💬 NLP

Neuron-Level Interventions for Gendered and Gender-Neutral Generation in Language Models

This paper introduces a neuron-level intervention method that identifies and manipulates gender-specific neurons, primarily concentrated in the earliest layers of language models, to achieve precise control over feminine, masculine, and gender-neutral generation while mitigating bias and preserving semantic meaning.

Original authors: Zhiwen You, Nafiseh Nikeghbal, Jana Diesner

Published 2026-06-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhiwen You, Nafiseh Nikeghbal, Jana Diesner

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 a large language model (like the ones that power chatbots) as a massive, bustling factory. Inside this factory, there are millions of tiny workers called neurons. When you ask the factory to write a sentence, these workers light up and pass messages to each other to build the final output.

The problem is that even when you ask for a neutral story (e.g., "The doctor treated the patient"), the factory sometimes accidentally adds gender bias, turning the doctor into a "he" or a "she" based on old stereotypes.

This paper is like a team of detectives who went inside the factory to find out exactly which workers are responsible for deciding whether a sentence sounds "masculine," "feminine," or "gender-neutral."

Here is the breakdown of their investigation using simple analogies:

1. The New Map: Three Colors, Not Just Two

Most previous studies only looked at the factory through a "binary" lens: Is the worker making things Red (Masculine) or Blue (Feminine)? They often ignored the Green (Gender-Neutral) workers.

The authors realized that in the real world, we need all three colors. They wanted to find the specific workers who handle:

Red: Words like he, man, father.
Blue: Words like she, woman, mother.
Green: Words like they, person, parent.

2. The Detective Work: Finding the "Specialists"

The researchers developed a new method to identify these specific workers. They didn't just guess; they used a "scorecard" to see which neurons fired up most strongly when the model was talking about a specific gender, while ignoring the others.

The Discovery:
They found that these gender-specialist workers aren't scattered randomly throughout the factory. Instead, they are huddled together in the very first few rooms (the early layers) of the factory.

Analogy: Imagine a relay race. The decision about "who is running this race" (the gender) is made almost immediately at the starting line, not at the finish line. Once the first few workers decide "This is a 'she' story," the rest of the factory just follows that lead.

3. The Experiment: The "Mute Button" Test

To prove they found the right workers, the researchers performed a "controlled experiment." They took a sentence and asked the model to rewrite it in a different gender (e.g., change a story about a "man" to a story about a "woman").

The Baseline: Without help, the model often gets confused or leaks words from the wrong category (e.g., saying "The woman firefighter" but then using "he" later).
The Intervention: The researchers used a "mute button" (masking) to silence all the workers except the ones responsible for the target gender.
- If they wanted a Feminine story, they silenced the "Masculine" and "Neutral" workers and let only the "Feminine" workers speak.
- Result: The model became much better at sticking to the requested gender. It was like turning off the background noise so the specific voice could be heard clearly.

4. The Results: Precision and Quality

The paper claims their method is better than previous attempts because:

Less Leaking: It stops the model from accidentally mixing genders (e.g., saying "she" when you asked for "they").
Meaning Preserved: The story didn't lose its original meaning; it just changed the gender labels correctly.
Efficiency: Because they found that the "gender workers" are concentrated in the early layers, they didn't have to shut down the whole factory, just a few specific rooms.

5. The Datasets: The "Training Grounds"

To test this, the authors built two new "training grounds" (datasets) filled with thousands of sentences.

One dataset was created by taking existing gendered sentences and rewriting them to be neutral.
The other was built from scratch using a dictionary of gendered and neutral terms to ensure every sentence was perfectly labeled as Red, Blue, or Green.
Humans checked these sentences to make sure they were grammatically correct and actually matched the intended gender.

Summary

In short, this paper says: "We found the specific switches in the AI's brain that control gender. They are mostly located at the very beginning of the thinking process. By flipping these specific switches and silencing the others, we can force the AI to write in a specific gender (male, female, or neutral) without breaking the sentence or losing the original meaning."

The authors provide the code and the datasets so others can try this "switch-flipping" technique themselves.

Technical Summary: Neuron-Level Interventions for Gendered and Gender-Neutral Generation in Language Models

Problem Statement
Large Language Models (LMs) frequently generate gendered language and reinforce stereotypes, even when prompted with neutral inputs. While prior research has extensively examined gender bias through a binary lens (feminine vs. masculine), there is limited understanding of how gender-neutral forms (e.g., singular "they," neutral job titles) are encoded within model representations. Furthermore, existing mitigation strategies often operate at the inference level (e.g., few-shot prompting, structured constraints) without explicitly locating the internal mechanisms responsible for gendered behavior. The specific question addressed is whether gender-related signals can be localized to specific neurons within the feed-forward networks (FFN) of Transformers and whether these neurons can be causally manipulated to steer generation toward feminine, masculine, or gender-neutral forms while preserving semantic meaning.

Methodology
The authors propose a neuron-level intervention framework involving three primary stages:

Neuron Activation and Identification:
The study focuses on the intermediate neurons within the Multi-Layer Perceptron (MLP) blocks of Transformer models. For three gender categories ( $g \in \{masculine, feminine, neutral\}$ ), the authors calculate activation statistics for every neuron across gender-specific text corpora. To identify neurons strongly tied to a specific gender while minimizing overlap, they introduce a Combined Exclusivity Score. This score aggregates three complementary measures for each neuron $j$ in layer $l$ :
- Effect Size (Cohen's $d$ ): The standardized difference in mean activations between the target gender and all other genders.
- Log-Odds Difference ( $\Delta$ ): The relative likelihood of positive activation for the target gender versus others.
- Relative Mean Difference ( $r$ ): The proportional difference in activation magnitude.
  A neuron is selected for a specific gender only if it achieves the highest exclusivity score for that category and meets predefined thresholds across all three metrics.
Controlled Generation via Masking:
To validate causality, the authors perform inference-time interventions. They generate "baseline" outputs and "masked" outputs. In the masked condition, they apply a "keep-only" strategy: for a target gender $g$ , neurons identified for $g$ remain active, while neurons identified for the other two genders are masked (deactivated). This tests whether the identified neurons are sufficient to steer the model toward the target gender form.
Evaluation Protocol:
The effectiveness of the intervention is evaluated using two datasets: GCGender (derived from existing binary datasets with added neutral versions) and InclusiveGender (curated using inclusive language dictionaries). Evaluation metrics include:
- Keyword-Based Analysis: Measuring the change in the ratio of gender-specific terms (masculine, feminine, neutral) in the output relative to the baseline.
- Human Evaluation: Assessing "Overall Idea Preservation" (semantic retention) and "Target Gender Realization" (correctness of gendered wording).

Key Contributions

Tripartite Gender Analysis: The work extends prior binary gender studies by explicitly identifying and controlling neurons for three categories: feminine, masculine, and gender-neutral.
Novel Identification Metric: The introduction of a combined exclusivity scoring system to isolate neurons that are unambiguously tied to a single gender category, addressing the challenge of overlapping activations.
New Datasets: The curation of InclusiveGender (8,600 sentences per category) and the expansion of GCGender to include gender-neutral variants, validated through human annotation for high quality and low ambiguity.
Causal Validation: A rigorous protocol demonstrating that masking specific neurons causally alters gendered output while maintaining semantic integrity.

Results
Experiments conducted on Llama-3.1-8B and Qwen2.5-7B yielded the following findings:

Layer Distribution: Gender-specific neurons are not evenly distributed. They concentrate heavily in the earliest layers (typically layers 0–3) of the models, with significantly smaller contributions from middle and later layers. This contrasts with competing methods (LAPE, sNeuron-TST), which tend to identify neurons in later layers.
Control Precision: The proposed method achieves more precise gender control than existing baselines. When targeting a specific gender, the method successfully preserves or increases the ratio of target gender terms while suppressing non-target terms. Competing methods exhibited higher "leakage," often inflating unrelated gender ratios (e.g., increasing neutral terms when targeting masculine).
Quality Trade-offs: Human evaluation showed that the intervention method significantly improves "Target Gender Realization" compared to baselines. While there was a minor, annotator-dependent decrease in "Overall Idea Preservation" (attributed to subjective judgments of phrasing changes), the core meaning was largely retained.
Dataset Sensitivity: The study noted that smaller training sets (as in GCGender) can lead to noisier neuron identification, particularly for the neutral category, suggesting that data volume impacts the precision of exclusivity filtering.

Significance and Claims
The paper claims to provide a simple yet effective approach for understanding how gender is encoded in LMs. By demonstrating that gender representations are localized to a small fraction of neurons (often <0.5% of total parameters) concentrated in early layers, the work offers a mechanistic explanation for gendered generation. The authors assert that their neuron-level intervention provides a clearer understanding of internal model behavior than output-only mitigation strategies. They position this work as a foundational step toward controlled gender intervention, offering a tool for both evaluating neuron-based bias mitigation and potentially reducing gender bias in future model deployments. The authors remain modest regarding limitations, acknowledging that their findings are currently restricted to 7–8B parameter models and that smaller datasets may introduce noise in neuron selection.