Algorithms

Choose a method by the resource you want to improve. FlashRSI supplies versioned resources, rollout, grading, and records; each method keeps its own proposal, selection, and memory policy.

Instructions and skills

Method	What it changes	How it searches	Settings / example
OPRO	Instructions	Proposes from scored history; returns the best candidate.	OPRO YAML
GEPA	Instructions	Reflects on trajectories and screens a candidate pool.	GEPA YAML
EvoPrompt	Instructions	Uses GA or DE population search on dev scores.	EvoPrompt YAML
ACE	Instructions / playbook	Accumulates feedback and reflections into a playbook.	ACE guide · YAML
SkillOpt	Skills	Uses fast/slow updates and adopts strict dev gains.	Quickstart YAML
Evo-Harness	Skills	Accumulates task-stream skills through native skill discovery.	Evo-Harness guide · YAML

Harness resources

These methods edit a broader resource scope. Start from their own configurations, including an Agent Worker capable of file edits.

Method	How it searches	Settings / example
AHE	Successive edits with diagnosis and edit audits.	AHE guide · YAML
ADAS	A design archive; retains proposals and selects the best.	ADAS guide · YAML
DarwinX	Branches and recombination with repeated outcome checks.	DarwinX guide · YAML
Meta-Harness	File-based history of code, scores, and traces.	Meta-Harness guide · YAML
ModularRSI	Independent module branches followed by integration.	ModularRSI guide · YAML
RRSI	History-guided revisions with score/cost admission.	RRSI guide · YAML

Match the Worker to the method

• Use algorithm.worker.type: llm when the method requests textual proposals.
• For an Agent Worker, set type: agent and name: miniagent, claude_code, or codex, with independent llm settings.
• Evo-Harness needs directory-based native skill discovery. ModularRSI needs explicit module mappings. Follow the linked method guide before switching harnesses.

Interpret results

The twelve implementations adapt methods to FlashRSI's resource and evaluation interfaces. Registration is a configuration entry point; consult each guide's validation and migration notes for differences from the original method and tested paths.

The paper's overall experiment table will be added when its final matrix is available. No aggregate result is claimed here.

To implement your own search policy, see Add an algorithm.