Research map / Multi-agent systems and games / LLM agents
WALL-E: World Alignment by NeuroSymbolic Learning improves World Model-based LLM Agents
Read the paperCheck your idea against 108,967 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.
Training-free "world alignment" that learns symbolic knowledge (code rules, knowledge graph, scene graph) from LLM-inferred rules and trajectories, turning an LLM into a neurosymbolic world model used in an LLM-based MPC agent for Mars and ALFWorld.
Problem
LLMs' prior knowledge often conflicts with a specific environment's dynamics, so they are unreliable as world models for model-based LLM agents in partially observable open-world settings.
Approach
Iteratively compare predicted and real transitions, induce action rules and knowledge/scene graphs with LLM inductive reasoning, compile them into executable Python code rules that override LLM predictions and give feedback, prune them via greedy maximum coverage, and use an LLM agent as a look-ahead planner that replans against the code-checked world model (RL-free MPC).
Key results
- On Mars, WALL-E 2.0 reaches average reward 6.7 and score 15.3% versus 8.3% for WALL-E 1.0 and 5.5 for IfR, after 5,600 environment steps.
- On ALFWorld (134 test tasks), WALL-E 2.0 achieves 98% success rate after four iterations, versus 95% for WALL-E 1.0 and RAFA.
- Learned symbolic knowledge raises the cover rate of mispredictions from 0.0% to 92.3% over iterations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add WALL-E 2.0 instead of WALL-E 1.0 | Mars agent · Mars (average) | Score (%): 8.3 -> 15.3 | Helped |
| use action rules and KG/SG together instead of action rules only (ablation) | WALL-E 2.0 · Mars | Reward: 5.1 -> 6.7 | Helped |
| use action rules and KG/SG together instead of action rules only (ablation) | WALL-E 2.0 · Mars | Score (%): 8.3 -> 15.3 | Helped |
| add action rules to KG/SG-only learning (ablation) | WALL-E 2.0 · Mars | Reward: 4.4 -> 6.7 | Helped |
| add action rules to KG/SG-only learning (ablation) | WALL-E 2.0 · Mars | Score (%): 5.2 -> 15.3 | Helped |
| add symbolic knowledge in the world model (ablation) | LLM agent with LLM world model · Mars tasks | Reward: 3.8 -> 6.7 | Helped |
| add symbolic knowledge in the world model (ablation) | LLM agent with LLM world model · Mars tasks | Score (%): 4.1 -> 15.3 | Helped |
| apply symbolic knowledge in the agent instead of no symbolic knowledge (ablation) | LLM agent without a world model · Mars tasks | Reward: 3.6 -> 5.5 | Helped |
| add symbolic knowledge to both the agent and world model instead of only the world model (ablation) | WALL-E 2.0 · Mars tasks | Reward: 6.7 -> 6.3 | Hurt |
| skip code-rule pruning (ablation) | WALL-E 2.0 · Mars | Reward: 6.7 -> 1.5 | Hurt |
| skip code-rule pruning (ablation) | WALL-E 2.0 · Mars | Score (%): 15.3 -> 1.6 | Hurt |
| reduce code-rule limit from no limit (6) to 5 (ablation) | WALL-E 2.0 · Mars | Reward: 6.7 -> 6.2 | Hurt |
Surprising findings
- Applying code rules to the world model gave larger gains than applying them only in the agent prompt.
- Using only action rules or only KG/SG reduced Mars performance, so both types are needed.
Limitations
- Performance gains are smaller when multiple environment mechanics are modified simultaneously, as rule representations may not capture complex general mechanics.
- High variance in the Survival setting due to cow arrows and initial spawn conditions.
- Results are far below DreamerV3's absolute reward, which is trained on 1M steps; the comparison is about sample efficiency.
Experimental setup
Datasets: Mars ALFWorld
Benchmarks: Mars ALFWorld
Metrics: reward score (%) success rate cover rate
Baselines: PPO DreamerV3 ReAct Reflexion Skill Library IfR WALL-E 1.0 RAFA AdaPlanner
Builds on: WALL-E 1.0 Mars IfR Voyager
Closest papers by meaning
- Reinforcement World Model Learning for LLM-based Agents (NeurIPS 2026 · similarity 0.89)
- WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment (NeurIPS 2024 · similarity 0.89)
- Agent2World: Learning to Generate Symbolic World Models via Adaptive Multi-Agent Feedback (NeurIPS 2026 · similarity 0.88)
- Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.88)
- From Words to Actions: Unveiling the Theoretical Underpinnings of LLM-Driven Autonomous Systems (ICML 2024 · similarity 0.88)
- Test-Time Adaptation for LLM Agents via Environment Interaction (ICLR 2026 · similarity 0.88)
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.