AI Paper Atlas

Research map / Multi-agent systems and games / Human-AI and goal-conditioned

Pairwise is Not Enough: Hypergraph Neural Networks for Multi-Agent Pathfinding

Rishabh Jain, Keisuke Okumura, Michael Amir, Pietro Lio, Amanda Prorok

ICLR 2026 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Learning-guided MAPF search · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace GNN layers with hypergraph attention layers (ablation)MAGAT+ · Dense Warehouse (max agent density)success rate (%): 2.3% -> 39.8%Improved
replace GNN layers with hypergraph attention layers (ablation)MAGAT+ · Dense Maze (max agent density)success rate (%): 72.7% -> 75.8%Improved
replace GNN layers with hypergraph attention layers (ablation)MAGAT+ · Dense Room (max agent density)Rel. SoC: 2.63 ± 0.07 -> 2.35 ± 0.06Improved
replace GNN layers with hypergraph attention layers (ablation)MAGAT+ · Dense Maze (max agent density)Rel. SoC: 2.07 ± 0.09 -> 1.79 ± 0.09Improved
replace GNN layers with hypergraph attention layers (ablation)MAGAT+ · Sparse Maze (max agent density)success rate (%): 93.8% -> 93.0%No clear effect
replace GNN layers with hypergraph attention layers (ablation)MAGAT+ · ost003d (max agent density)Rel. SoC: 1.06 ± 0.00 -> 1.05 ± 0.00No clear effect
replace GNN layers with hypergraph attention layers (ablation)MAGAT+ · Sparse Warehouse (max agent density)success rate (%): 96.9% -> 100.0%Improved
add hypergraph attention instead of pairwise attention, grow agents in unimportant group (ablation)MAGAT+ vs HGNN · hand-crafted Scenario 1coefficient of variation of attention, Group 0: 21.85% -> 6.24%Improved
add hypergraph attention instead of pairwise attention (ablation)MAGAT+ vs HGNN · hand-crafted Scenario 2relative influence of agent 3 over agent 4: 13% more -> 802% moreImproved
add RL-based temperature sampling to HMAGAT (ablation)HMAGAT · small maps, highest agent densitysuccess rateHurt
add each other module (hypergraph, online expert, post-training) incrementally (ablation)MAGAT+ to HMAGAT · small maps, highest agent densitysuccess rate and Rel. SoCImproved
use PIBT collision shieldingMAPF-GPT-DDG vs MAPF-GPT (85M) · Sparse Maze (128 agents)success rate (%): 6.3% -> 76.6% (DDG over 85M without shielding)Mixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy