Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Beyond Rewards: a Hierarchical Perspective on Offline Multiagent Behavioral Analysis
Shayegan Omidshafiei, Andrei Kapishnikov, Yannick Assogba, Lucas Dixon, Been Kim
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 |
|---|---|---|---|
| use hierarchical joint+local latents (MOHBA) instead of an LSTM final-hidden-state encoding | MOHBA · Hill-climbing domain | APL (lower better): 4.8 -> 1.3 | Improved |
| use hierarchical joint+local latents (MOHBA) instead of an LSTM final-hidden-state encoding | MOHBA · Coordination game | APL (lower better): 2.9 -> 2.4 | Improved |
| use hierarchical latents instead of flat-VAE feeding only z_omega to policies (remove local latents and local KL) (ablation) | MOHBA · Hill-climbing domain | APL (lower better): 6.9 -> 1.3 | Improved |
| use hierarchical latents instead of flat-VAE feeding only z_omega to policies (ablation) | MOHBA · Coordination game | APL (lower better): 5.0 -> 2.4 | Improved |
| use MOHBA joint latent instead of LSTM hidden state for K-means clustering | MOHBA · Hill-climbing domain | ICTD (lower better): 0.47 -> 0.17 | Improved |
| use MOHBA joint latent instead of LSTM hidden state for K-means clustering | MOHBA · Coordination game | ICTD (lower better): 0.34 -> 0.16 | Improved |
| add hierarchy (local latents) over flat-VAE for joint clustering quality (ablation) | MOHBA · Hill-climbing domain | ICTD (lower better): 0.18 -> 0.17 | No clear effect |
| add hierarchy (local latents) over flat-VAE for joint clustering quality (ablation) | MOHBA · Coordination game | ICTD (lower better): 0.16 -> 0.16 | No clear effect |
| flat-VAE baseline still recovers the 3 coordination clusters, whereas LSTM hidden states do not | LSTM vs VAE baseline · Coordination game | latent cluster interpretability (qualitative) | Mixed |
| predict dispersion class from z_omega only (instead of full trajectory) | MOHBA · Hill-climbing agent dispersion (5 classes) | validation accuracy: 54.27% | Improved |
| predict total-return class from z_omega only | MOHBA · Hill-climbing total return (5 classes) | validation accuracy: 60.98% | Improved |
| cyclically anneal the KL weight beta from 0 to target value | MOHBA training · training stability | KL vanishing (qualitative) | Improved |
Closest papers by meaning
- Learning Generalizable Skills from Offline Multi-Task Data for Multi-Agent Cooperation (ICLR 2025 · similarity 0.87)
- Agent-Omit: Adaptive Context Omission for Efficient LLM Agents (ICML 2026 · similarity 0.87)
- $\texttt{Multi}^2$: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments (ICML 2026 · similarity 0.87)
- Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)
- Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models (NeurIPS 2025 · similarity 0.87)
- A2PO: Towards Effective Offline Reinforcement Learning from an Advantage-aware Perspective (NeurIPS 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.
