Macaron-A2UI-Tall

Macaron-A2UI: A Model for Generative UI in Personal Agents

<p align="center"> <b>📰&nbsp;&nbsp;<a href="https://huggingface.co/mindlab-research/Macaron-A2UI-Tall/blob/main/Macaron_A2UI.pdf">Technical Report</a></b> &nbsp;&nbsp;&nbsp; | &nbsp;&nbsp;&nbsp; <b>📄&nbsp;&nbsp;<a href="https://arxiv.org/abs/2605.24830">Paper</a></b> </p> </p>

This repository contains the LoRA adapter weights for Macaron A2UI Tall.

Macaron A2UI Tall is a LoRA adapter trained to generate valid A2UI v0.8 cards from user context. It is designed for dynamic UI generation in personal-agent scenarios, where a model converts conversation context, product state, and available actions into one structured UI card.

This release corresponds to Macaron A2UI Tall.

Highlights

• A2UI v0.8 card generation: generates structured UI cards that can be consumed by an A2UI-compatible renderer.
• LoRA adapter release: lightweight adapter weights for continued training, inspection, and adaptation.
• Context-aware UI generation: takes user intent, conversation context, product state, and available actions as input.
• GRPO post-training: this release is produced with GRPO on top of a Qwen3-30B A2UI initialization.
• Validation-first design: outputs should be checked by the provided A2UI v0.8 validator before rendering.

Model Overview

Adapter Details

Model Variants

You are currently viewing the Tall release.

Quickstart

This repository contains adapter weights only. Load the corresponding foundation checkpoint first, then attach this adapter with PEFT.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen3-30B-A3B-Instruct-2507"
adapter_id = "mindlab-research/Macaron-A2UI-Tall"

tokenizer = AutoTokenizer.from_pretrained(
    base_model_id,
    trust_remote_code=True,
)

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

messages = [
    {
        "role": "system",
        "content": "You are an A2UI v0.8 card generation model. Output exactly one valid A2UI JSON card."
    },
    {
        "role": "user",
        "content": "<USER_CONTEXT_JSON>",
    },
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    do_sample=False,
)

response = tokenizer.decode(
    outputs[0][inputs.input_ids.shape[-1]:],
    skip_special_tokens=True,
)

print(response)

Output Contract

Macaron A2UI Tall is trained to output:

• valid JSON;
• a top-level object of the form {"text_response": "...", "a2ui": [...]};
• no Markdown code fences;
• no extra explanation outside the JSON object;
• only A2UI actions and components supported by the calling product surface.

The a2ui field is expected to contain A2UI v0.8 messages such as beginRendering, surfaceUpdate, dataModelUpdate, or deleteSurface.

The model targets A2UI v0.8. Compatibility with later protocol revisions is not guaranteed without additional validation or fine-tuning.

Evaluation

We evaluate Macaron A2UI on internal A2UI v0.8 card-generation benchmarks and product-aligned task suites.

Public benchmark numbers and reproduction details are being standardized and will be added in a future revision of this model card.

At the moment, this repository should be interpreted as an adapter release first. Evaluation methodology, task definitions, and comparable public results are still being consolidated.

Limitations

Macaron A2UI Tall is specialized for A2UI generation and is not intended as a general-purpose chat model.

Known limitations:

• may generate valid JSON that is still semantically weak;
• may hallucinate actions if the action space is underspecified;
• may fail on A2UI versions other than v0.8;
• requires external validation before production rendering;
• should not be used for irreversible or safety-critical UI actions without user confirmation.

License

The adapter weights are released under MIT.

This adapter is trained on top of Qwen/Qwen3-30B-A3B-Instruct-2507. Users are responsible for complying with both:

1. the adapter license;
2. the license of the corresponding foundation checkpoint.

Citation

@misc{kong2026macaron_a2ui,
  author = {Fancy Kong and Congjie Zheng and Murphy Zhuang and Rio Yang and Sueky Zhang and Hao Fu and Gene Jin and Andrew Chen and Pony Ma and {Mind Lab}},
  title = {Macaron-A2UI: A Model for Generative UI in Personal Agent},
  year = {2026},
  howpublished = {Mind Lab: A Lab for Experiential Intelligence},
  note = {https://macaron.im/mindlab/research/macaron-a2ui-generative-ui-personal-agent}
}

Contact

[email protected]