllm-mutationmixture-of-expertsopen-sourcesakana

Related loops

More architecture and optimizer search →
FoundationarXiv
Self-training learned optimizers
  1. Learned optimizers update other learned optimizers
  2. Training keeps best performing optimizer parameters
  3. Better optimizers train each other faster
  4. repeat
Architecture and optimizer searchMetz et al. (Google Research, Brain Team) · 2021
PaperarXiv
Genesys
  1. Designer agents propose language model architectures
  2. Verifier agents pre-train and evaluate designs
  3. Verified results feed evolutionary population
  4. Language models search for better architectures
  5. repeat
Architecture and optimizer searchCheng, Clark, Richardson (Allen Institute for AI, Dartmouth) · 2025
PaperarXiv
DiscoPOP
  1. GPT-4 proposes preference-optimization losses
  2. Candidate losses fine-tune language models
  3. Evaluation scores feed back into prompt
  4. Best loss aligns other LLMs
  5. repeat
Architecture and optimizer searchSakana AI, FLAIR (Oxford), University of Cambridge · 2024