GEPS v5: Evolutionary Portfolio Search FAQs

Question 1

What is the purpose of the Failure Ledger?

Accepted Answer

The Failure Ledger tracks which generation channels fail to produce high-quality ideas and uses Thompson Sampling to adjust channel weights, improving the pipeline's efficiency over time.

Question 2

Can I use different models for generation and judging?

Accepted Answer

Yes, the skill provides a configuration prompt to specify model preferences, such as using Claude Opus for generation and Gemini or GPT-4 for the judge pool.

Question 3

Do I need external Python packages to run this?

Accepted Answer

No, GEPS v5 is designed to run using Python 3.10+ standard libraries only to minimize dependency issues, though it requires a separate llm_runner script for API calls.

Question 4

How does the tournament system work?

Accepted Answer

It employs a FIDE Swiss-style tournament where various LLM judges evaluate pairs of ideas. The results are aggregated using a Bradley-Terry model to provide a statistically sound ranking of finalists.

Question 5

What is the 'Graph-Guided' part of GEPS v5?

Accepted Answer

It uses a literature corpus to build a concept graph and identifies 'structural holes'—areas where connections or ideas are missing—to guide the AI in generating truly novel research concepts.

GEPS v5: Evolutionary Portfolio Search

Key Features

Use Cases

GEPS v5: Evolutionary Portfolio Search

Key Features

Use Cases