BehR-WorldModel-Textworld-Llama3.1-8B
A behavior-consistent text-based world model for TextWorld.
- Paper: Beyond State Consistency: Behavior Consistency in Text-Based World Models (arXiv:2604.13824)
- Serving and evaluation repository: https://github.com/Ricardo-H/behr-wm
Model details
| Field | Value |
|---|---|
| Base model | X1AOX1A/WorldModel-Textworld-Llama3.1-8B |
| Base revision | 0cc3d6046690ea2f947dca9843531fbcaf7e19c9 |
| Architecture | LlamaForCausalLM (Llama-3.1-8B) |
| Environment | TextWorld, via AgentGym |
Given the agent-visible interaction history and the agent's next action, the model predicts the next environment observation. It can therefore act as a text simulator for agent rollouts in TextWorld.
Transformers usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Ricardo-H/BehR-WorldModel-Textworld-Llama3.1-8B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)
Evaluation
The paper evaluates this model on TextWorld with single-step exact match, task-success and consistency metrics. See the paper and the serving and evaluation repository for the evaluation protocol and reported results.
Intended use and limitations
- Intended for research on world models and agent simulation in text-based environments.
- Evaluated only on TextWorld; behavior outside that distribution is unvalidated.
- Predicted observations may be plausible but factually wrong. Do not treat outputs as ground truth about the real environment.
- The model inherits biases and failure modes from Llama 3.1 8B and the base world model.
License and provenance
This model is a derivative of X1AOX1A/WorldModel-Textworld-Llama3.1-8B at revision 0cc3d6046690ea2f947dca9843531fbcaf7e19c9, which is itself derived from Meta Llama 3.1. Use is governed by the Llama 3.1 Community License and by the terms of the base world model; check the base repository before redistribution or commercial use. The serving and evaluation repository is Apache-2.0; these model weights are not.
Built with Llama.
The base world model and the WebShop/TextWorld data splits originate from From Word to World: Can Large Language Models be Implicit Text-based World Models? (arXiv:2512.18832).
Citation
@article{huang2026behrwm,
title = {Beyond State Consistency: Behavior Consistency in Text-Based World Models},
author = {Huang, Youling and Chen, Guanqiao and Yao, Junchi and Wang, Lu and
Yang, Fangkai and Du, Chao and Zhao, ChenZhuo and Zhao, Pu and
Lin, Qingwei and Rajmohan, Saravan and Zhang, Dongmei},
journal = {arXiv preprint arXiv:2604.13824},
year = {2026},
url = {https://arxiv.org/abs/2604.13824}
}