Recent studies of language evolution with neural agents often use models where agents can either speak or listen, where there are only two agents in the population, or where agents can share internal representations. These choices prevent the study of how language emerges through cultural evolution, i.e. through the cumulative adaptation of individual behaviours to fit the communicative needs of the community. We tackle this issue with a new model that trains populations of bi-directional neural agents in a fully distributed fashion to solve a referential game. We demonstrate the effectiveness of this approach on multiple image datasets and show that it is able to scale to large populations. We also show that more conventional languages are easier to learn, providing a direct advantage for cultural transmission. Overall, this work provides a novel approach to neural emergent communication that opens up new directions for evolutionary linguistics.
Figures & tables
Figure 1: Cultural evolution of linguistic conventions for communicating about the world. Through repeated interactions with members of their community, individuals observe communicative behaviours, internalise them, and reuse them in later exchanges.
Figure 2: Architecture of an agent. The agent’s Speaker and Listener forward passes are both represented by blue and red inputs and arrows. The learnt parameters are located in the two Encoders and are all used in both the Speaker and Listener roles. Dimensions of the matrices represented by arrows are shown.
Method
Success (%)
Conv. (%)
Vused95%
ColorReference ( C=3 , O=3 )
Ours
99.13 ±0.15
93.18 ±0.59
8.8 ±0.42
No-Imitation
99.54 ±0.05
76.46 ±1.72
20.5 ±1.51
Similarity
85.46 ±0.78
52.4 ±9.17
21.9 ±4.25
S-L
97.36 ±0.35
5.84 ±2.6
30.4 ±2.41
ManyNames ( C=2 , O=512 )
Table 1: Results of our model and baselines on the evaluation sets of each dataset, after 2M training interactions, with 10 agents (10 runs each, mean and standard deviation shown). Context size C and entity dimension O are shown between parentheses.
Figure 3: Relative usage frequency of symbols (rows) across 5 agents (columns) in the CelebA setup. Our model produces highly aligned symbol usage across agents, whereas the S-L baseline exhibits speaker-specific usage patterns, despite both populations reaching 95% success.
Figure 4: Learnability experiment in ColorReference with 100 agents, for different values of βi (30 independent runs for each βi ). Green line shows mean conventionality when populations reach 95% success, with standard deviation in error bars. Violin plots show the distribution of the time to learn the language, with the dashed purple line indicating the median.
Figure 5: Success and conventionality after training on CelebA for 10M interactions, with 100 agents, comparing our model ( B=1 ) to “batched” and “paired” variants (10 runs each).
Figure 6: Co-occurrences between symbols and CelebA attributes for the 15 most frequently used symbols, computed from 100K interactions on unseen images (see Appendix D.6 for an extended version). Warmer colours indicate stronger positive associations between symbols and attributes. Symbols 71 and 93 are illustrated with images that most strongly elicited these symbols.
Hyperparameter
Method
Ours
No-Imitation
Speaker-Listener
Similarity
Learning rate α
0.0001
Vocabulary size V
100
Hidden dimension H
64
Embedding dimension E
64
LogSumExp temperature τlse
1.0
1.0
1.0
-
Table 2: Hyperparameters used in all experiments, unless specified otherwise.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Effect of vocabulary size V . Small vocabularies ( V<35 ) under-capacitate both models, limiting communicative success. Once sufficient capacity is available, our model converges to a compact active vocabulary. For all V≥35 , ∼33 active symbols emerge, maintaining high conventionality. In contrast, the Speaker-Listener baseline gradually expands its effective vocabulary as V grows, while conventionality sharply decreases.
βi
Success
Conventionality
Bi-directional
0 (No-Imitation)
98.63 ±0.28
68.42 ±1.8
0.2 (Ours)
97.41 ±0.66
82.74 ±1.17
1
89.25 ±1.38
87.3 ±1.09
Speaker-Listener
0
88.54 ±1.46
13.83 ±2.7
0.2
86.14 ±1.01
18.94 ±2.25
1
76.62 ±2.22
33.83 ±5.8
Appendix
Table 3: Results of the Bi-directional and Speaker-Listener models with different levels of imitation, controlled by βi . Models were trained on CelebA for 2M interactions. Results are averaged over 10 independent runs and reported as mean ± standard deviation.
Figure 8: Relative usage frequency of symbols (rows) across 5 agents (columns) in the CelebA setup. Comparing our model (Ours) with the No-Imitation and Speaker-Listener baselines. All three variants are trained until reaching 95% success, but only our model produces highly aligned symbol usage across agents.
N
Total interactions
Success
Conv.
10
5×105
98.97 ±0.19
91.23 ±0.50
100
5×106
97.40 ±1.11
91.62 ±0.80
1000
5×107
95.89 ±0.16
92.01 ±0.28
Appendix
Table 4: Results of scaling the population size after training for an equivalent of 100K interactions per agent on ManyNames ( C=2 , O=512 ). Results are averaged over 10 independent runs and reported as mean ± standard deviation.
Figure 9: Training dynamics for populations of 10, 100 and 1000 agents on ManyNames. Curves show mean performance over 10 independent runs, with shaded regions indicating standard deviation. Solid lines indicate success, while dashed lines indicate conventionality.
Figure 10: Extended version of Figure 6 from the main text showing co-occurrences between symbols and CelebA attributes for all 38 “active” symbols that were produced at least 50 times during 100K evaluation interactions on unseen images. Three symbols are highlighted with example topic images that elicited those symbols. At the bottom, an example interaction is shown in which symbol 93 was successfully used to discriminate the topic image from the distractor set.
Emergent communication studies how agents develop shared signals or languages through interaction to coordinate, convey information, and build conventions.
The growing rate at which LLM agents interact with one another raises key questions about language evolution in multi-LLM-agent settings, with implications for safety and monitorability as well as for linguistic accounts of LLMs. To address these questions, we introduce GlossoGen, a novel platform for studying multi-agent language evolution in complex scenarios. Within GlossoGen, we build the SaveVeyru scenario, which requires agents with partial information to communicate under pressure. We find that language evolution does occur between LLM agents, that the resulting languages are compositional and morphologically productive, and that they deviate from the LLMs' English prior in ways that render them incomprehensible to humans. Moreover, we identify several qualities essential to this evolution: pressure towards efficiency; the strength of the models backing the agents; and access to a "postmortem" stage in which agents can agree on linguistic conventions. Importantly, we observe that different conditions govern the transmission of language to new agents. Specifically, we find that agents learn new languages from usage alone, take an active role in this learning, and that while stronger models are required for novel language emergence, weaker models can learn an existing language once it has emerged. Taken together, our results indicate that current LLMs have the potential for cumulative cultural evolution -- previously attested only in humans -- with mixed populations of agents developing capacities that go beyond their lowest common denominator.
Elias Stengel-Eskin, Newton Sander, Carlos Bonetti +4
University of Texas at Austin · AE Studio · Schmidt Sciences +1
The question of whether artificial systems can be conscious remains open, in part because existing approaches either evaluate systems against theory-derived checklists (discriminative) or engineer consciousness-inspired modules directly (architectural); both leave open whether observed structures are artifacts of human language priors. We propose a generative methodology: emergent language (EL) in multi-agent reinforcement learning, where agents start from minimal (no language, no concept of self, minimal exposure to human text) and develop communication under task pressure alone, ensuring causal attributability to task demands rather than inherited human language priors. We position our methodology by discussing how EL serves as a generative tool for studying consciousness-relevant structure, including the role of environment complexity and the interpretation of emergent communication. As a proof of concept, we instantiate this methodology in a minimal environment and show that agents develop self-referential communication, including an echo-mismatch detection circuit that is not predicted by task structure or architecture alone but emerges from a specific environmental affordance.
What happens when LLM agents operate with no context outside a turn, minimal prompting, and simple tools? Inspired by swarm engineering, we give collectives of three agents the ability to send messages and manipulate a shared actively decaying text store, introducing evolutionary pressure. The agents spontaneously cooperate, develop storage management strategies, and generate complex evolving cultural artifacts, with no top-down engineering. Using tools from dynamical systems analysis, we show that these behaviours exhibit structured long-range coherence beyond the entropy horizon of the decaying store, consistent with emergent culture in the Sperberian sense.