LLM-based agents autonomously develop shared alien lexicon
A new framework called Neuro-Symbolic Lexical Discovery (NSLD) has been introduced by researchers, enabling LLM-based agents to autonomously create a shared lexicon of alien terms through a referential game involving out-of-distribution visual references. Each agent utilizes a fixed CLIP vision encoder, a dedicated FAISS vector index, and a text-only LLM. The alien vocabulary identified is linked to natural language through semantic closeness in the embedding space, thereby enriching human vocabulary. Simulations featuring up to twenty agents and ten visual references demonstrate consensus, with convergence dynamics analyzed through three models achieving R² > 0.95.
Key facts
- NSLD framework proposed for lexical discovery in unknown environments
- Populations of LLM-based agents play referential game over out-of-distribution visual referents
- Each agent uses CLIP vision encoder, FAISS vector index, and text-only LLM
- Discovered alien words anchored to natural language via semantic proximity
- Consensus reached in simulations with up to twenty agents and ten visual referents
- Convergence dynamics characterized by three analytical models with R² > 0.95
- Application: planetary or deep-sea exploration where entities have no human language name
- Framework autonomously self-organises shared alien lexicon
Entities
Institutions
- arXiv