This repository contains the official implementation for the TMLR 2026 paper:

Can VLMs Reason Robustly? A Neuro-Symbolic Investigation Weixin Chen, Antonio Vergari, Han Zhao Paper link: https://arxiv.org/abs/2603.23867

VLC is a simple neuro-symbolic pipeline for visual deductive reasoning. A VLM first recognizes object concepts from an image, and an SDD circuit then executes the task rule exactly over the predicted concepts.

VLC/
  config/                 Prompt, few-shot, and rule files
  eval/                   Evaluation scripts
  finetune/               Fine-tuning scripts
  models/sdd.py           Rule parser and SDD compiler
  environment.yml         Conda environment

Datasets and checkpoints are not included. They can be downloaded from https://drive.google.com/drive/folders/1m70uXTPRUwg6Q7yBiN7A55dBnC-KbyR4?usp=sharing. Add them as folders or symlinks under the repository root:

datasets/
  MNMath_Add_3digit/{train,val,test}/
  MNMath_Add_5digit/{train,val,test}/
  MNMath_Add_7digit/{train,val,test}/
  MNLogic_XOR_3digit/{train,val,test}/
  MNLogic_XOR_5digit/{train,val,test}/
  MNLogic_XOR_7digit/{train,val,test}/
  KandLogic_3obj/{train,val,test}/
  KandLogic_5obj/{train,val,test}/
  KandLogic_7obj/{train,val,test}/

checkpoints/
  MNMath_Add_3digit_Qwen_Qwen2.5-VL-7B-Instruct_End2end/
  MNMath_Add_3digit_Qwen_Qwen2.5-VL-7B-Instruct_Comp/
  MNLogic_XOR_3digit_Qwen_Qwen2.5-VL-7B-Instruct_End2end/
  MNLogic_XOR_3digit_Qwen_Qwen2.5-VL-7B-Instruct_Comp/
  KandLogic_3obj_Qwen_Qwen2.5-VL-7B-Instruct_End2end/
  KandLogic_3obj_Qwen_Qwen2.5-VL-7B-Instruct_Comp/

Each split should contain matching .png images and .joblib labels. The labels are expected to follow the format used by the released datasets: label for the task target and meta.concepts for object concepts.

conda env create -f environment.yml
conda activate qwen

If you already have a compatible Qwen2.5-VL environment, make sure it includes pysdd, qwen-vl-utils, transformers, trl, peft, bitsandbytes, torch, and torchvision.

Run the fine-tuning jobs used for the main experiments:

bash finetune/finetune.bash

The script fine-tunes end-to-end VLMs and VLC's concept recognizers, respectively, on the smallest training task in each family: MNMath_Add_3digit, MNLogic_XOR_3digit, and KandLogic_3obj. Checkpoints are saved under checkpoints/.

To enable Weights & Biases logging:

REPORT_TO=wandb bash finetune/finetune.bash

Run the three main experiment families:

bash eval/eval_mnmath.bash
bash eval/eval_mnlogic.bash
bash eval/eval_kandlogic.bash

By default, these scripts use:

MODEL_NAME_OR_PATH=Qwen/Qwen2.5-VL-7B-Instruct
LLM_NAME_OR_PATH=Qwen/Qwen2.5-7B-Instruct
SEEDS="0 1 2 3 4"

Override them from the command line if needed:

MODEL_NAME_OR_PATH=Qwen/Qwen2.5-VL-3B-Instruct SEEDS="0" bash eval/eval_mnmath.bash

Outputs are written to:

results/eval/       Few-shot prompting, Prism, and VLC results
results/finetune/   Fine-tuned end-to-end and fine-tuned VLC results

@article{chen2026vlc,
  title={Can VLMs Reason Robustly? A Neuro-Symbolic Investigation},
  author={Chen, Weixin and Vergari, Antonio and Zhao, Han},
  journal={Transactions on Machine Learning Research (TMLR)},
  year={2026}
}