---
license: cc-by-4.0
library_name: peft
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3-VL-8B-Instruct
tags:
- 3d
- multi-modal
- chain-of-thought
- neuro-symbolic
---

# APEIRIA: Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs

**APEIRIA** (ἄπειρον, *unlimited* in Greek) is a neuro-symbolic 3D multi-modal LLM framework designed to bridge the gap between interpretable neuro-symbolic reasoning and the flexibility of end-to-end 3D MLLMs. It distills symbolic reasoning patterns into MLLMs using a natural language chain-of-thought (CoT) approach.

The model was presented in the paper [Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs](https://huggingface.co/papers/2606.01215), which was accepted to ICML 2026. This repository contains the released APEIRIA model checkpoint as a LoRA adapter based on the [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) backbone.

- **Paper:** [Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs](https://huggingface.co/papers/2606.01215)
- **Project Page:** [https://matthewdm0816.github.io/Apeiria_Open](https://matthewdm0816.github.io/Apeiria_Open)
- **Repository:** [https://github.com/oceanflowlab/APEIRIA](https://github.com/oceanflowlab/APEIRIA)

## Model Description

APEIRIA introduces a three-stage curriculum to progressively build reasoning capabilities:
1. **3D Perception Alignment**: Grounds object visual-geometric features to the LLM.
2. **CoT-SFT**: Teaches query decomposition and stepwise verification from symbolic program traces.
3. **CoT-RL**: Extends reasoning patterns to open-set concepts and deeply nested instructions.

By transferring reasoning patterns rather than concept-specific knowledge, APEIRIA preserves transparent reasoning and modular interchangeability of planning and perception components.

## Usage

Please refer to the [official GitHub repository](https://github.com/oceanflowlab/APEIRIA) for environment setup, data preparation, and detailed instructions on running inference or training.

## Citation

If you find this work useful, please consider citing:

```bibtex
@inproceedings{mo2026,
  title={Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs},
  author={Mo, Wentao and Liu, Yang},
  booktitle={International Conference on Machine Learning},
  year={2026}
}
```

## Acknowledgements

This code builds upon previous 3D MLLMs and foundation models, including [Chat-Scene](https://github.com/ZzZZCHS/Chat-Scene), [SegDINO3D](https://github.com/IDEA-Research/SegDINO3D), and [Mask3D](https://github.com/jonasschult/mask3d). It utilizes the [SGLang](https://github.com/sgl-project/sglang) library for fast multi-modal generation.