💻 computer science

CounterCount: A Diagnostic Framework for Counting Bias in Vision Language Models

The paper introduces CounterCount, a diagnostic framework that reveals Vision-Language Models' reliance on object priors over visual evidence in counterfactual counting tasks and proposes an inference-time attention modulation strategy to significantly improve their accuracy.

Original authors: Reem Alzahrani, Hassan Alshanqiti, Bushra Bin Hemid, Zaid Alyafeai, Abdelrahman Eldesokey, Bernard Ghanem

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Reem Alzahrani, Hassan Alshanqiti, Bushra Bin Hemid, Zaid Alyafeai, Abdelrahman Eldesokey, Bernard Ghanem

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart, well-read parrot that has memorized millions of books about the world. You show this parrot a picture of a rabbit and ask, "How many ears does this rabbit have?" Because the parrot has read so many stories about rabbits, it immediately answers, "Two!" without really looking at the picture. It's relying on its memory (what it thinks a rabbit should look like) rather than its eyes (what is actually in the photo).

Now, imagine you take that same picture of the rabbit and use a digital editor to give it four ears. If you ask the parrot again, a truly observant animal would say, "Four!" But this smart parrot, still stuck in its memory, stubbornly says, "Two!" It ignores the visual evidence because its "book knowledge" is so strong it overpowers what it sees.

This is exactly the problem the paper CounterCount investigates.

The Problem: The "Book Smarts" vs. "Eyes" Conflict

The researchers found that modern AI models (called Vision-Language Models) are great at reading and talking, but they often fail at simple counting tasks when the picture contradicts what they've learned from their training data. They are like that parrot: they trust their internal "rulebook" more than the actual image in front of them.

The Solution: A Diagnostic Test Kit

To prove this, the team built a special test kit called CounterCount.

The Setup: They created pairs of images. One is a normal picture (e.g., a car with 4 wheels). The other is a "counterfactual" picture where they digitally changed a specific part (e.g., the same car now has 6 wheels).
The Test: They asked the AI, "How many wheels does this car have?"
The Result: The AI did great on the normal pictures. But on the weird, edited pictures, it often failed, sticking to the "normal" answer (4) instead of counting the actual wheels (6). This proved the AI was ignoring the visual evidence in favor of its pre-learned biases.

The "Why": The AI is Looking, But Not Listening

You might think the AI fails because it can't see the extra wheels. The researchers dug deeper and found that wasn't the case.

The Metaphor: Imagine the AI is a student taking a test. The student is looking at the right part of the diagram (the wheels), but their brain is distracted by a loud voice in their head shouting, "Cars have 4 wheels!" The student sees the 6 wheels, but the "loud voice" (the language bias) drowns out the visual signal.
The Proof: By looking at the AI's internal "attention maps" (which show what the AI is focusing on), the researchers saw that the AI was looking at the wheels. However, it wasn't giving those wheels enough "mental weight" to override its memory.

The Fix: Turning Up the Volume on the Eyes

The researchers developed a clever trick called Attention Modulation.

The Analogy: Think of the AI's brain as a sound mixer with many sliders. Some sliders control the "visual evidence" (the picture), and others control the "language priors" (the memory).
The Action: They created a method to manually turn up the volume on the sliders controlling the specific parts of the image being counted (like the wheels or ears) and turn down the volume on the background noise or the distracting "memory" voice.
The Outcome: When they applied this "volume control" while the AI was answering, it suddenly got much better at counting the edited images. It improved accuracy by up to 8%. It didn't need to be retrained or taught new facts; it just needed to be told, "Hey, look harder at what you are actually seeing right now."

Summary

In short, this paper shows that even the smartest AI models can be "blind" to reality if their internal knowledge is too strong. They built a test to prove this, found that the models were looking at the right things but not listening to them, and fixed it by giving the visual evidence a louder voice in the AI's decision-making process.

Technical Summary: CounterCount: A Diagnostic Framework for Counting Bias in Vision Language Models

Problem Statement
Vision-Language Models (VLMs) have demonstrated strong capabilities in multimodal reasoning, largely by leveraging Large Language Models (LLMs) as textual backbones. However, a critical ambiguity remains: when a VLM answers a visual question, is the prediction grounded in the specific visual evidence, or is it driven by statistical regularities and "canonical priors" learned from text corpora? This issue is particularly acute in counting tasks. While counting appears simple, it requires models to identify entities, distinguish them from distractors, and aggregate visual evidence. Current VLMs often struggle with counting, and prior analyses suggest that even when relevant objects receive attention, models may fail to utilize that evidence correctly, defaulting instead to prototypical knowledge (e.g., assuming a rabbit has two ears regardless of the image content). Existing benchmarks often conflate counting errors with other factors like prompt bias or semantic ambiguity, making it difficult to isolate failures driven specifically by the conflict between visual evidence and language priors.

Methodology
To address this, the authors introduce CounterCount, a diagnostic framework designed to isolate conflicts between visual evidence and object-level count priors.

Dataset Construction:
- Counterfactual Pairs: The dataset consists of 168 paired images (336 total) spanning 10 semantic categories (e.g., mammals, transportation, currency). Each pair includes a "factual" image with canonical object attributes and a "counterfactual" (CF) image where a count-relevant attribute (e.g., number of ears, wheels, or blades) is systematically altered using image-editing models (Nano-Banana Pro).
- Localization: Unlike standard benchmarks, CounterCount provides region-level annotations. For each CF image, the authors generate segmentation masks and bounding boxes (BBs) for the modified attributes. These are mapped to the visual token grid of the vision encoder to create "Mask-BB" regions, isolating the specific visual tokens required for the count.
- Evaluation Formats: Each instance is evaluated using Open-Ended (OE) prompts and Multiple-Choice Questions (MCQ) with balanced distractors to distinguish between generic miscounting and prior-driven errors.
Diagnostic Analysis:
- The authors evaluate recent open-weight VLMs (Qwen3-VL, Gemma3) and a closed-source model (Claude-haiku-4.5).
- They analyze attention mechanisms to determine if failures stem from a lack of localization or a failure to weight localized evidence. Results indicate that while models often attend to the relevant regions, they underweight these tokens relative to the background or irrelevant context during reasoning.
Intervention Strategy:
- The paper proposes a Unified Inference-Time Attention Modulation strategy. This method operates on the raw attention logits of the LLM backbone before the softmax layer.
- It allows for the amplification ( $\alpha > 1$ ), suppression ( $0 < \beta < 1$ ), or masking ( $\beta = 0$ ) of specific visual token sets (Target) while simultaneously dampening background tokens.
- This formulation unifies various interventions (amplification, suppression, masking) into a single continuous logit-modulation space, allowing for targeted rebalancing of modality contributions without retraining.

Key Contributions

CounterCount Dataset: A controlled diagnostic dataset featuring paired factual and counterfactual images with localized annotations (masks and bounding boxes) specifically designed to test visual grounding against canonical priors.
Diagnostic Insight: Empirical evidence showing that VLMs exhibit strong performance on factual images but suffer consistent accuracy degradation on counterfactual images. This confirms a reliance on object-level priors even when contradictory visual evidence is present.
Attention Analysis: Demonstration that counting failures are not solely due to missing visual evidence or poor localization; rather, models attend to relevant regions but fail to sufficiently weight them against language priors and background noise.
Training-Free Intervention: A unified attention modulation strategy that improves counterfactual counting accuracy by up to 8% across multiple VLMs by reweighting count-relevant visual tokens at inference time.

Results

Performance Drop: On factual images, models achieve high accuracy (e.g., Qwen3-VL-32B: ~87%, Gemma3: ~66–81%). However, on counterfactual images, accuracy drops significantly (e.g., Qwen3-VL-32B: ~48%, Gemma3: ~34–38%), with high rates of bias toward the canonical count.
Modulation Efficacy: Applying attention modulation (specifically amplifying target tokens and dampening/masking background tokens) improves counterfactual accuracy. For instance, Qwen3-VL-32B saw an 8.26% improvement in Open-Ended accuracy.
Bias Reduction: The intervention successfully shifts predictions away from canonical priors toward visual evidence, reducing the bias rate in several configurations.
Generalization: The findings hold across different model scales (4B to 32B) and both open-weight and closed-source architectures, suggesting a systemic issue in current VLM reasoning.
Factual Stability: The modulation strategies generally do not degrade performance on factual images and, in some cases, improve it, indicating the method enhances focus on semantically relevant tokens without harming general understanding.

Significance
The paper claims that CounterCount exposes a persistent reasoning failure in current VLMs where counting predictions are driven by prior knowledge rather than visual evidence, even when the visual evidence is explicit and localized. The significance lies in:

Isolating the Failure Mode: Providing a controlled setting to distinguish between perceptual limitations and modality imbalance (the underuse of visual evidence in favor of linguistic priors).
Mechanistic Insight: Showing that failures occur not because models cannot "see" the count, but because they fail to prioritize the visual tokens associated with the count during the reasoning process.
Design Guidance: Offering a training-free, inference-time mechanism to probe and mitigate these biases, providing a pathway for designing future VLMs with stronger visual grounding. The authors note that while the intervention improves accuracy, it does not fully eliminate the bias, suggesting that deeper architectural changes may be required for robust visual grounding.