Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Semantics and Spatiality of Emergent Communication
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces semantic consistency and spatial meaningfulness as prerequisites for meaningful emergent communication, proving reconstruction-optimal protocols are semantically consistent while discrimination-optimal protocols need not be.
Problem
Emergent communication protocols can solve discrimination or reconstruction tasks with near-perfect accuracy yet be counterintuitive; the paper asks whether optimal protocols under these common objectives have messages whose equivalence classes contain semantically similar inputs.
Approach
Defines semantic consistency (inputs sharing a message are, in expectation, closer than random pairs, equivalently explained variance is positive) and a stricter spatial meaningfulness that also considers distances between messages; proves reconstruction-optimal senders satisfy these while discrimination-optimal senders need not, since its objective depends only on message class sizes.
Key results
- On Shapes, reconstruction reaches TopSim 0.34 vs 0.09 for discrimination, with discrimination accuracy 61.96% vs reconstruction 31.64%.
- On MNIST, reconstruction achieves 88.00% discrimination accuracy with message variance 371.63 vs random baseline 1029.57, indicating semantic consistency.
- Within reconstruction runs on Shapes, correlation between compositionality and task performance is negative (-0.24), reported as counter-intuitive.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use reconstruction rather than discrimination training | Emergent communication agents · Shapes validation | TopSim: 0.09 { $\pm 0.01$} -> 0.34 { $\pm 0.02$} | Helped |
| use reconstruction rather than discrimination training | Emergent communication agents · Shapes validation | message variance: 2280.24 { $\pm 157.38$} -> 1334.38 { $\pm 78.05$} | Helped |
| use discrimination rather than reconstruction training | Emergent communication agents · Shapes validation | discrimination accuracy: 31.64 { $\pm 2.51$} -> 61.96 { $\pm 4.78$} | Helped |
| use reconstruction rather than discrimination training | Emergent communication agents · MNIST validation | message variance: 1226.14 { $\pm 42.4$} -> 371.63 { $\pm 2.1$} | Helped |
| use reconstruction rather than discrimination training | Emergent communication agents · MNIST validation | discrimination accuracy: 78.76 { $\pm 10.4$} -> 88.00 { $\pm 0.6$} | Helped |
| use reconstruction rather than discrimination training | Emergent communication agents · MNIST validation | TopSim: 0.360 { $\pm 0.036$} -> 0.365 { $\pm 0.042$} | No clear effect |
| add label-aware distractor selection to discrimination training | Discrimination agents · MNIST validation | discrimination accuracy: 78.76 { $\pm 10.4$} -> 87.10 { $\pm 5.4$} | Helped |
| use label-aware distractor selection rather than unsupervised discrimination | Discrimination agents · MNIST validation | message variance: 1226.14 { $\pm 42.4$} -> 1381.72 { $\pm 41.6$} | Hurt |
| use reconstruction rather than supervised discrimination training | Emergent communication agents · MNIST validation | TopSim: 0.269 { $\pm 0.044$} -> 0.365 { $\pm 0.042$} | Helped |
Surprising findings
- Uniformly distributed message assignments can achieve globally optimal discrimination loss regardless of message content.
- On Shapes, discrimination-trained agents used fewer unique messages than reconstruction agents despite theory favoring maximal message diversity.
Limitations
- Theory assumes unrestricted receiver hypothesis classes and optimal solutions, ignoring architecture and optimization error.
- Discrimination results are for the vanilla independent-distractor variant; other variants are analyzed only in the appendix.
- Assumes Euclidean distance in input space reflects semantic similarity.
Experimental setup
Datasets: MNIST Shapes
Benchmarks: Shapes validation set MNIST validation set
Metrics: discrimination accuracy topographic similarity empirical message variance unique messages message purity
Baselines: random message baseline discrimination game supervised discrimination game
Builds on: Lewis signaling games k-means clustering infoNCE Gumbel-Softmax
Closest papers by meaning
- Speaking Your Language: Spatial Relationships in Interpretable Emergent Communication (NeurIPS 2024 · similarity 0.89)
- Bridging semantics and pragmatics in information-theoretic emergent communication (NeurIPS 2024 · similarity 0.88)
- Narrowing the Collaboration Gap, Probably (NeurIPS 2026 · similarity 0.87)
- Lewis's Signaling Game as beta-VAE For Natural Word Lengths and Segments (ICLR 2024 · similarity 0.87)
- Learning to Communicate Through Implicit Communication Channels (ICLR 2025 · similarity 0.87)
- Unsupervised Translation of Emergent Communication (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.