💬 NLP

Unveiling the "Fairness Seesaw": Discovering and Mitigating Gender and Race Bias in Vision-Language Models

This paper systematically investigates gender and race bias in Vision-Language Models by analyzing internal hidden states and probability distributions, discovering that bias often persists in confidence scores and specific residual streams, and proposes a post-hoc framework called RES-FAIR to mitigate these biases by recalibrating hidden state dynamics.

Original authors: Jian Lan, Udo Schlegel, Tanveer Hannan, Gengyuan Zhang, Haokun Chen, Thomas Seidl

Published 2026-02-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jian Lan, Udo Schlegel, Tanveer Hannan, Gengyuan Zhang, Haokun Chen, Thomas Seidl

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

The "Fairness Seesaw": Why AI Models are Secretly Biased (and How to Fix Them)

Imagine you are training a new assistant. You teach them everything about the world, but because they learned from the internet—which is full of stereotypes—they’ve picked up some bad habits.

Even if you ask this assistant a polite question like, "Who is more likely to be a great scientist?" and they answer, "Both men and women are equal," they might actually be "lying" to you. Deep down, in their "brain," they might still be 75% sure that the answer is "men."

This paper, "Unveiling the Fairness Seesaw," explores exactly this phenomenon in Vision-Language Models (the AI that can "see" images and "talk" about them).

1. The Discovery: The "Fairness Paradox"

The researchers found something sneaky called the Fairness Paradox.

Think of it like a person who says, "I don't believe in stereotypes," but when you look at their heart rate or their nervous fidgeting, you can tell they actually do.

The AI might give you a perfectly "fair" text answer (the surface level), but if you look at its confidence scores (the internal math), it is still heavily leaning toward biased answers. It’s saying the right thing, but it doesn't believe it.

2. The "Fairness Seesaw"

The researchers looked inside the AI's "brain" (its layers) and found that fairness is incredibly unstable. They called this the Fairness Seesaw.

Imagine a seesaw in a playground. On one side, you have "Fairness Knowledge" (the part of the AI that knows everyone is equal), and on the other, you have "Bias Knowledge" (the part that holds onto old stereotypes).

As information travels through the AI's layers:

The Middle Layers: The seesaw is mostly balanced. The AI is being "fair."
The Final Layers: Suddenly, the seesaw tilts violently! Just before the AI speaks, the "Bias" side slams down, and the "Fairness" side flies up. This is why the AI might think a man is a pilot, even if it eventually says "both are equal."

3. The Solution: RES-FAIR (The "Internal Filter")

How do you fix a seesaw that keeps tilting the wrong way? You can't easily retrain the whole AI—that's too expensive and difficult. Instead, the researchers created a "post-hoc" fix called RES-FAIR.

Think of RES-FAIR as a high-tech filter placed right at the end of the AI's thought process.

Instead of letting the AI's thoughts flow freely, RES-FAIR looks at the "residual streams" (the tiny streams of information moving through the brain). It identifies the "biased" stream and says, "Hey, that's a stereotype! Let's subtract that," and then it finds the "fair" stream and says, "Let's boost this part instead."

It’s like having a polite editor standing behind the assistant, catching biased thoughts before they ever reach the microphone.

4. Does it work?

The researchers tested this on some of the most advanced AI models (like LLaVA and Qwen).

The Result: The AI became much more honest. It didn't just say things were fair; its internal "confidence" actually aligned with fairness.
The Best Part: The AI didn't lose its intelligence. It could still answer general questions about the world (like "What color is this car?") just as well as before. It just became a much more socially responsible assistant.

Summary in a Nutshell

The Problem: AI models often "perform" fairness on the surface while remaining biased in their "inner thoughts."
The Cause: Fairness and bias fight like a seesaw inside the AI's layers.
The Fix: A mathematical filter (RES-FAIR) that identifies and removes biased "thought streams" without needing to rebuild the whole brain.

Technical Summary: Unveiling the “Fairness Seesaw”

1. Problem Statement

While Vision-Language Models (VLMs) have achieved significant success, their internal mechanisms regarding social biases (specifically gender and race) remain a "black box." Existing research often focuses on surface-level frequency (e.g., how often a model correctly identifies an attribute in an image) or text-only biases. This paper identifies two critical gaps in current VLM evaluation:

The Fairness Paradox: Models may generate "fair" text labels (e.g., "both genders are equal") while maintaining highly skewed, biased confidence scores (miscalibration) in their internal probability distributions.
Generative Bias: Previous methods (like image retrieval) fail to capture the open-ended, generative biases present when models are asked about general social categories.

The authors characterize the internal dynamics of these biases as a "Fairness Seesaw," where fairness knowledge fluctuates unstablely across transformer layers and residual streams.

2. Methodology

The authors approach the problem through two phases: Discovery (analysis) and Mitigation (intervention).

A. Discovery & Evaluation Framework:

New MCS Task: They propose a Multiple-Choice Selection (MCS) task using modified versions of the PAIRS and SocialCounterfactuals (SCF) datasets. Unlike previous work, they use prompts that treat images as references rather than absolute evidence, forcing the model to reveal its underlying societal "opinions."
Confidence Analysis: They measure not just the chosen label, but the KL Divergence of the model's probability distribution against a uniform distribution ( $P_{target} = \{0.5, 0.5\}$ ) to quantify how "certain" a model is about its biased or fair responses.

B. Mitigation (RES-FAIR):
The core contribution is RES-FAIR (RESidual Flow Adjustment for Inference Recalibration), a post-hoc framework that operates on the model's hidden states without requiring retraining.

Subspace Decomposition: Based on the theory that semantic concepts are linear directions in latent space, they decompose the residual streams ( $\Delta_1, \Delta_2$ ) of each layer into "fair" and "biased" components.
Orthogonal Projection: For a given hidden vector $v_i$ $v_{i}$ , the method:
1. Identifies the mean biased direction ( $v_{biased}$ ) and fair direction ( $v_{fair}$ ) by aggregating residuals from a dataset.
2. Uses an orthogonal projection matrix to annihilate the biased component ( $v_{debias} = v_i - P_{biased}v_i$ ).
3. Amplifies the fair component ( $v_{new} = \lambda_2 P_{fair} v_{debias}$ ) to stabilize the representation.

C. Training-based Baseline:
They also implement an Uncertainty-Enhanced DPO (Direct Preference Optimization) method that adds a KL divergence constraint to the loss function to align both preference and confidence.

3. Key Contributions

Systematic Discovery: First study to analyze both the response and the confidence levels (probability distributions) of SOTA VLMs regarding gender and race.
New Evaluation Protocol: Construction of new prompts and MCS tasks that allow for explicit evaluation of a model's general societal opinion.
The "Fairness Seesaw" Insight: Discovery that fairness knowledge peaks in intermediate layers and undergoes "knowledge erosion" in final layers, with conflicting information residing in different residual streams.
RES-FAIR Framework: A novel, post-hoc, and parameter-free method to disentangle and adjust biased residual flows.

4. Results

Bias Presence: Even SOTA models like LLaVA-NeXT and Qwen2.5-VL exhibit significant bias. When provided with a single-gender image, models' fairness levels drop drastically as they over-rely on visual cues to reinforce stereotypes.
Effectiveness of RES-FAIR:
- Fairness: Significantly increased the proportion of "Fair" responses across both gender and race categories.
- Calibration: Successfully reduced KL Divergence, meaning the models' confidence levels became more aligned with fair distributions.
- Performance Preservation: Unlike many debiasing methods, RES-FAIR maintained general reasoning abilities, showing negligible drops in standard VQA (Visual Question Answering) accuracy.
Comparison: RES-FAIR outperformed standard LoRA fine-tuning and Temperature Scaling (a common calibration method) in achieving both fairness and reliable confidence.

5. Significance

This work shifts the paradigm of VLM fairness research from "correctness" (did the model label the person correctly?) to "social ethics" (does the model hold biased views?). By uncovering the "Fairness Seesaw," the authors provide a mathematical lens to understand how social information is processed within the transformer architecture. Their post-hoc approach offers a highly efficient way to deploy safer, more ethical multi-modal models in real-world applications without the massive computational cost of retraining.