Research map / Multi-agent systems and games / LLM agents
WALL-EVE: World Alignment with Rule Learning in Visual Environments
Siyu Zhou, Tianyi Zhou, Yijun Yang, Deheng Ye, Chengqi Zhang, Jing Jiang, Guodong Long
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 rules to the world model of the MPC agent (ablation) | WALL-EVE MPC agent · Minecraft TechTree | success rate (%): 38 -> 69 | Improved |
| add rules to the agent only (ablation) | WALL-EVE · Minecraft TechTree | success rate (%): 37 -> 61 | Improved |
| also add rules to the agent on top of rule-aligned world model (ablation) | WALL-EVE · Minecraft TechTree | success rate (%): 69 -> 67 | Hurt |
| add LLM world model without rules to MPC (ablation) | LLM agent · Minecraft TechTree | success rate (%): 37 -> 38 | No clear effect |
| add visual modality input (ablation) | WALL-EVE · Minecraft TechTree | success rate (%): 59 -> 69 | Improved |
| add the world model (ablation) | WALL-EVE · Minecraft TechTree | success rate (%): 61 -> 69 | Improved |
| add rule set pruning (ablation) | WALL-EVE · Minecraft TechTree | success rate (%): 12 -> 69 | Improved |
| raise rule set limit from 2 to 4 (ablation) | WALL-EVE · Minecraft TechTree | success rate (%): 52 -> 64 | Improved |
| remove rule cap (no limit, 6 rules) vs limit 4 (ablation) | WALL-EVE · Minecraft TechTree | success rate (%): 64 -> 69 | Improved |
| use GPT-4o vision instead of fine-tuned Mask R-CNN for perception (ablation) | WALL-EVE · ALFWorld (visual) | success rate (%): 75 -> 64 | Hurt |
| use GPT-4o vision instead of fine-tuned Mask R-CNN for perception (ablation) | RAFA · ALFWorld (visual) | success rate (%): 72 -> 56 | Hurt |
| add learned rules to the LLM world model (ablation) | LLM world model · ALFWorld | success/failure accuracy (%): 73 -> 94 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- Optimus-1 — Optimus-1: Hybrid Multimodal Memory Empowered Agents Excel in Long-Horizon Tasks (NeurIPS 2024) · better on 4 of 4 shared results
- AdaPlanner — AdaPlanner: Adaptive Planning from Feedback with Language Models (NeurIPS 2023) · better on 2 of 2 shared results
- Reflexion — Reflexion: language agents with verbal reinforcement learning (NeurIPS 2023) · better on 2 of 2 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
- WALL-E: World Alignment by NeuroSymbolic Learning improves World Model-based LLM Agents (NeurIPS 2025 · similarity 0.94)
- Dyn-O: Building Structured World Models with Object-Centric Representations (NeurIPS 2025 · similarity 0.88)
- DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning (ICML 2025 · similarity 0.87)
- VAGEN: Reinforcing World Model Reasoning for Multi-Turn VLM Agents (NeurIPS 2025 · similarity 0.87)
- Steve-Eye: Equipping LLM-based Embodied Agents with Visual Perception in Open Worlds (ICLR 2024 · similarity 0.87)
- Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld (CVPR 2024 · similarity 0.87)
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.
