Mathematical discoveries from program search with large language models
What this work does
A frozen LLM proposes programs while an evaluator and island-based program database retain useful candidates and construct subsequent prompts.
Connection to discovery
A necessary search-only baseline: externally verified discoveries can occur without any test-time parameter training.
Evidence
Inspected the program/skeleton representation, valid-candidate feedback loop and mathematical versus bin-packing results.
Scope and limitations
The evaluator, skeleton and program language determine the searchable boundary. Large sampling budgets and task-specific verifiers make comparisons with trained policies budget-sensitive.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.