Research map / Multi-agent systems and games / LLM agents

Agent2World: Learning to Generate Symbolic World Models via Adaptive Multi-Agent Feedback

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

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.

Proposes Agent2World, a tool-augmented multi-agent framework (Deep Researcher, Model Developer, Testing Team) that generates symbolic world models (PDDL and Python code) and uses its test-feedback loop for verifier-guided rejection sampling to SFT Llama-3.1-8B.

Problem

Generating symbolic world models (PDDL domains or executable simulators) from natural language is hard because correctness is behavioral and verifiable supervision is scarce, while static validators miss behavior-level errors.

Approach

A three-role agent pipeline where a Deep Researcher fills specification gaps via web search, a Model Developer implements the model, and a Testing Team adaptively writes unit tests and play-tests simulations; the resulting multi-turn feedback trajectories, filtered by a verifier, serve as SFT data.

Key results

• On Text2World with gpt-4.1-mini, Agent2World achieves 93.1% executability and 75.4 average F1, above Text2World EC=3 (78.2, 60.1).
• On CWMB with gpt-4.1-mini, overall normalized return R reaches 0.4811 versus 0.3649 for GIF-MCTS*.
• SFT of Llama-3.1-8B on the generated trajectories raises CWMB overall R from 0.2296 to 0.3197 and Text2World executability from 27.7 to 44.6.

Surprising findings

• Without the Unit Tester, CWMB discrete-space accuracy drops by 0.3008, the largest ablation effect.
• Llama-3.1-8B with Text2World EC=3 and GIF-MCTS achieves near-zero executability on PDDL.

Limitations

• Multi-agent loop incurs higher token and compute cost than single-agent generation.
• Winnability relies on human evaluation and automatic estimates show limited agreement.
• Training data is small (1526 trajectories) and SFT is applied to a single 8B model.

Experimental setup

Datasets: Text2World CWMB ByteSized32 LIMA

Benchmarks: Text2World CWMB ByteSized32

Metrics: executability structural similarity component-wise F1 prediction accuracy normalized return technical validity specification compliance winnability physical reality alignment

Baselines: Text2World WorldCoder GIF-MCTS ByteSized32 baseline Best-of-N Self-Consistency Direct Generation Agent2World single-agent

Builds on: ReAct Text2World GIF-MCTS AgentGen LlamaFactory

Closest papers by meaning

• WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment (NeurIPS 2024 · similarity 0.89)
• Test-Time Adaptation for LLM Agents via Environment Interaction (ICLR 2026 · similarity 0.89)
• Revisiting Multi-Agent World Modeling from a Diffusion-Inspired Perspective (NeurIPS 2025 · similarity 0.89)
• Agent Planning with World Knowledge Model (NeurIPS 2024 · similarity 0.88)
• WALL-E: World Alignment by NeuroSymbolic Learning improves World Model-based LLM Agents (NeurIPS 2025 · similarity 0.88)
• Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.88)