Neuralese
Emergent, task-optimized communication codes learned by neural agents—often discrete symbol systems—that function like a language for coordination but are not designed for human interpretability.
In AI, the term refers to the learned message protocols developed by communicating neural agents in RL (Reinforcement Learning) or MARL (Multi-Agent Reinforcement Learning) settings, where differentiable or discretized channels (e.g., Gumbel-Softmax or straight-through estimators over symbol vocabularies) enable agents to discover signaling schemes that optimize reward; these codes exhibit properties studied in game-theoretic signaling (e.g., Lewis signaling equilibria), information bottleneck trade-offs, and compositionality, and are analyzed via probing, topographic similarity, referential games, and translation models to natural language for interpretability and oversight. Practically, neuralese underpins coordination in multi-robot systems, emergent tool use, and decentralized control, while theoretically it serves as a testbed for understanding how structured communication arises under inductive biases, pressures for generalization, and environmental symmetries; current research targets aligning these emergent languages with human semantics (e.g., via auxiliary supervision, pragmatic reasoning, or translation bridges) and mitigating risks such as covert channels and reward-hacking through uninterpretable codes.
First use is commonly traced to 2017 in work explicitly framing “neuralese” and its translation, with popularity peaking circa 2017–2019 alongside the surge in emergent communication research in MARL (Multi-Agent Reinforcement Learning).
Key contributors include Jacob Andreas, Dan Klein, and Sergey Levine (who formalized and popularized the “Translating Neuralese” framing); Jakob N. Foerster and colleagues (deep differentiable communication in MARL); Igor Mordatch and Pieter Abbeel (emergent communication for embodied agents); and Angeliki Lazaridou and Marco Baroni (referential games and compositionality), with broader contributions from DeepMind and academic groups advancing evaluation metrics, grounding, and natural language alignment.
