Alphabell.
Paper · Architecture and optimizer search

RSI-Opt

RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model TrainingA meta-agent recursively improves the prompts, procedures, and code library of an optimizer agent that monitors and adjusts neural network training dynamics.

RSI-Opt
Figure 1 of the paper, arXiv. Source
The loop

A meta-agent rewrites the policy and code library of an optimizer agent based on evidence from earlier training runs. The improved optimizer agent then manages subsequent model training tasks more effectively.

The loop
RSI-Opt
Paper · arXiv
  1. Meta-agent analyzes past training runs
  2. Evolves optimizer agent policy and code
  3. Optimizer agent manages new training runs
  4. Improves training of subsequent models
  1. Meta-agent analyzes past training runs
  2. Evolves optimizer agent policy and code
  3. Optimizer agent manages new training runs
  4. Improves training of subsequent models
↻ The improved system does the next round, and the loop turns again.

Why it is a road to recursion

Automating the evolution of optimizer agents allows training procedures to accumulate experience and improve recursively across runs.

Evidence

Increases representative raw-space summary z-scores from +10.48 to +14.33 on the evolution suite with Opus 5.5.

optimizer-searchmeta-agentrecursive-improvement

Related loops

More architecture and optimizer search →
Self-training learned optimizers
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
ShinkaEvolve
ProjectGitHub
ShinkaEvolve
  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
  5. repeat
Architecture and optimizer searchSakana AI · 2025
Genesys
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