💬 NLP

Fairness Evaluation and Inference Level Mitigation in LLMs

This paper proposes a dynamic, reversible, inference-time framework that mitigates fairness issues and behavioral inconsistencies in large language models by adaptively masking context-aware neuron activations, offering a flexible and memory-efficient alternative to static pruning or costly retraining methods.

Original authors: Afrozah Nadeem, Mark Dras, Usman Naseem

Published 2026-04-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Afrozah Nadeem, Mark Dras, Usman Naseem

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 Problem: The "Drunk" Conversation

Imagine you are talking to a very smart, well-read friend (the AI). At the start of the conversation, they are polite, fair, and sensible. But as you talk for a long time, something strange happens.

If you make a few subtle, slightly biased comments early on, or if the conversation drifts into a tricky topic, your friend starts to "catch" those biases. They might start repeating stereotypes, getting angry, or saying hurtful things they wouldn't have said at the beginning.

This is what happens with Large Language Models (LLMs). They are great at chatting, but in long conversations, their internal "memories" can get polluted. They start amplifying harmful ideas or becoming inconsistent, not because they are "evil," but because their internal gears (neurons) are getting stuck on the wrong tracks.

The Old Solutions: The "Brute Force" Approach

Previously, scientists tried to fix this in two main ways, both of which had big flaws:

The "Rewrite the Book" Method (Training): They tried to re-teach the AI from scratch with better data.
- Analogy: This is like hiring a new teacher for a school and spending years retraining them. It's expensive, slow, and once the school year starts, you can't easily change the teacher's mind if a new problem pops up.
The "Permanent Amputation" Method (Static Pruning): They tried to find the specific "bad" neurons in the AI's brain and cut them out forever.
- Analogy: Imagine your friend has a habit of making jokes about a specific topic. The old method says, "Let's surgically remove the part of their brain that makes jokes." The problem? That same part of the brain might also be needed for telling funny jokes about other things or for understanding complex history. By cutting it out, you make the friend less smart and less coherent, and you can't put the brain part back if the conversation changes.

The New Solution: The "Smart Traffic Cop"

The authors of this paper propose a new, clever way to handle this called Dynamic Neuron Suppression.

Instead of cutting out brain parts or retraining the AI, they built a real-time traffic cop that sits inside the AI's brain while it is talking.

Here is how it works, step-by-step:

1. The Radar (Behavioral Detection)

The traffic cop is constantly scanning the conversation. It asks: "Is the AI starting to say something biased or harmful right now?"

If the AI is being polite, the cop does nothing.
If the AI starts drifting toward a stereotype (e.g., "Old people are useless"), the cop raises a red flag.

2. The Detective (Bias Neuron Identification)

Once the flag is raised, the cop doesn't just guess. It investigates which specific gears (neurons) are causing the trouble.

Analogy: It's like a mechanic listening to a car engine. They don't just say "the engine is bad." They pinpoint exactly which piston is misfiring. Crucially, they check if that piston is misfiring only because of the current conversation, or if it's a permanent defect.

3. The Memory Check (Concept Testing)

This is the paper's secret sauce. The cop checks the AI's "short-term memory."

Analogy: Imagine your friend is telling a story. Sometimes they get carried away and say something rude. The cop checks: "Did they say this because they are naturally rude, or because the conversation got heated?"
If the AI is just having a temporary "bias moment" due to the context, the cop knows to be gentle. If the AI is genuinely stuck in a biased loop, the cop knows to act stronger.

4. The Dimmer Switch (Dynamic Masking)

Instead of cutting the wire (permanent pruning), the cop uses a dimmer switch.

When the AI starts to say something biased, the cop gently dims the volume of those specific neurons.
When the conversation shifts back to a safe topic, the cop brightens the switch again, turning the neurons back on.
Result: The AI stays smart and coherent because its brain parts are still there, just temporarily quieted when they are being dangerous.

Why This is a Game Changer

The paper tested this on many different languages (like Urdu, Punjabi, and English) and in long, multi-turn conversations.

It's Reversible: If the conversation changes, the AI can "wake up" the neurons it just quieted. It doesn't lose its intelligence.
It's Context-Aware: It knows the difference between a harmless joke and a harmful stereotype based on the flow of the chat.
It's Fast: It doesn't require retraining the AI. It happens instantly while the AI is typing, like a real-time filter.

The Bottom Line

Think of this new method as a smart, invisible guardian for AI conversations. It doesn't try to lobotomize the AI or rewrite its entire personality. Instead, it gently nudges the AI back on track the moment it starts to drift into bias, ensuring that even after hours of chatting, the AI remains fair, safe, and coherent.

It's the difference between firing a employee for one bad day (old method) and having a supportive manager who steps in to guide them back on track the moment they make a mistake (new method).

1. Problem Statement

Large Language Models (LLMs) often exhibit undesirable behaviors such as bias amplification, inconsistency, and the propagation of harmful stereotypes. While traditional mitigation strategies (e.g., Supervised Fine-Tuning, RLHF, data filtering) occur during training, they suffer from significant limitations:

Computational Cost: They require extensive retraining and data curation.
Irreversibility: Once deployed, changes are permanent and cannot adapt to new contexts.
Static Nature: They fail to address the dynamic, cumulative nature of bias in multi-turn conversations.

The authors argue that bias in LLMs is not a fixed property of a single output but a phenomenon that accumulates and re-emerges across dialogue turns due to context priming, anaphora, and memory carry-over. Existing inference-time methods (e.g., prompt engineering, static neuron pruning) are often too shallow, static, or irreversible, leading to semantic loss or reduced coherence when applied globally.

2. Methodology: Dynamic Neuron Suppression (DNS)

The paper proposes Dynamic Neuron Suppression, an inference-time framework that detects and selectively gates bias-carrying neuron activations in real-time without retraining the model. The framework operates through a four-stage pipeline:

A. Behavioral Detection (Stage 1)

Goal: Identify when biased behavior emerges in the model's output.
Mechanism: A turn-level bias score ( $S_t$ ) is computed using an LLM-as-a-Judge protocol (primary: GPT-3.5-Turbo; auxiliary: Claude).
Output: A quantitative signal indicating the degree of ideological polarization, stereotyping, or harmful framing in the current turn.

B. Bias Neuron Identification (Stage 2)

Goal: Trace the specific internal neurons responsible for the detected bias.
Mechanism: Uses Integrated Gradients to attribute the bias score ( $S_t$ ) to internal neuron activations.
Decomposition: Importance scores are split into two components:
- $m_{local}$ : Relevance to the current turn.
- $m_{carry}$ : Persistent influence from earlier dialogue turns (memory carry-over).
Aggregation: Scores are pooled across tokens, instances, and paraphrased instructions to identify robust "bias neurons."

C. Concept-Based Testing / Memory Consistency Probe (Stage 3)

Goal: Distinguish between transient bias spikes and persistent bias memory to avoid over-suppression.
Mechanism: A Memory Consistency Probe evaluates whether identified bias neurons preserve their activation patterns across turns. It uses contrastive queries to measure a memory score ( $C_t$ ), quantifying how strongly a neuron retains biased concepts over time.
Significance: This ensures the system only suppresses neurons when bias is actively persisting, preserving the model's ability to recall factual knowledge when bias is absent.

D. Dynamic Neuron Masking (Stage 4)

Goal: Modulate the influence of identified neurons during generation.
Mechanism: Instead of permanently removing neurons (static pruning), the framework applies a reversible gating mechanism.
- A gating coefficient $g^{(t)}_l = \sigma(\alpha S_t + \beta C_t)$ is calculated for each layer $l$ .
- This coefficient adaptively scales the activation of bias-sensitive neurons.
- If bias signals ( $S_t$ ) or memory persistence ( $C_t$ ) are high, the gate closes (suppressing the neuron). If bias is absent, the gate opens, restoring full model capacity.
Implementation: The intervention is applied exclusively to the Feed-Forward Network (FFN) neurons in the transformer decoder, leaving attention mechanisms and model parameters untouched.

3. Key Contributions

Reframing Bias as Dynamic: The paper formalizes bias as an accumulating, context-dependent phenomenon in multi-turn dialogues, rather than a static defect of single-turn outputs.
Inference-Time Reversibility: Introduces a lightweight, training-free framework that dynamically gates neurons. Unlike static pruning, it allows the model to recover its original reasoning capabilities when bias is not present.
Memory-Aware Mitigation: The novel Memory Consistency Probe distinguishes between local context bias and long-term memory carry-over, preventing the "over-suppression" that degrades coherence in long dialogues.
Multilingual & Multi-Turn Evaluation: Comprehensive evaluation across six demographic attributes (Gender, Race, Appearance, Disability, Religion, Age) and multiple languages (including low-resource Pakistani languages like Urdu, Punjabi, Sindhi) using benchmarks like FairMT-Bench, BBQ, and Political Compass Test (PCT).

4. Experimental Results

The framework was evaluated on Mistral-7B and DeepSeek-7B models against baselines including Prompt Engineering, Output Filtering, Vector Steering, and Static Neuron Pruning.

Bias Reduction: Dynamic Neuron Masking consistently achieved the lowest bias scores across all languages and tasks.
- Example: In the FairMT-Bench (Multi-turn), the proposed method reduced bias scores significantly (e.g., from ~42.1 to 21.1 for Mistral on Anaphora tasks) compared to static pruning.
- It outperformed static pruning, which often degraded performance in later turns due to accumulated context drift.
Preservation of Utility: Unlike static pruning, which often caused semantic loss and reduced coherence, the dynamic approach maintained:
- Knowledge Retention: High accuracy in factual recall.
- Faithfulness & Relevance: Responses remained contextually grounded.
- Fluency: No significant degradation in language quality.
Ablation Studies: Removing the Memory Probe or using static masking led to a sharp decline in coherence and an increase in bias rebound, confirming the necessity of the dynamic, memory-aware components.
Efficiency: The method adds minimal overhead (~1.3 ms per turn) compared to the baseline forward pass, making it suitable for real-time deployment.

5. Significance

This work represents a paradigm shift in LLM safety and fairness:

From Static to Dynamic: It moves beyond the "one-size-fits-all" static pruning approach, acknowledging that neurons can be beneficial in neutral contexts but harmful in biased ones.
Scalability: By operating at inference time without retraining, it offers a practical solution for deploying fair AI in diverse, real-world conversational settings where context evolves rapidly.
Interpretability: The method provides a transparent mechanism (neuron tracing) to understand why and when bias occurs, offering a path toward more controllable and accountable AI systems.

In conclusion, the authors demonstrate that dynamic, reversible, and memory-aware neuron masking is a superior strategy for mitigating bias in multi-turn conversations, effectively balancing fairness with the preservation of model utility and coherence.