Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Emergent Tool Use From Multi-Agent Autocurricula
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 |
|---|---|---|---|
| Pretrain with multi-agent hide-and-seek self-play | Policy initialization for the transfer suite · Lock and Return, Sequential Lock, Construction From Blueprint | final reward and learning speed | Helped |
| Pretrain with multi-agent hide-and-seek self-play | Policy initialization for transfer tasks · Object Counting | performance | Hurt |
| Pretrain with multi-agent hide-and-seek self-play | Policy initialization for transfer tasks · Shelter Construction | final reward | No clear effect |
| Randomize team sizes, object counts and shapes, initial positions, and walls | Hide-and-seek training environment · Hide-and-seek | emergent strategy stages: 2 stages -> 6 stages | Helped |
| Increase batch size to at least 32,000 (ablation) | Hide-and-seek PPO training · Ramp defense | training speed and convergence | Helped |
| Use an omniscient value function with access to the full state (ablation) | Hide-and-seek value function · Emergent autocurriculum | strategy progression | Helped |
| Replace a masked value function with an omniscient value function (ablation) | Hide-and-seek value function · Emergent autocurriculum | maximum stage reached: stage 3 -> stage 4 | Helped |
| Use self-attention over entities instead of embedding and pooling entities (ablation) | Hide-and-seek policy architecture · Emergent autocurriculum | sample efficiency | Helped |
| Use self-attention over entities instead of embedding and pooling entities (ablation) | Hide-and-seek policy architecture · Emergent autocurriculum | wall-clock time to convergence | Hurt |
| Use separate policy parameters per agent instead of shared parameters (ablation) | Hide-and-seek self-play · Six-stage skill progression | sample efficiency | Hurt |
| Increase count-based exploration state representation from box locations to higher-dimensional observations (ablation) | Count-based exploration · Hide-and-seek environment | agent and box movement | Hurt |
| Use Random Network Distillation instead of count-based exploration in the full-state setting (ablation) | Intrinsic-motivation baseline · Hide-and-seek environment | agent and object movement | Helped |
Closest papers by meaning
- Mimicking To Dominate: Imitation Learning Strategies for Success in Multiagent Games (NeurIPS 2024 · similarity 0.88)
- CoMAS: Co-Evolving Multi-Agent Systems via Interaction Rewards (ICLR 2026 · similarity 0.87)
- A Single Goal is All You Need: Skills and Exploration Emerge from Contrastive RL without Rewards, Demonstrations, or Subgoals (ICLR 2025 · similarity 0.87)
- Individual Contributions as Intrinsic Exploration Scaffolds for Multi-agent Reinforcement Learning (ICML 2024 · similarity 0.86)
- Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.86)
- FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics (NeurIPS 2026 · 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.