💬 NLP

LoRA for Gender-Inclusive Rewriting and Activation Steering for Counter-Narrative Generation

This paper introduces the IHLC system for the LT-EDI 2026 Shared Task, which combines LoRA fine-tuning for gender-inclusive rewriting and a compute-efficient activation steering technique using PCA-derived principal directions for counter-narrative generation, achieving high scores while analyzing key limitations like semantic drift and bias leakage.

Original authors: Akhil Rajeev P, Manoj Balaji J

Published 2026-07-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Akhil Rajeev P, Manoj Balaji J

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 are talking to a very smart, very well-read robot that has read almost every book on the internet. This robot is great at writing stories and answering questions, but it has a problem: it learned from human books, and sometimes those books contain old-fashioned, unfair ideas about boys and girls. If you ask the robot to write a story about a "fireman," it might stubbornly insist on that word, even though "firefighter" is better. If you ask it to argue against a mean statement like "girls are bad at math," it might accidentally agree with the mean idea or sound too robotic. This is the world of Artificial Intelligence (AI) and Natural Language Processing (NLP), where scientists try to teach computers to speak more fairly and inclusively. The big challenge is figuring out how to tweak the robot's brain so it naturally chooses kind, neutral words without having to rebuild its entire brain from scratch.

In this paper, a team of researchers named IHLC tried to solve this problem for a special contest called LT-EDI 2026. They focused on two main tasks: first, simply rewriting biased sentences to be fair (like changing "fireman" to "firefighter"), and second, creating a "counter-narrative"—a polite, persuasive response to a mean or biased statement. For the first task, they used a standard method called LoRA, which is like giving the robot a small, specialized cheat sheet to help it learn new rules quickly. But for the second task, they tried something much more experimental and clever called Activation Steering.

Think of the robot's brain as a giant, complex machine with thousands of gears turning inside. Usually, to change how the machine works, you have to take it apart and replace the gears (this is called "fine-tuning"). But the researchers asked: "What if we could just give the machine a gentle nudge while it's running?" They found a specific "nudge" direction by comparing how the robot's brain reacts to a mean sentence versus a kind one. They calculated the difference between these two reactions and created a "steering vector"—imagine it as a magnetic force that pushes the robot's thoughts toward kindness and inclusivity. They injected this force into the robot's brain while it was writing, without changing any of its original gears.

The results were a mix of success and interesting glitches. For the simple rewriting task, their system did very well, scoring 80.00% and ranking 3rd out of 9 teams. For the harder task of writing counter-narratives, they scored 78.12% and ranked 6th out of 7 teams. The system was great at being polite and understanding the context, but sometimes it got a little too excited by the "nudge." When they pushed the steering too hard, the robot started repeating itself, fusing words together (like "exhibitimpulsiveness"), or drifting away from the original meaning to give a long lecture instead of a direct answer.

The researchers found that while this "nudge" method is a powerful and efficient way to make AI more inclusive, it's not perfect. It sometimes leaves behind tiny bits of the original bias, or it changes the sentence so much that it loses its original flavor. They suggest that in the future, they might need to combine this "nudge" with other tools to keep the robot's answers both fair and faithful to the original question. It's a promising step toward making AI a kinder conversationalist, but it shows that even with a gentle push, the robot still needs careful guidance to get it just right.

Technical Summary: IHLC at LT-EDI 2026

Problem Statement
The paper addresses the challenge of generating gender-inclusive language and counter-narratives to biased statements, a core objective of the LT-EDI 2026 shared task. The task is divided into two subtasks: (A) rewriting gendered or biased sentences into gender-neutral variants while preserving meaning, and (B) generating empathetic, persuasive counter-narratives to overtly biased statements. The authors note that while standard fine-tuning is effective, there is a need for compute-efficient methods to adjust model behavior without expensive weight updates, particularly for counter-narrative generation where balancing politeness, coherence, and factual correction is complex.

Methodology
The IHLC submission employs distinct strategies for the two subtasks, utilizing the Gemma-3-4B-it model as the foundation.

Subtask A (Gender-Inclusive Rewriting): The team utilized standard Low-Rank Adaptation (LoRA) fine-tuning. This approach involved training the model on parallel datasets of gendered and neutral sentence pairs to learn direct mapping between biased and inclusive terminology.
Subtask B (Counter-Narrative Generation): The core contribution of this paper is a compute-efficient activation-steering approach combined with constrained prompting.
- Activation Steering: Drawing on representation engineering protocols (Zou et al., 2023; Turner et al., 2023), the authors derived a steering direction by performing Principal Component Analysis (PCA) on the difference vectors of hidden activations between biased sentences and their neutral counterfactuals. This steering vector was injected into a specific intermediate layer (Layer 16) of the model during inference via a forward hook. The injection strength was controlled by a coefficient $\alpha$ .
- Prompt Engineering: To ensure the output met evaluation criteria, the system employed two prompt templates: a "DEI prompt" for flexible, explanatory outputs and a "Strict prompt" for deterministic, single-line outputs to avoid penalization by automated judges.
- Decoding: The system used a mixture of greedy decoding and low-temperature sampling with repetition penalties to mitigate text degeneration and looping, which were observed when steering coefficients were too high.

Key Results
The system was evaluated using a hybrid framework involving an LLM-as-a-judge and human oversight.

Subtask A: The LoRA-based approach achieved an 80.00% average score, ranking 3rd out of 9 participating teams. The score was consistent across Gender Assumption removal (GA), Gender Neutrality (GN), and Quality & Relevance (QR).
Subtask B: The activation-steering method achieved an average score of 78.12%, ranking 6th out of 7 participants.
- Strengths: The system excelled in Politeness & Respectfulness (PR: 84.89%) and Contextual Counter-Narrative Coherence (CCNC: 84.79%).
- Weaknesses: The Quality & Relevance (QR) score was significantly lower (64.68%), indicating issues with semantic preservation.

Failure Mode Analysis
Through manual evaluation of 47 outputs, the authors identified five primary failure modes inherent to the activation-steering approach:

Faithfulness–Edit Trade-off (53%): The model often replaced original sentence structures rigidly (e.g., defaulting to "People of all gender identities...") rather than making minimal edits, altering the semantic domain (e.g., changing "science" to "STEM fields").
Layer-Specific Sensitivity/Entanglement (23%): Steering at a single layer failed to cleanly neutralize concepts, often transforming negative traits into related but still biased concepts (e.g., "naive" becoming "innocence and vulnerability").
Residual Bias Leakage (11%): Because steering relied on final-token representations, bias encoded in earlier context tokens sometimes persisted, leading to the neutralization of the subject but the retention of the harmful premise.
Over-Steering Instability (9%): High steering coefficients caused the model to abandon the input's syntactic structure entirely, shifting into generalized social commentary rather than specific counter-narratives.
Repetition and Degeneration (4%): Strong perturbations occasionally disrupted token boundaries, causing fused words or looping text.

Significance and Claims
The paper positions its work as a demonstration of both the promise and limitations of activation steering for gender-inclusive generation.

Promise: The method successfully generates polite, context-aware counter-narratives without the computational cost of full fine-tuning, achieving competitive results in fairness metrics (PR and CCNC).
Limitations: The authors modestly claim that while activation steering is a viable alternative to weight updates, it currently struggles with semantic preservation (QR) and requires precise tuning to avoid instability. They note that single-layer steering risks residual bias leakage and that the approach is currently limited to English, restricting cultural generalizability.
Future Direction: The authors suggest that future improvements require integrating semantic-preservation constraints (such as contrastive loss), human-in-the-loop calibration for steering directions, and extension to multilingual tracks.

The paper concludes that while activation steering offers a lightweight mechanism for bias mitigation, it is not yet a complete replacement for weight-updating methods like Instruction Finetuning, particularly when high fidelity to the original text structure is required.

Technical Summary: IHLC at LT-EDI 2026

More like this