xiamoent/Agent-G2-alfworld-7b

Agent-G2 ALFWorld 7B

Agent-G2 ALFWorld 7B is an ALFWorld-specialized language-agent checkpoint initialized from Qwen2.5-7B-Instruct and post-trained with Agent-G2: Gaussian Guidance for Agentic Reinforcement Learning.

Agent-G2 samples an expert-prefix depth for each task from an adaptive Gaussian distribution. The distribution is updated from rollout statistics already collected for policy optimization, without additional probe rollouts or a learned depth predictor.

Project Page · Code · Model Collection · Training Data

Important: This checkpoint is designed for research in the sandboxed ALFWorld text environment. It is not a general-purpose chat model or a controller for a physical robot.

Model Details

Item	Description
Base model	Qwen/Qwen2.5-7B-Instruct
Architecture	Qwen2ForCausalLM
Checkpoint format	BF16 Safetensors
Configured context length	32,768 tokens
Target environment	ALFWorld / ALFRED text environment
Post-training	Agent-G2 with GRPO
Language	English
Required output format	<think>...</think><action>...</action>

Although the tokenizer metadata contains a larger generic maximum length, the model configuration declares 32,768 positions. The public 1.5B reference recipe uses substantially shorter prompts and responses.

Evaluation

The Agent-G2 project reports the following ALFWorld success rates for this 7B checkpoint:

Task group	Success rate
Pick	100.0%
Look	100.0%
Clean	100.0%
Heat	100.0%
Cool	100.0%
Pick Two	91.7%
All tasks	98.4%

Agent-G2 uses expert-prefix guidance as a training mechanism rather than an inference-time dependency. The project evaluates the learned policy without supplying an expert trajectory.

These results are reported by the Agent-G2 repository and have not been independently reproduced in this model card. Evaluation variance is not currently available. Results may vary with the ALFWorld version, task split, prompt template, action history, random seed, and decoding configuration.

Intended Use

This checkpoint is intended for:

• reproducing Agent-G2 results in the ALFWorld text environment;
• research on long-horizon language agents and agentic reinforcement learning;
• studying adaptive expert-prefix guidance;
• evaluating action selection over an environment-provided admissible action set.

For faithful evaluation, use the ALFWorld environment, prompt template, action parser, and rollout loop provided by the Agent-G2 repository. A standalone generation only demonstrates that the checkpoint loads successfully; it does not reproduce the interactive benchmark.

Environment Interface

At every environment step, provide the task, current observation, recent history, and admissible actions. The released parser expects English output containing reasoning and one action selected from the current admissible set:

<think>Reason about the observation and admissible actions.</think>
<action>put apple 1 in/on fridge 1</action>

Missing tags or outputs containing Chinese characters are marked invalid by the released ALFWorld parser. The action inside <action>...</action> must match an action that the environment currently allows.

Quick Start

pip install -U transformers accelerate torch

The following example performs one ALFWorld-style generation step. Replace the placeholders with state supplied by the environment:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "xiamoent/Agent-G2-alfworld-7b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)
model.eval()

task_description = "<ALFWorld task>"
current_observation = "<current observation>"
admissible_actions = ["<admissible action 1>", "<admissible action 2>"]
actions_text = ", ".join(admissible_actions)

prompt = f"""
You are an expert agent operating in the ALFRED Embodied Environment.
Your task is to: {task_description}
Your current observation is: {current_observation}
Your admissible actions of the current situation are: [{actions_text}].

Now take one action. Enclose your reasoning within <think> </think> tags, then
present one admissible action within <action> </action> tags.
""".strip()

inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=True,
        temperature=0.4,
        top_p=0.8,
        top_k=20,
        repetition_penalty=1.05,
    )

new_tokens = output_ids[0, inputs["input_ids"].shape[-1]:]
response = tokenizer.decode(new_tokens, skip_special_tokens=True)
print(response)

The released checkpoint's generation_config.json defaults to temperature 0.7. The example uses temperature 0.4 to match the project's evaluation setting.

Training

Agent-G2 uses expert ALFWorld trajectories as prefix guidance during training, followed by policy rollouts and GRPO updates. The guidance depth is sampled per task from a Gaussian distribution estimated online from existing rollout statistics. Prefix guidance is not required at inference time.

The associated data release includes 3,553 ALFWorld expert trajectories with action lengths from 3 to 15. See the Agent-G2 repository and training data for the public implementation and expert-prefix store.

The repository currently publishes paper-locked ALFWorld training commands for the 1.5B setup, but not a separate 7B hyperparameter file or the exact checkpoint-selection rule for this Hub artifact. This card therefore does not attribute the 1.5B-specific hyperparameters to the 7B checkpoint.

Limitations

• The model is specialized for the text-based ALFWorld environment and may not generalize to other simulators or physical environments.
• It can produce malformed or inadmissible actions; environment-side validation is required.
• Performance is sensitive to prompt formatting, observation history, decoding settings, random seed, and environment configuration.
• The reported evaluation does not include variance across repeated runs.
• The model may inherit factual errors, biases, and other limitations from the base model and training data.
• This checkpoint should not directly control physical systems or be used for consequential real-world actions without independent safety mechanisms.

Citation

If you find this checkpoint useful, please cite Agent-G2:

@misc{wang2026agentg2,
  title         = {Agent-G$^2$: Gaussian Guidance for Agentic Reinforcement Learning},
  author        = {Zixuan Wang and Yanrui Miao and Zhengxi Lu and Teng Pan and Yiwen Qiu
                   and Hongxing Li and Peng Qiu and Ruiqing Zhang and Yongliang Shen},
  year          = {2026},
  eprint        = {2608.23318},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2608.23318},
  note          = {Accepted at EMNLP 2026 Main Conference},
}

The paper has been accepted to the EMNLP 2026 Main Conference. A public paper link will be added when available.

Acknowledgements

Agent-G2 builds on verl-agent, veRL, and ALFWorld.