LLM-Guided Search for Deletion-Correcting Codes
Researchers have utilized an LLM-guided evolutionary search method called FunSearch to discover functions for constructing deletion-correcting codes. This approach successfully identified a function that generates the conjectured-optimal Varshamov-Tenengolts code for single deletions. For multiple deletions and quaternary edit codes, the discovered functions offer improvements over existing methods but are empirical heuristics. The study also explored design choices for LLM-guided search, finding that prioritizing sampling more functions over longer reasoning traces per function was more effective, and that deduplicating logically identical functions was crucial for search diversity. AI
IMPACT Demonstrates LLM-guided search for novel code design, potentially accelerating discovery in information theory.