Shaping Shared Languages: Human and Large Language Models' Inductive Biases in Emergent Communication

TL;DR

This work investigates how the inductive biases of humans and large language models (LLMs) shape emergent referential languages in a classic referential game. It extends prior emergent-communication work by comparing language evolution in Human-Human, LLM-LLM, and Human-LLM pairings, using a 27-meaning space and in-context learning for LLMs. The results show that structured, referentially grounded vocabularies emerge under all conditions, with human-optimized languages being more diverse and efficient, while LLM-optimized vocabularies tend toward degeneracy; crucially, Human-LLM collaboration yields vocabularies that are more human-like, suggesting cross-domain flexibility. The study highlights communicative success as a driving reward signal and argues for training LLMs with human interaction to maintain alignment and grounding in human communication. Together, these findings illuminate how LLM biases influence language dynamics and point to human-in-the-loop training as a pathway to robust human–machine communication.

Abstract

Languages are shaped by the inductive biases of their users. Using a classical referential game, we investigate how artificial languages evolve when optimised for inductive biases in humans and large language models (LLMs) via Human-Human, LLM-LLM and Human-LLM experiments. We show that referentially grounded vocabularies emerge that enable reliable communication in all conditions, even when humans \textit{and} LLMs collaborate. Comparisons between conditions reveal that languages optimised for LLMs subtly differ from those optimised for humans. Interestingly, interactions between humans and LLMs alleviate these differences and result in vocabularies more human-like than LLM-like. These findings advance our understanding of the role inductive biases in LLMs play in the dynamic nature of human language and contribute to maintaining alignment in human and machine communication. In particular, our work underscores the need to think of new LLM training methods that include human interaction and shows that using communicative success as a reward signal can be a fruitful, novel direction.

Figures (5)

• Figure 1: The experimental blocks in our experiment. Participants go through the exposure and guessing block twice before labelling each of the 15 training stimuli in the labelling block. The communication block is performed for 4 rounds each consisting of 30 tasks $T$, where participants alternate speaker-listener roles for each stimulus once. Participants label 27 (15 original and 12 novel) stimuli in the testing block. Image is adopted from kouwenhoven-etal-2025-searching.
• Figure 2: Left: Humans learn the language by being exposed to stimuli and the corresponding signals in the exposure block. Right: LLMs learn the same vocabulary by virtue of in-context learning. A JSON-like structure containing the signal-meaning mappings is prepended to each prompt to serve as learning stimuli.
• Figure 3: A prompt snippet used for labelling and guessing. During communication, we add a communicativeSuccess attribute, update the system prompt to inform about the communicative task, and instruct that 'Communicative success is important'.
• Figure 4: The average communicative performance (PecrCom) per round across the conditions. Communication steadily increases over rounds except for the LLM-LLM condition, in which coordination happens in the first round but does not increase afterwards.
• Figure 5: An overview of structure metrics used to measure compositional structure in the languages produced in the testing block. Generally, languages optimised for LLMs differ from those optimised for humans. Languages optimised for both mediate these differences. The asterisks indicate whether an independent Welch's t-test reveals a significant difference between the conditions where $\ast\ p < .05$, $\ast\!\!\ast\ p < .01$, $\ast\!\!\ast\!\!\ast\ p < .001$. TopSim and WordLength are normalised to values between 0 and 1 for visualisation purposes.