Salesforce/cogalign-internvl2_5-mpo-1b

Why Vision Language Models Struggle with Visual Arithmetic? Towards Enhanced Chart and Geometry Understanding

Vision Language Models (VLMs) have achieved remarkable progress in multimodal tasks, yet they often struggle with visual arithmetic, seemingly simple capabilities like object counting or length comparison, which are essential for relevant complex tasks like chart understanding and geometric reasoning. In this work, we first investigate the root causes of this deficiency through a suite of probing tasks focusing on basic visual arithmetic. Our analysis reveals that while pre-trained vision encoders typically capture sufficient information, the text decoder often fails to decode it correctly for arithmetic reasoning. To address this, we propose CogAlign, a novel post-training strategy inspired by Piaget's theory of cognitive development. CogAlign trains VLMs to recognize invariant properties under visual transformations. We demonstrate that this approach significantly improves the performance of three diverse VLMs on our proposed probing tasks. Furthermore, CogAlign enhances performance by an average of 4.6% on CHOCOLATE and 2.9% on MATH-VISION, outperforming or matching supervised fine-tuning methods while requiring only 60% less training data. These results highlight the effectiveness and generalizability of CogAlign in improving fundamental visual arithmetic capabilities and their transfer to downstream tasks.

Quick start

Loading models can be done easily with transformers:

from transformers import AutoTokenizer, AutoModel, AutoConfig
path = "Salesforce/cogalign-internvl2_5-mpo-1b"
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    use_flash_attn=True,
    trust_remote_code=True).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)

Then, we define some functions for inference

# Adapted from https://huggingface.co/OpenGVLab/InternVL2_5-1B-MPO
import copy
import pandas as pd
from datasets import load_dataset
import requests
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode

# Taken from InternVL's code
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform

def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio

def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_file, input_size=448, max_num=12, is_url=False):
    if is_url:
        image = Image.open(requests.get(image_file, stream=True).raw)
    else:
        image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

After that, we can try an example on the CHOCOLATE datset:

chocolate = load_dataset("khhuang/CHOCOLATE")["test"]
chocolate_df = pd.DataFrame(chocolate)
chocolate_df_lvlm = chocolate_df.loc[chocolate_df.split=="LVLM",:]

instance =  chocolate_df_lvlm.iloc[2]
caption = ' '.join(instance.sentences)

url = instance.image_path
pixel_values = load_image(url, max_num=12, is_url=True).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)

prompt = f"""
You are given a chart and a caption, you are tasked to detect whether the caption is factually
consistent with the chart. A caption is factually consistent with the chart if it describes the datapoints within the charts without factual errors (e.g. wrong label, value, trends).
[Start of Caption]
{caption}
[End of Caption]
For the above caption, you should respond 'Answer: Yes' if it is factually consistent with the chart. Otherwise, respond 'Answer: No'. Do not provide explanation or other thing.
"""
question = f'<image>\n{prompt}'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')

License information

This release is for research purposes only in support of an academic paper. This repository is licensed under the noncommercial license CC-BY-NC 4.0.

Citation

If you find CogAlign useful in your research, please consider citing:

@misc{huang-etal-2025-cogalign,
    title = "Why Vision Language Models Struggle with Visual Arithmetic? Towards Enhanced Chart and Geometry Understanding",
    author = "Huang, Kung-Hsiang  and
      Qin, Can  and
      Qiu, Haoyi  and
      Laban, Philippe  and
      Joty, Shafiq  and
      Xiong, Caiming  and
      Wu, Chien-Sheng",
    year = "2025",
    eprint={2502.11492},
    archivePrefix = "arXiv",
    primaryClass={cs.AI}
}