Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Compositional languages emerge in a neural iterated learning model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Apply NIL with both agents reset between generations | Referential-game agents · Referential game, validation set sizes 8, 16, and 32 | validation performance: 0.136 -> 0.847; 0.132 -> 0.755; 0.102 -> 0.558 | Helped |
| Reset Alice only instead of neither agent (ablation) | Referential-game agents · Referential game, validation set sizes 8, 16, and 32 | validation performance: 0.136 -> 0.598; 0.132 -> 0.280; 0.102 -> 0.210 | Mixed |
| Reset Bob only instead of neither agent (ablation) | Referential-game agents · Referential game, validation set sizes 8, 16, and 32 | validation performance: 0.136 -> 0.094; 0.132 -> 0.152; 0.102 -> 0.116 | Mixed |
| Use 1500 Alice pre-training rounds rather than 100 (ablation) | NIL · Referential game, first ten generations | mean topological similarity: 0.225 -> 0.575 | Helped |
| Increase Alice pre-training from 1500 to 8000 rounds (ablation) | NIL · Referential game, first ten generations | mean topological similarity: 0.575 -> 0.443 | Hurt |
| Use 120 Bob pre-training batches rather than 10 (ablation) | NIL · Referential game, first ten generations | mean topological similarity: 0.415 -> 0.591 | Helped |
| Increase Bob pre-training from 120 to 800 batches (ablation) | NIL · Referential game, first ten generations | mean topological similarity: 0.591 -> 0.448 | Hurt |
| Pre-train neural agents on a high-topological-similarity language rather than a low-similarity language | Alice and Bob · Referential game language-learning experiments | learning speed | Helped |
| Use both-agent NIL rather than no iterated learning | Referential-game agents · Referential game | topological similarity | Helped |
| Use weak pre-training with an appropriate number of rounds rather than very short or very long pre-training (ablation) | NIL · Referential game | topological similarity | Mixed |
| Increase topological similarity of the emergent language | Referential-game agents · Held-out validation game | validation performance: correlation coefficient 0.928; p-value 3.8*10^{-104} | Helped |
Closest papers by meaning
- A compressive-expressive communication framework for compositional representations (NeurIPS 2025 · similarity 0.91)
- Iterated Learning Improves Compositionality in Large Vision-Language Models (CVPR 2024 · similarity 0.90)
- Bias Amplification in Language Model Evolution: An Iterated Learning Perspective (NeurIPS 2024 · similarity 0.88)
- CtD: Composition through Decomposition in Emergent Communication (ICLR 2025 · similarity 0.87)
- Lewis's Signaling Game as beta-VAE For Natural Word Lengths and Segments (ICLR 2024 · similarity 0.86)
- Bridging semantics and pragmatics in information-theoretic emergent communication (NeurIPS 2024 · similarity 0.85)
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.