Instructions to use kmingl/IoA-Prober-8B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use kmingl/IoA-Prober-8B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="kmingl/IoA-Prober-8B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("kmingl/IoA-Prober-8B")
model = AutoModelForCausalLM.from_pretrained("kmingl/IoA-Prober-8B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use kmingl/IoA-Prober-8B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "kmingl/IoA-Prober-8B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kmingl/IoA-Prober-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/kmingl/IoA-Prober-8B

SGLang

How to use kmingl/IoA-Prober-8B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "kmingl/IoA-Prober-8B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kmingl/IoA-Prober-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "kmingl/IoA-Prober-8B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kmingl/IoA-Prober-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use kmingl/IoA-Prober-8B with Docker Model Runner:
```
docker model run hf.co/kmingl/IoA-Prober-8B
```

Agree to the IoA Responsible Use License

Access requires accepting the model license. Human dialogue analysis requires every participant's informed consent, equal access to questions and aggregated results, and participant confirmation before action. Surveillance, employment decisions, asymmetric monitoring, and adversarial elicitation are prohibited. Hugging Face shares your username and email with the repository owner when you request access.

Review the LICENSE before requesting access. These conditions also apply to derivatives and hosted services. Access is approved automatically after you accept the conditions.

Illusion of Alignment logo

IoA-Prober-8B

Illusion of Alignment: Detecting Hidden Disagreement in Collaborative Dialogue

IoA-Prober-8B generates diagnostic multiple-choice questions to surface hidden disagreement in collaborative dialogue. Its input is the observable transcript and public participant profiles. Participants answer independently; a difference in option choices provides a signal for discussion and confirmation.

Code and IoA-Suite · License · Upstream notice

Model and training

The model is based on Qwen3-8B, with the Qwen3 causal language model architecture and bfloat16 weights.

Training uses a supervised warm-start on 300 IoA-Suite training dialogues, followed by GRPO on all 1,200 training dialogues. The warm-start retains questions that elicit a meaningful, implicit, non-redundant divergence. GRPO optimizes per-dialogue F1 through participant simulation and the validity filter. Both stages run for two epochs. The validation split contains 120 dialogues.

Evaluation

On the 300-dialogue IoA-Suite test split, the paper reports P = 50.5%, R = 53.9%, F1 = 51.8%. These are means of per-dialogue scores. Precision excludes accepted but unattributed divergences, and recall counts distinct planted misalignments recovered.

Evaluation uses a Gemini-3-Flash participant simulator and GPT-5.4 validity filter, both at temperature 0.2, with one simulator rollout per participant. The prober uses temperature 0.7 with thinking enabled. See the repository's evaluation instructions and data description.

Usage

Accept the license on this model page, then authenticate with hf auth login. For an OpenAI-compatible inference server, use the model name kmingl/IoA-Prober-8B, enable Qwen3 thinking, and configure the endpoint in the evaluation repository.

The repository's ioa.prober_messages(case) builds the system and user messages from an IoA-Suite record. For Transformers inference, use the tokenizer's chat template with add_generation_prompt=True and enable_thinking=True, and generate at temperature=0.7. Preserve the prompt's JSON output schema.

The published generation_config.json uses temperature 0.7. Inference servers may override generation settings; set the temperature explicitly when reproducing an evaluation.

Intended use and safeguards

Use the model in symmetric collaboration where all participants seek mutual understanding. Before analyzing a dialogue, obtain informed consent from every participant and disclose that the model will infer unstated assumptions. Give every participant access to generated questions and aggregated results. Require participant confirmation before acting on a suspected disagreement.

The license prohibits surveillance, employment decisions, asymmetric monitoring, and adversarial elicitation. It applies to derivatives and hosted services. Deployments must enforce consent, visibility, and human confirmation; a download gate or a license alone cannot prevent misuse.

Limitations

IoA-Suite uses synthetic English dialogues with two participants and planted ground truth. Its F1 is a controlled proxy for detection capability, not a complete measure of hidden disagreement in real collaboration. See the data description for split characteristics and source reuse. Synthesized personas and model inferences may reflect biases. An MCQ split does not establish a person's actual beliefs, motives, competence, or performance.

License

The IoA contributions use the custom, RAIL-style IoA Responsible Use License 1.0. The upstream Qwen3-8B Apache-2.0 license is preserved in LICENSE-Qwen3-8B. See NOTICE for attribution and modifications.

Citation

@misc{liu2026illusion,
  title={Illusion of Alignment: Detecting Hidden Disagreement in Collaborative Dialogue}, 
  author={Kaiming Liu and Fuwen Luo and Ziyue Wang and Jinrui Ju and Yuxuan Liu and Xuanyu Lei and Yunghwei Lai and Peng Li and Yang Liu},
  year={2026},
  eprint={2608.08210},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2608.08210}, 
}