EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making

🌐 Project Page · 📑 Paper · 💡 Idea · 📊 Results · 📋 Release Plan · 📝 Citation

---

EVOKE is a post-training method that elicits the world knowledge already inside pretrained LLM agents, so that they decide by the consequences of their actions rather than by contextual habits. It holds the state fixed, swaps in alternative goals, and trains the policy to rank the same candidate actions under each goal, with no world-model module and no inference-time planning.

The EVOKE training loop. At collected states, actions are executed and assessed under alternative goals with the state, history, and available actions fixed. The policy learns by contrastive ranking on aggregated preferences; actions can switch between positive and competing across goals.

📰 Updates

• 2026-10-01: 📄 The paper is on arXiv.
• 2026-09: 🏠 Repository and project page are live. Code, models, and data are coming soon.

✨ Highlights

• 🚀 Best across the board — best average on ALFWorld, WebShop, and search-based QA with all three backbones.
• 🌍 Transfers to unseen environments — 91.1% / 93.1% / 84.6% on unseen ALFWorld games (Qwen2.5-3B / 7B / Qwen3-1.7B).
• 🧠 Elicits, not adds, knowledge — the backbone already encodes action consequences; EVOKE makes the policy decide by the goal rather than by habit.

💡 From Prediction to Preference

Do agents need to predict consequences to use them? (a) Predict consequences. World models learn to predict future observations, adding cost and compounding errors in planning. (b) Single-goal supervision. The knowledge is already in pretrained LLMs, but one goal per state lets the policy fit habits that fail to transfer. (c) EVOKE. Change the goal at a fixed state. Preferences flip, so habits fail and the policy must use its knowledge of consequences.

📊 Results

Main Results

Main results on ALFWorld and WebShop (left) and search-based QA (right) across three backbones.

Generalization to unseen ALFWorld games.

Analysis

All analyses use Qwen2.5-3B on ALFWorld unless noted.

Decision-supervision ablations. Removing alternative goals or replacing ranking with imitation hurts unseen success. Micro success (%) and average unseen steps, averaged over training seeds.

Data efficiency. EVOKE outperforms SFT at every budget.	Iterative improvement. Every round improves every backbone.

From knowledge to goal-directed decisions. (a) Action consequences are linearly decodable before any ALFWorld training. (b) EVOKE makes the fewest habitual errors, i.e., choosing the action that is correct for another goal in the same context.

Execution on unseen games. EVOKE succeeds earlier and wastes fewer actions: (a) cumulative success over steps, (b) invalid actions per game, (c) revisits per successful game.

🤗 Models

Model	Base model
EVOKE-ALFWorld-3B	Qwen2.5-3B-Instruct
EVOKE-ALFWorld-7B	Qwen2.5-7B-Instruct

Both are full models and load directly with Transformers or vLLM:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Gnonymous/EVOKE-ALFWorld-3B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto"
)

🏠 ALFWorld Evaluation

Installation

Requires Linux, Python 3.11, an NVIDIA GPU, a C/C++ compiler, and CMake. The setup script creates two environments, one for inference and one for ALFWorld:

bash scripts/setup.sh

Data

Download the TextWorld evaluation resources from the official ALFWorld releases:

python3 scripts/download_data.py --data-root /path/to/alfworld

Run

.venv/inference/bin/python evaluate_alfworld.py \
  --model Gnonymous/EVOKE-ALFWorld-3B \
  --data-root /path/to/alfworld/json_2.1.1 \
  --output /path/to/results/evoke-3b \
  --gpus 0

For 7B, use --model Gnonymous/EVOKE-ALFWorld-7B; a local model directory also works.

• Quick check: add --num-episodes 6 --seeds 20260911.
• Run the three seeds on separate GPUs: --gpus 0,1,2.
• Environments installed with scripts/setup.sh --env-root PATH: run PATH/inference/bin/python and pass --alfworld-python PATH/alfworld/bin/python.

📄 License

The code is released under Apache 2.0. Each model follows the license of its base model: Qwen Research License for 3B and Apache 2.0 for 7B.

📋 Release Plan

Last updated: 2026-10-01

• Paper on arXiv
• Evaluation code
• Trained models (Hugging Face)
• Training code
• Goal-intervention preference data
• Scripts to reproduce the main results and analyses

📝 Citation

@article{guo2026evoke,
  title={EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making},
  author={Guo, Yuhan and Liu, Jinming and Xu, Liang and Li, Ziqiang and Huang, Jianguo and Wang, Zhicheng and Zhu, Hu and Chen, Qiuyu and Wei, Yuntao and Jin, Xin and Zeng, Wenjun},
  journal={arXiv preprint arXiv:2609.38334},
  year={2026}
}

📬 Contact

Questions and suggestions are welcome — please open an issue.