Research map / Multi-agent systems and games / LLM agents
DeepHA: Scaling Action Chains Elicits Deep Hierarchical Agents
Zihao Wang, Muyao Li, Kaichen He, Haowei Lin, Xiaojian Ma, Anji Liu, Yitao Liang
Read the paperCheck your idea against 137,217 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.
Proposes DeepHA, a unified hierarchical agent architecture that operates across a mixture of heterogeneous action spaces by generating action chains.
Problem
Autonomous agents are usually constrained to a single predefined action space, which limits generalization across tasks and causes compounding errors from decoupled policy execution.
Approach
A unified architecture that flexibly generates action chains spanning heterogeneous action spaces within one hierarchical policy, instead of a single fixed action space with decoupled execution.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add a Motion action before the raw action (ablation) | DeepHA · Minecraft | ASR (%): 30.0 -> 39.2 | Improved |
| add a Grounding action before the raw action (ablation) | DeepHA · Minecraft | ASR (%): 30.0 -> 42.6 | Improved |
| use full Grounding, Motion, Raw chain (depth 3) instead of depth 2 Grounding chain (ablation) | DeepHA · Minecraft | ASR (%): 42.6 -> 48.9 | Improved |
| add efficient trajectory-compressed memory vs no memory (ablation) | DeepHA · Minecraft | ASR (%): 37.1 -> 39.6 | Improved |
| add efficient memory vs no memory (ablation) | DeepHA · Minecraft | Bo10 (%): 68.3 -> 77.3 | Improved |
| use efficient memory instead of full sliding-window memory (ablation) | DeepHA · Minecraft | context tokens: 7948 -> 1976 | Improved |
| use efficient memory instead of full memory (ablation) | DeepHA · Minecraft | Bo10 (%): 78.5 -> 77.3 | No clear effect |
| use Eager mode instead of Direct mode with Motion outcome (ablation) | DeepHA · Minecraft embodied | ASR (%): 32.9 -> 37.6 | Improved |
| use Eager mode instead of Direct mode with Grounding outcome (ablation) | DeepHA · Minecraft embodied | ASR (%): 37.7 -> 40.7 | Improved |
| use Greedy mode with Raw outcome instead of Direct Raw (ablation) | DeepHA · Minecraft embodied | ASR (%): 34.5 -> 48.9 | Improved |
| use Eager (Grounding) instead of Greedy (Raw) mode (ablation) | DeepHA · Minecraft | FPS: 1.2 -> 6.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- OpenHA — OpenHA: A Series of Open-Source Hierarchical Agentic Models in Minecraft (ICML 2026) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Training One Model to Master Cross-Level Agentic Actions via Reinforcement Learning (CVPR 2026 · similarity 0.90)
- Optimus-1: Hybrid Multimodal Memory Empowered Agents Excel in Long-Horizon Tasks (NeurIPS 2024 · similarity 0.88)
- From an LLM Swarm to a PDDL-empowered Hive: Planning Self-executed Instructions in a Multi-modal Jungle (ICLR 2025 · similarity 0.88)
- Latent Action Reparameterization for Efficient Agent Inference (NeurIPS 2026 · similarity 0.87)
- ToolChain*: Efficient Action Space Navigation in Large Language Models with A* Search (ICLR 2024 · similarity 0.86)
- G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · 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.
