Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Emergence of Linguistic Communication from Referential Games with Symbolic and Pixel Input
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 |
|---|---|---|---|
| increase maximum message length from 2 to 5 (ablation) | VisA referential game · VisA training data | communicative success: 92.0% -> 98.2% | Helped |
| increase maximum message length from 2 to 10 (ablation) | VisA referential game · VisA training data | communicative success: 92.0% -> 98.5% | Helped |
| increase maximum message length from 2 to 10 (ablation) | VisA referential game · VisA test data | communicative success: 74.2 -> 81.6 | Helped |
| use uniformly sampled feature combinations for novel objects | VisA referential game, max length 10 · VisA uniform chimera | communicative success: 47.0 -> 42.6 | Hurt |
| sample distractors from target-specific context co-occurrence statistics rather than uniformly (ablation) | VisA referential game · Context-dependent distractors | training dynamics | Mixed |
| reduce distractors from 20 to 2 (ablation) | MuJoCo pixel referential game · Game B versus game A | test communicative success: 93.6 -> 93.4 | No clear effect |
| reduce distractors from 20 to 2 (ablation) | MuJoCo pixel referential game · Game B versus game A | topographic rho: 0.13 -> 0.006 | Hurt |
| give speaker and listener different viewpoints (ablation) | MuJoCo pixel referential game · Game C versus game B | test communicative success: 93.4 -> 85.7 | Hurt |
| balance the numbers of shapes and colors (ablation) | MuJoCo pixel referential game · Game D versus game C | test communicative success: 85.7 -> 89.9 | Helped |
| train visual encoders from raw pixels using only communication reward | End-to-end ConvNet agents · MuJoCo pixel games A–D | test communicative success: 93.6; 93.4; 85.7; 89.9 | Helped |
| use gold attribute classifiers instead of learned visual representations | MuJoCo pixel game A · Game A | communicative success: 93.7 -> 95.6 | Helped |
| use gold attribute classifiers instead of learned visual representations | MuJoCo pixel games B–D · Games B–D | communicative success: 93.4 -> 99.4; 85.7 -> 98.8; 89.9 -> 98.5 | Helped |
Closest papers by meaning
- Emergence of Language with Multi-agent Games: Learning to Communicate with Sequences of Symbols (NeurIPS 2017 · similarity 0.93)
- Multi-Agent Cooperation and the Emergence of (Natural) Language (ICLR 2017 · similarity 0.91)
- Emergent Communication of Generalizations (NeurIPS 2021 · similarity 0.91)
- Compositional languages emerge in a neural iterated learning model (ICLR 2020 · similarity 0.90)
- A compressive-expressive communication framework for compositional representations (NeurIPS 2025 · similarity 0.89)
- Lewis's Signaling Game as beta-VAE For Natural Word Lengths and Segments (ICLR 2024 · similarity 0.89)
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.