Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Bridging semantics and pragmatics in information-theoretic 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.
Extends the VQ-VIB information-bottleneck emergent communication framework to pragmatic reference games on ManyNames images, showing that a mix of utility, informativeness and complexity pressures yields a lexicon that best aligns with English naming data.
Problem
How a shared human-like lexicon can emerge from local context-sensitive pragmatic interactions among artificial agents, and how semantics and pragmatics co-evolve.
Approach
Trains self-play speaker-listener agents in a pragmatic reference game with a multi-objective IB-style loss (utility, informativeness via MSE, complexity via a variational MI bound) without a given lexicon, then evaluates the emergent lexicon in a lexical semantic naming setting against ManyNames English naming data.
Key results
- Best English-aligned model (lambda_U=0.005, lambda_I=0.98, lambda_C=0.015) reaches NID 0.55 vs English, lexicon size gap 22 and complexity gap 1.48 bits.
- Utility-only model (lambda_U=1) reaches utility 0.95 but NID is poor (>0.8), with lexicon size gap 559.
- Informativeness-only model (lambda_I=1) yields over 1.5K words versus about 400 English topnames and complexity above 6 bits.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use utility maximization only, removing informativeness and complexity pressures (ablation) | VQ-VIB · ManyNames | NID: 0.55 -> 0.84 | Hurt |
| Use utility maximization only, removing informativeness and complexity pressures (ablation) | VQ-VIB · ManyNames | Utility: 0.72 -> 0.95 | Helped |
| Maximize informativeness only, removing utility and complexity pressures (ablation) | VQ-VIB · ManyNames | NID: 0.55 -> 0.60 | Hurt |
| Maximize informativeness only, removing utility and complexity pressures (ablation) | VQ-VIB · ManyNames | MSE: 0.23 -> 0.13 | Helped |
| Minimize complexity only, removing utility and informativeness pressures (ablation) | VQ-VIB · ManyNames | Utility: 0.72 -> 0.58 | Hurt |
| Minimize complexity only, removing utility and informativeness pressures (ablation) | VQ-VIB · ManyNames | NID: 0.55 -> 0.99 | Hurt |
| Evaluate the best-aligned model on prototypical objects only | VQ-VIB λ*Eng · ManyNames prototypical objects | NID: 0.55 (±0.00) -> 0.45 (±0.02) | Helped |
Surprising findings
- Highly informative systems also achieve high utility, indicating utility and informativeness are only partially in competition.
- Small complexity pressure (lambda_C around 0.02) produces large drops in lexicon size, so the complexity term acts at a very small scale.
Limitations
- Evaluated only on English data and only the lexicon, not syntax or compositional meaning.
- Best model's NID of 0.55 shows only moderate alignment with English speakers.
- Pretrained ResNet18 classification features may leak some linguistic knowledge.
Experimental setup
Datasets: ManyNames Visual Genome (detector fine-tuning)
Benchmarks: ManyNames
Metrics: utility (accuracy) negative MSE complexity (bits) lexicon size normalized information distance (NID)
Baselines: utility-only emergent communication (lambda_U=1) informativeness-only (lambda_I=1) complexity-only (lambda_C=1)
Builds on: VQ-VIB information bottleneck Rational Speech Act Bottom-Up attention
Closest papers by meaning
- Semantics and Spatiality of Emergent Communication (NeurIPS 2024 · similarity 0.88)
- Improving Cooperation in Language Games with Bayesian Inference and the Cognitive Hierarchy (AAAI 2025 · similarity 0.87)
- Implicit Intelligence - Evaluating Agents on What Users Don’t Say (ICML 2026 · similarity 0.87)
- Lewis's Signaling Game as beta-VAE For Natural Word Lengths and Segments (ICLR 2024 · similarity 0.87)
- Speaking Your Language: Spatial Relationships in Interpretable Emergent Communication (NeurIPS 2024 · similarity 0.87)
- Thought Communication in Multiagent Collaboration (NeurIPS 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.