Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 64 of the 68 references it cites were found in a registry or an index; 42 of the 65 sources its claims cite were checked for support.

Towards Human-Like Emergent Communication via Utility, Informativeness, and Complexity

Open access
Tucker, Mycal, Shah, Julie, Levy, Roger, Zaslavsky, Noga · 2025 · Open Mind

Abstract

Two prominent, yet contrasting, theoretical views are available to characterize the underlying drivers of language evolution: on the one hand, task-specific utility maximization; on the other hand, task-agnostic communicative efficiency. The latter has recently been grounded in an information-theoretic tradeoff between communicative complexity and informativeness, known as the Information Bottleneck (IB) principle. Here, we integrate these two views and propose an information-constrained emergent communication framework that trades off utility, informativeness, and complexity. To train agents within our framework, we develop a method, called Vector-Quantized Variational Information Bottleneck (VQ-VIB), that allows agents to interact using information-constrained discrete communication embedded in a continuous vector space. We test this approach in three domains and show that pressure for informativeness facilitates faster learning and better generalization to novel domains. At the same time, limiting complexity yields better alignment with actual human languages. Lastly, we find that VQ-VIB outperforms previously proposed emergent communication methods; we posit that this is due to the semantically-meaningful communication embedding space that VQ-VIB affords. Overall, our work demonstrates the role of cognitively-motivated optimality principles in inducing aspects of human-like communication among artificial agents.

Overview

CoChat Summary

TL;DR

The paper integrates utility maximization with the Information Bottleneck tradeoff between informativeness and complexity into an information-constrained emergent communication framework, trained via a new method called VQ-VIB. Across color naming, open-domain image naming, and 2D navigation, informativeness pressure improved learning and generalization, while complexity limits produced more human-like communication.

Key findings

  • In the color reference game, VQ-VIB_N was near the IB-optimal bound and more human-like than utility-only baselines: human-agent gNID 0.151 (0.00) vs. Onehot 0.42 (0.16), VQ-VAE 0.38 (0.15), and IB-optimal 0.18 (0.10); efficiency loss was 0.024 (0.00) vs. 0.06 (0.01) and 0.08 (0.01).
  • In ManyNames open-domain communication, increasing informativeness improved out-of-distribution generalization; for C = 32, VQ-VIB_N with n = 4 reached up to roughly 60% OOD accuracy, while Onehot and Proto peaked at approximately 17% and 13%.
  • Limiting complexity improved alignment with human language: relative representation alignment with GloVe peaked when EC informativeness matched English response informativeness, and increasing n lowered distortion but worsened alignment.
  • In 2D navigation, increasing λ_I led to faster convergence and higher reward, increasing λ_C led to less complex communication and coarser spatial discretization, and VQ-VIB achieved greater utility for the same complexity than other architectures.

Methods

  • The authors propose the ICEC objective trading off utility, informativeness, and complexity, and implement VQ-VIB (VQ-VIB_N and VQ-VIB_C) using stochastic embedding/quantization, learnable codebooks, variational complexity bounds, and an entropy tie-breaking loss.
  • Color experiment: speaker observed noisy target colors and listener selected target vs. distractor; λ_U = 1.0, λ_I = 1.0, then λ_C was increased after 40 epochs for 50 epochs; results compared to WCS languages and IB-optimal systems over five runs.
  • ManyNames experiment: 25,000 images with roughly 36 English responses each, 442 topnames, 20% of topnames used for training and semantically distinct test images held out; ResNet 512-d features plus VAE noise; C = 2 during training, evaluated at C = 16 and 32; λ_U = 1, λ_C = 0.01, varied λ_I. 2D navigation used MADDPG with speaker observing target location, listener observing only its own location, reward as negative Euclidean distance, and varied λ_I or annealed λ_C.

Limitations

  • Stated in the paper: it is unclear how human speakers and listeners could estimate informativeness without direct access to others’ mental states, motivating future Theory of Mind work; agents were trained with a single utility function, leaving multi-task utility as future work; and rigorously defining how “meaningful” or “semantic” an EC space is remains challenging.
  • Evident from the text: the experiments are simulated reference games and a simplified 2D navigation world, not direct tests of human language evolution; alignment with human language is measured indirectly through WCS color data, GloVe embeddings, and English naming responses.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1162/opmi_a_00188

Cites

64 of 68 references verified1 citation errors · 3 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1. In our driving example, a passenger wishes to communicate with a driver about driving straight. High utility is achieved if the steering wheel’s angle is within the gray region. Focusing on systems with two words (yellow angles map to word w1, purple angles map to word w2) illustrates the tension between maximizing utility and maximizing informativeness. (a) Utility-maximizing communication divides angles according to whether they are within the gray region. As θ → 0, this will yield non-informative communication that does not generalize to other tasks. (b–c) Maximally informative communication uses the two signals to divide the circle by half, yielding maximal reconstruction accuracy of the angle. However, this strategy does not distinguish between a system with sub-optimal but still non-trivial utility (b) and a system that leads to the worst-case utility in this task (c).
Figure 1. In our driving example, a passenger wishes to communicate with a driver about driving straight. High utility is achieved if the steering wheel’s angle is within the gray region. Focusing on systems with two words (yellow angles map to word w1, purple angles map to word w2) illustrates the tension between maximizing utility and maximizing informativeness. (a) Utility-maximizing communication divides angles according to whether they are within the gray region. As θ → 0, this will yield non-informative communication that does not generalize to other tasks. (b–c) Maximally informative communication uses the two signals to divide the circle by half, yielding maximal reconstruction accuracy of the angle. However, this strategy does not distinguish between a system with sub-optimal but still non-trivial utility (b) and a system that leads to the worst-case utility in this task (c).CC BY · PubMed Central
Figure 2. (a) A general communication model in which a speaker (S) and a listener (L) learn to communicate by trading off utility, informativeness, and complexity (see main text for further detail). (b) The three tradeoff parameters, (λU, λI, λC), control a variety of important communicative behaviors that can be mapped on a simplex. The region in the middle of the simplex, rather than the extremes, is expected to capture human-like communication.
Figure 2. (a) A general communication model in which a speaker (S) and a listener (L) learn to communicate by trading off utility, informativeness, and complexity (see main text for further detail). (b) The three tradeoff parameters, (λU, λI, λC), control a variety of important communicative behaviors that can be mapped on a simplex. The region in the middle of the simplex, rather than the extremes, is expected to capture human-like communication.CC BY · PubMed Central
Figure 3. a) VQ-VIB employs a stochastic embedder (Pθ) that samples a continuous embedding, z, based on parameters extracted from a meaning, m, and a stochastic quantizer (Qψ) that maps z to a discrete word, w, using a codebook, ζ. The neural net mapping from m to θ, and the codebook ζ are learnable parameters. b) The two-stage process first maps from a potentially high-dimensional input space into an embedding space, where discrete vectors, ζ, are used for quantization. In this work, we propose two implementations of VQ-VIB by providing different Pθ and Qψ.
Figure 3. a) VQ-VIB employs a stochastic embedder (Pθ) that samples a continuous embedding, z, based on parameters extracted from a meaning, m, and a stochastic quantizer (Qψ) that maps z to a discrete word, w, using a codebook, ζ. The neural net mapping from m to θ, and the codebook ζ are learnable parameters. b) The two-stage process first maps from a potentially high-dimensional input space into an embedding space, where discrete vectors, ζ, are used for quantization. In this work, we propose two implementations of VQ-VIB by providing different Pθ and Qψ.CC BY · PubMed Central
Figure 4. We conducted experiments in 3 domains: (a) a color-naming reference game, (b) a reference game of natural images, accompanied by human-provided names, and (c) a 2D navigation environment.
Figure 4. We conducted experiments in 3 domains: (a) a color-naming reference game, (b) a reference game of natural images, accompanied by human-provided names, and (c) a 2D navigation environment.CC BY · PubMed Central
Figure 5. Reference game setup for our color experiment. Given a noisily-observed target color, the speaker communicated to a listener, which predicted, from a set of the target and a distractor color, the target. Utility was based on team accuracy.
Figure 5. Reference game setup for our color experiment. Given a noisily-observed target color, the speaker communicated to a listener, which predicted, from a set of the target and a distractor color, the target. Utility was based on team accuracy.CC BY · PubMed Central
Figure 6. Color reference game results. Within the ICEC framework, we annealed λC to generate a spectrum of near-IB-optimal naming systems, closely matching human-like diversity (a). Within the range of human-like complexities, VQ-VIB𝒩 tended to achieve greater utility, for the same complexity, as other methods (b). At low (c–d) and high (e–f) complexity levels, VQ-VIB𝒩 used fewer or more distinct tokens for colors, and closely aligned with human languages at similar complexity levels. Notably, the tokens were embedded in a semantically-meaningful communication space, as visualized using 2D PCA (c, e).
Figure 6. Color reference game results. Within the ICEC framework, we annealed λC to generate a spectrum of near-IB-optimal naming systems, closely matching human-like diversity (a). Within the range of human-like complexities, VQ-VIB𝒩 tended to achieve greater utility, for the same complexity, as other methods (b). At low (c–d) and high (e–f) complexity levels, VQ-VIB𝒩 used fewer or more distinct tokens for colors, and closely aligned with human languages at similar complexity levels. Notably, the tokens were embedded in a semantically-meaningful communication space, as visualized using 2D PCA (c, e).CC BY · PubMed Central
Figure 7. Reference game setup for ManyNames domain. The speaker observed a randomly-drawn candidate image and communicated to a listener, which decoded the speaker’s meaning and predicted which candidate image was the target.
Figure 7. Reference game setup for ManyNames domain. The speaker observed a randomly-drawn candidate image and communicated to a listener, which decoded the speaker’s meaning and predicted which candidate image was the target.CC BY · PubMed Central
Figure 8. Team accuracy on OOD inputs as a function of λI (a) or informativeness (b). By increasing λI, we increased the communicative informativeness for all models, which in turn increased utility on OOD inputs.
Figure 8. Team accuracy on OOD inputs as a function of λI (a) or informativeness (b). By increasing λI, we increased the communicative informativeness for all models, which in turn increased utility on OOD inputs.CC BY · PubMed Central
Figure 9. Visualization of VQ-VIB𝒩 communication spaces at low (a) or high (b) informativeness. Each point represents the 2D PCA projection of the EC token for training inputs (blue) or OOD inputs (red). At low informativeness, there were few commonly-used tokens, and the meanings associated with each token was highly stochastic. At higher informativeness, distinct images were more often encoded differently, and the communication space was semantically-interpretable, which often generalized to OOD inputs.
Figure 9. Visualization of VQ-VIB𝒩 communication spaces at low (a) or high (b) informativeness. Each point represents the 2D PCA projection of the EC token for training inputs (blue) or OOD inputs (red). At low informativeness, there were few commonly-used tokens, and the meanings associated with each token was highly stochastic. At higher informativeness, distinct images were more often encoded differently, and the communication space was semantically-interpretable, which often generalized to OOD inputs.CC BY · PubMed Central
Figure 10. Functional alignment (a) and visualization of embeddings (b) for GloVe and EC models. Functional alignment improved as distortion decreased until reaching the estimated English informativeness, at which point performance plateaued.
Figure 10. Functional alignment (a) and visualization of embeddings (b) for GloVe and EC models. Functional alignment improved as distortion decreased until reaching the estimated English informativeness, at which point performance plateaued.CC BY · PubMed Central
Figure 11. Relative representation alignment (ρ) between EC and GloVe embedding spaces. VQ-VIB𝒩 models had greater alignment than other models, and alignment peaked when EC model informativeness matched English response informativeness.
Figure 11. Relative representation alignment (ρ) between EC and GloVe embedding spaces. VQ-VIB𝒩 models had greater alignment than other models, and alignment peaked when EC model informativeness matched English response informativeness.CC BY · PubMed Central
Figure 12. Training curves for VQ-VIB𝒩 and VQ-VIB𝒞. During training, mean reward increased for all methods, but increasing λI (different curves) led to higher mean rewards.
Figure 12. Training curves for VQ-VIB𝒩 and VQ-VIB𝒞. During training, mean reward increased for all methods, but increasing λI (different curves) led to higher mean rewards.CC BY · PubMed Central
Figure 13. Varying complexity solutions in the 2D world environments. (a) Utility vs. complexity reflect task-agnostic tradeoffs between referring to specific target locations (high utility) and using less complex communication. Among discrete methods, VQ-VIB methods outperformed other architectures. Visualizations of VQ-VIB𝒞 agent traces at high (c) and low (e) complexity show how agents discretized the space into increasingly large regions. Visualization of the communication for these agents (b and d) show that communication vectors were embedded in a meaningful space: nearby tokens (in communication space) referred to nearby targets (in physical space). A video of communication evolution for VQ-VIB𝒩 shows smooth evolution between high and low complexity solutions.
Figure 13. Varying complexity solutions in the 2D world environments. (a) Utility vs. complexity reflect task-agnostic tradeoffs between referring to specific target locations (high utility) and using less complex communication. Among discrete methods, VQ-VIB methods outperformed other architectures. Visualizations of VQ-VIB𝒞 agent traces at high (c) and low (e) complexity show how agents discretized the space into increasingly large regions. Visualization of the communication for these agents (b and d) show that communication vectors were embedded in a meaningful space: nearby tokens (in communication space) referred to nearby targets (in physical space). A video of communication evolution for VQ-VIB𝒩 shows smooth evolution between high and low complexity solutions.CC BY · PubMed Central
Figure A.1 VQ-VIB𝒩 (a) and VQ-VIB𝒞 (b) speaker architectures for combinatorial codebooks.
Figure A.1 VQ-VIB𝒩 (a) and VQ-VIB𝒞 (b) speaker architectures for combinatorial codebooks.CC BY · PubMed Central
Figure D.1 Informativeness vs. complexity (a) and utility vs. complexity (b) in the color reference game for models trained with REINFORCE. We observed similar trends as when training via backpropagation: all models were similarly IB-optimal (a), and VQ-VIB𝒩 tended to achieve higher utility, for the same complexity, as other models.
Figure D.1 Informativeness vs. complexity (a) and utility vs. complexity (b) in the color reference game for models trained with REINFORCE. We observed similar trends as when training via backpropagation: all models were similarly IB-optimal (a), and VQ-VIB𝒩 tended to achieve higher utility, for the same complexity, as other models.CC BY · PubMed Central
Figure D.2 Informativeness and utility vs. complexity for onehot and VQ-VAE models, for different hardcoded codebook sizes. Without informativeness or complexity pressures, varying the architecture was the only way to indirectly control important model behaviors.
Figure D.2 Informativeness and utility vs. complexity for onehot and VQ-VAE models, for different hardcoded codebook sizes. Without informativeness or complexity pressures, varying the architecture was the only way to indirectly control important model behaviors.CC BY · PubMed Central
Figure E.1 OOD accuracy vs. distortion for VQ-VIB𝒩 and VQ-VIB𝒞 (different columns) for different combinatorial codebook factors (different rows). For all architectures, OOD accuracy as distortion decreased, and using greater n supported lower distortion. Note how increasing n shifted all curves to the left, corresponding to lower MSE.
Figure E.1 OOD accuracy vs. distortion for VQ-VIB𝒩 and VQ-VIB𝒞 (different columns) for different combinatorial codebook factors (different rows). For all architectures, OOD accuracy as distortion decreased, and using greater n supported lower distortion. Note how increasing n shifted all curves to the left, corresponding to lower MSE.CC BY · PubMed Central
Figure E.2 OOD accuracy vs. distortion for onehot and Proto. architectures, and varying C. As in Figure E.1 for VQ-VIB models, decreasing distortion was associated with increased OOD utility.
Figure E.2 OOD accuracy vs. distortion for onehot and Proto. architectures, and varying C. As in Figure E.1 for VQ-VIB models, decreasing distortion was associated with increased OOD utility.CC BY · PubMed Central
Figure E.3 Functional alignment for different architectures for varying amounts of alignment data, N. Increasing N improved alignment slightly, but performance typically plateaued after reaching English response informativeness.
Figure E.3 Functional alignment for different architectures for varying amounts of alignment data, N. Increasing N improved alignment slightly, but performance typically plateaued after reaching English response informativeness.CC BY · PubMed Central
Figure E.4 Relative representation alignment (ρ) between EC and GloVe embedding spaces. Results for VQ-VIB𝒞 models are overlayed over data from Figure 11. VQ-VIB𝒞 exhibits similar, although slightly worse, alignment trends to VQ-VIB𝒩.
Figure E.4 Relative representation alignment (ρ) between EC and GloVe embedding spaces. Results for VQ-VIB𝒞 models are overlayed over data from Figure 11. VQ-VIB𝒞 exhibits similar, although slightly worse, alignment trends to VQ-VIB𝒩.CC BY · PubMed Central
Figure E.5 Relative representation alignment for VQ-VIB𝒩 and VQ-VIB𝒞, for varying combinatorial token settings (n). Increasing n decreased the distortion for both model types (note the changed x axis scale) but worsened alignment.
Figure E.5 Relative representation alignment for VQ-VIB𝒩 and VQ-VIB𝒞, for varying combinatorial token settings (n). Increasing n decreased the distortion for both model types (note the changed x axis scale) but worsened alignment.CC BY · PubMed Central
Figure F.1 Team utility as a function of training episode for different model architectures (different plots) and different λI values (different curves). Increasing λI improved performance for all model types.
Figure F.1 Team utility as a function of training episode for different model architectures (different plots) and different λI values (different curves). Increasing λI improved performance for all model types.CC BY · PubMed Central

Verification

Claims the paper makes

27 Valid · 3 Not valid · 16 Not checked · 155 Own results and interpretations

107 more claims · 5 with their cited sources checked

Create a free account to see all 201 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 47
  • It may state two separate claims as one. 51
  • The quote only partly states this claim. 2
  • Borderline: one of the tests for a substantial claim is unclear. 1

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 59 references

Research works printed but not linked: 3

  1. Chaabouni, 2021, Emergent communication at scale
  2. Kay, 2009, The World Color Survey
  3. Malloy, 2022, Resource-rational cognitive modelling: An information-theoretic approach

Cited by

See all 2

Metadata

Authors
Tucker, M., Shah, J., Levy, R., & Zaslavsky, N.
Affiliations
  • Anthropic (United States) (US)
  • Massachusetts Institute of Technology (US)
  • Department of Brain & Cognitive Sciences, Massachusetts Institute of Technology
  • New York University (US)
Corresponding author
Tucker, Mycal
Venue
Open Mind
Publisher
MIT Press
Published
2025
Volume
9
Pages
418-451
ISSN
2470-2986 (ISSN-L)
OpenAlex
W4409248873
Cited by
2
References
59
Citations by year
2025: 2
FWCI
12.87
Type
journal article
Language
en
Open access
Locations
  • Open Mind (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DSpace@MIT (Massachusetts Institute of Technology) (submittedVersion, cc-by)
Funders
  • Simons Center for the Social Brain, Massachusetts Institute of Technology
  • Quest for Intelligence, Massachusetts Institute of Technology
Topics
  • Social Sciences › Cultural Studies › Language and cultural evolution
  • Physical Sciences › Computer Science › Artificial Intelligence › Evolutionary Algorithms and Applications
  • Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
Keywords
emergent communication, language evolution, information bottleneck theory, informativeness, communicative complexity, utility maximization, variational information bottleneck, vector quantization, artificial agents, principle of optimality

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Towards Human-Like Emergent Communication via Utility, Informativeness, and Complexity. https://app.cochat.ai/papers/10.1162/opmi_a_00188
The paper
Tucker, M., Shah, J., Levy, R., & Zaslavsky, N. (2025). Towards Human-Like Emergent Communication via Utility, Informativeness, and Complexity. Open Mind. https://doi.org/10.1162/opmi_a_00188

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.