Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Emergent Discrete Communication in Semantic Spaces
Mycal Tucker, Huao Li, Siddharth Agrawal, Dana Hughes, Katia P. Sycara, Michael Lewis, Julie Shah
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use learnable prototype tokens instead of one-hot tokens | Multi-agent communication · Triangle, uniform prior, σ=0.9, c=9 | self-play reward: -201 -> -124 | Improved |
| Use learnable prototype tokens instead of one-hot tokens | Multi-agent communication · Triangle, skewed prior, σ=0.9, c=9 | self-play reward: -136 -> -100 | Improved |
| Use learnable prototype tokens in a 3-dimensional channel for 9 targets | 9 Points · 9 Points, σ=0, c=3 | self-play reward: -96 -> -60 | Improved |
| Use learnable prototype tokens instead of one-hot tokens | 9 Points · 9 Points, σ=0.5, c=9 | self-play reward: -132 -> -100 | Improved |
| Use prototype tokens instead of continuous communication | Uniform · Uniform | reward: -97.8 -> -113.6 | Hurt |
| Use prototype tokens instead of one-hot tokens | CIFAR10 reference game · Label Out | accuracy: 50% (0.8) -> 62% (1.2) | Improved |
| Use prototype communication with BERT embeddings instead of continuous communication | CIFAR10 reference game · AMT Out | accuracy: 53% (0.6) -> 56% (0.6) | Improved |
| Use prototype tokens instead of BERT-based embeddings as the human-interpreted communication | Human listeners · Human understanding of held-out tokens | accuracy: 66% -> 70% | No clear effect |
| Remove noise from prototype-agent training (ablation) | CIFAR10 reference game · Label Out | accuracy: 58% (1.0) -> 55% (1.0) | Hurt |
| Remove non-identity reward structure from prototype-agent training (ablation) | CIFAR10 reference game · Label Out | accuracy: 58% (1.0) -> 53% (1.0) | Hurt |
| Remove both communication noise and non-identity reward structure (ablation) | CIFAR10 reference game · AMT Out | accuracy: 56% (0.6) -> 53% (0.5) | Hurt |
| Replace class-name embeddings with animal/vehicle binary labels (ablation) | Proto (BERT) · CIFAR10 reference game | Label In accuracy: 93% (0.7) -> 61% (0.3) | Hurt |
Closest papers by meaning
- Emergence of Language with Multi-agent Games: Learning to Communicate with Sequences of Symbols (NeurIPS 2017 · similarity 0.90)
- Trading off Utility, Informativeness, and Complexity in Emergent Communication (NeurIPS 2022 · similarity 0.89)
- A compressive-expressive communication framework for compositional representations (NeurIPS 2025 · similarity 0.88)
- Semantics and Spatiality of Emergent Communication (NeurIPS 2024 · similarity 0.88)
- Bridging semantics and pragmatics in information-theoretic emergent communication (NeurIPS 2024 · similarity 0.88)
- Language Grounded Multi-agent Reinforcement Learning with Human-interpretable Communication (NeurIPS 2024 · 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.
