Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Emergent Translation in Multi-Agent Communication
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 |
|---|---|---|---|
| Use the communication-based model instead of CNN-AvgMax | Word-level translation · Bergsma500 | P@1: 53.00 -> 56.39 | Helped |
| Use the communication-based model instead of NMT with neighboring pairs | Our translation model · Multi30k Task 1 EN-DE | BLEU: 3.07 -> 6.49 | Helped |
| Use a multilingual full community instead of a single language pair | Communication-based translation · Multi30k Task 1 | BLEU: 3.85 -> 4.83; 5.36 -> 7.21; 5.20 -> 7.09; 5.87 -> 8.10; 4.31 -> 6.55; 3.92 -> 5.15 | Helped |
| Use a fair community instead of a single language pair | Communication-based translation · Multi30k Task 1 | BLEU: 3.85 -> 3.73; 5.36 -> 5.56; 5.20 -> 4.81; 5.87 -> 5.96; 4.31 -> 5.08; 3.92 -> 4.00 | Mixed |
| Initialize with pretrained captioning models and fix the speaker module (ablation) | Our translation model · Multi30k Task 2 EN-DE | BLEU: 14.81 -> 14.93 | Helped |
| Use pretrained speaker and image encoder, both fixed (ablation) | Our translation model · COCO & STAIR EN-JA | BLEU: 33.26 -> 35.25 | Helped |
| Remove the shared image encoder between language pathways (ablation) | Our communication model · COCO & STAIR JA-EN | BLEU | Hurt |
| Increase the number of image distractors (ablation) | Our communication model · Sentence-level translation | BLEU | No clear effect |
| Replace the proposed listener loss with MSE loss (ablation) | Our communication model · Multi30k Task 2 EN-DE | BLEU | Hurt |
| Replace the proposed listener loss with pairwise ranking loss (ablation) | Our communication model · Sentence-level translation | BLEU | Hurt |
| Use normalized image vectors with pairwise ranking loss (ablation) | Our communication model · Sentence-level translation | BLEU | Hurt |
| Use more distractors during training (ablation) | Our communication model · Sentence-level learning speed | learning speed | Helped |
Closest papers by meaning
- Multi-Agent Cooperation and the Emergence of (Natural) Language (ICLR 2017 · similarity 0.91)
- Unsupervised Translation of Emergent Communication (AAAI 2025 · similarity 0.89)
- Learning to Model the World With Language (ICML 2024 · similarity 0.88)
- Speaking Your Language: Spatial Relationships in Interpretable Emergent Communication (NeurIPS 2024 · similarity 0.88)
- Language Grounded Multi-agent Reinforcement Learning with Human-interpretable Communication (NeurIPS 2024 · similarity 0.87)
- Learning Multi-Object Positional Relationships via Emergent Communication (AAAI 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.