Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Speaking Your Language: Spatial Relationships in Interpretable 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.
Shows that emergent-communication agents in a modified referential game can develop spatial (positional) references to targets within a sequence window, and interprets the emergent language with an NPMI-based collocation analysis.
Problem
Emergent communication has not yet produced languages with spatial deixis, i.e. relative positional references to parts of an observation, which are valuable for shared context and interpretability.
Approach
A referential game over a windowed integer sequence forces senders to describe a target's position relative to other elements; NPMI-based collocation (PMI_nc and PMI_c) segments messages and maps n-grams to integers and relative positions, producing a human-readable dictionary that is validated by receiver accuracy.
Key results
- Agent pairs achieve above 98% accuracy on the referential task across 16 seeds.
- Non-compositional positional dictionary (begin, begin+1, end-1, end) yields about 90% average receiver accuracy versus 20% chance.
- Compositional-P dictionaries reach up to 78% receiver accuracy, while Compositional-NP stays near chance (22% average).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NPMI-identified non-compositional positional messages | Spatial referential agents · Non-Compositional Positional dictionary | accuracy: 90% ±3% (average); 94% (maximum); random chance 20% | Helped |
| Use NPMI-identified non-compositional integer messages | Spatial referential agents · Non-Compositional Integer dictionary | accuracy: 36% ±0.4%; random chance 20% | Helped |
| Add identified positional components to compositional messages (ablation) | Spatial referential agents · Compositional dictionary | accuracy: 22% ± 2% -> 30% ± 21%; maximum 28% -> 78% | Helped |
| Replace compositional positional components with 0 (ablation) | NPMI dictionary evaluation · Compositional dictionary | accuracy: 30% ± 21% -> 22% ± 2%; maximum 78% -> 28% | Hurt |
| Evaluate on sequences shortened by 40 after training length 60 | Spatial referential agents · Sequence length generalization | accuracy difference: -15.8% | Hurt |
| Evaluate on sequences shortened by 40 after training length 100 | Spatial referential agents · Sequence length generalization | accuracy difference: -5.2% | Hurt |
| Increase hidden size from 64 to 128 for sequence length 100 | Spatial referential agents · Training sequence length 100 | training and validation accuracy: over 90% | Helped |
| Vary vocabulary size or training sequence length | Spatial referential agents · Generalization evaluation | task accuracy | No clear effect |
Surprising findings
- Compositional spatial references emerged in only 25% of agent pairs, and non-compositional integer messages in about 45% of runs.
- Non-compositional monolithic messages are used for sparse observation types (begin, end), which the authors hypothesise is due to observation sparsity and compressibility.
Limitations
- Average accuracy of non-compositional integer and compositional-positional dictionaries is only about 33%, indicating incomplete interpretation.
- Environment is a simple synthetic referential game; non-trivial compositional functions cannot be analysed with NPMI.
- Performance declines when evaluated on sequences over 50% shorter than training sequences.
Experimental setup
Datasets: synthetic integer sequences of length 60 with 5-integer observation windows
Benchmarks: synthetic referential game test set (20k samples) NPMI-derived dictionary accuracy datasets
Metrics: accuracy NPMI translation accuracy
Baselines: random chance (20%) Compositional-NP dictionary ablation
Builds on: EGG toolkit Gumbel-Softmax referential game
Closest papers by meaning
- Learning Multi-Object Positional Relationships via Emergent Communication (AAAI 2024 · similarity 0.89)
- Semantics and Spatiality of Emergent Communication (NeurIPS 2024 · similarity 0.89)
- Bridging semantics and pragmatics in information-theoretic emergent communication (NeurIPS 2024 · similarity 0.87)
- Do Vision-Language Models Represent Space and How? Evaluating Spatial Frame of Reference under Ambiguities (ICLR 2025 · similarity 0.86)
- Unsupervised Translation of Emergent Communication (AAAI 2025 · similarity 0.86)
- Lewis's Signaling Game as beta-VAE For Natural Word Lengths and Segments (ICLR 2024 · similarity 0.86)
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.