AI agents invent alien languages using NSLD framework for exploration

A team of researchers (Sendra-Arranz et al.) introduced the Neuro-Symbolic Lexical Discovery (NSLD) framework, enabling populations of autonomous agents to invent shared vocabularies for entities never before named in any human language. The system is critical for deep-sea or planetary exploration, where agents must refer to unknown phenomena. Each agent uses a frozen CLIP vision encoder to perceive visual referents, a private FAISS vector index for memory, and a text-only LLM for reasoning. Through a referential game—a classic communication task—agents self-organize a shared 'alien' lexicon by aligning their private embeddings via semantic proximity in the CLIP space.

In simulations with up to 20 agents and 10 out-of-distribution visual referents, the framework achieved reliable consensus. Convergence dynamics were characterized by three analytical models, all with R² > 0.95, providing a theoretical basis for pre-deployment planning in autonomous missions. The discovered words are perceptually grounded yet remain anchorable to natural language, effectively extending human vocabulary. This work bridges multi-agent systems, neuro-symbolic AI, and lexical semantics, offering a practical approach for exploration teams that cannot rely on predefined dictionaries.