Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Interpretable agent communication from scratch (with a generic visual processor emerging on the side)
Roberto Dessi, Eugene Kharitonov, Marco Baroni
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 |
|---|---|---|---|
| add data augmentation (different views for Sender and Receiver) (ablation) | Communication game (separate CNNs) · Gaussian blobs | game accuracy (%): 43.4% -> 0.8% (chance) | Improved |
| add data augmentation (ablation) | Communication game (separate CNNs) · ILSVRC-2012 val game | game accuracy (%): 91.2% -> 81.5% | Hurt |
| add data augmentation (ablation) | Communication game · ILSVRC-val protocol | nMI: 0.50 -> 0.58 | Improved |
| share CNN weights between Sender and Receiver (ablation) | Communication game with augmentation · ILSVRC-2012 val game | game accuracy (%): 81.5% -> 82.2% | No clear effect |
| increase training distractors from 1 to 127 (ablation) | Game +aug -shared · ILSVRC-2012 val game | game accuracy (%): 7.3% -> 81.5% | Improved |
| increase training distractors from 63 to 127 (ablation) | Game +aug -shared · OOD set game | game accuracy (%): 66.0% -> 72.0% | Improved |
| reduce vocabulary size from 2048 to 512 (ablation) | Game +aug -shared · ILSVRC-2012 val game | game accuracy (%): 81.5% -> 68.8% | Hurt |
| increase vocabulary size from 2048 to 4096 (ablation) | Game +aug -shared · ILSVRC-2012 val game | game accuracy (%): 81.5% -> 77.7% | No clear effect |
| use game-trained Sender CNN features instead of SimCLR features | linear evaluation · ILSVRC-2012 linear eval | top-1 acc: 60.6% -> 60.2% | No clear effect |
| train communication game instead of discretizing SimCLR | referential game · OOD set game | game accuracy (%): 47.4% -> 72.0% | Improved |
| apply augmentation at test time to a model trained without augmentation (ablation) | Game -aug -shared · ILSVRC-2012 val game | game accuracy (%): 91.2% -> 1.9% | Hurt |
Closest papers by meaning
- Learning to Communicate and Solve Visual Blocks-World Tasks (AAAI 2019 · similarity 0.86)
- The Curious Robot: Learning Visual Representations via Physical Interactions (ECCV 2016 · similarity 0.86)
- Learning Multi-Object Positional Relationships via Emergent Communication (AAAI 2024 · similarity 0.84)
- MATA: A Trainable Hierarchical Automaton System for Multi-Agent Visual Reasoning (ICLR 2026 · similarity 0.84)
- A Multimodal Automated Interpretability Agent (ICML 2024 · similarity 0.84)
- Expressive Multi-Agent Communication via Identity-Aware Learning (AAAI 2024 · similarity 0.83)
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.
