IHLC@LT-EDI 2026: Steering Toward Inclusivity - A Representation Engineering for Gender-Neutral Rewriting

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, short

Summary

The IHLC team submitted an English-only system to the LT-EDI ACL 2026 Shared Task on Gender-Inclusive Language Generation and Counterfactual/Counter Narrative Generation. This system employs an activation steering approach combined with carefully engineered prompt templates to generate gender-neutral rewrites and empathetic counter-narratives. For Task A, Gender Inclusive Language Generation, the system achieved an average score of 80.00% and ranked 3rd. In Task B, Counter Narrative Generation, it scored an average of 78.12% and secured the 6th rank. The submission details the system's design, experimental setup, and the shared task's evaluation protocol. It also includes an analysis of observed strengths and failure modes from both automatic and human-checked evaluations, suggesting areas for future improvement.

Key takeaway

For NLP Engineers developing language models for inclusive communication, this work highlights effective strategies. You should consider integrating activation steering and carefully engineered prompt templates into your systems. These techniques can significantly improve your model's ability to generate gender-neutral rewrites and empathetic counter-narratives, as demonstrated by achieving 80.00% in gender-inclusive generation. Prioritize evaluating your system with both automatic and human checks to ensure robust performance.

Key insights

Activation steering and prompt engineering can generate gender-neutral and empathetic language.

Principles

• Activation steering guides language model output.
• Engineered prompts enhance specific text generation.
• Combine automatic and human evaluation for language tasks.

Method

The system combines an activation steering approach with engineered prompt templates to produce gender-neutral rewrites and empathetic counter-narratives.

In practice

• Implement activation steering for controlled text generation.
• Develop tailored prompt templates for inclusive language.
• Benchmark language models on gender-neutral rewriting tasks.

Topics

• Gender-Inclusive Language
• Counter Narrative Generation
• Activation Steering
• Prompt Engineering
• NLP Bias Mitigation
• LT-EDI Shared Task

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• Want Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation — Activation steering on early LLM layers improves synthetic data diversity and downstream performance for low-resource languages.
• A Critical Discourse Analysis of Gender Representation in Software Engineering Education Videos on YouTube — Online software engineering tutorials perpetuate a "hidden curriculum" that associates technical agency with masculinity, creating a symbolic barrier.
• Greater accessibility can amplify discrimination in generative AI — Voice interfaces for LLMs introduce new discrimination pathways via paralinguistic cues, amplifying biases beyond text.
• Refusal Lives Downstream of Persona in Chat Models — Compliant model persona directly gates refusal behavior in instruction-tuned chat models at late-layer expression stages.
• Structured Intent as a Protocol-Like Communication Layer: Cross-Model Robustness, Framework Comparison, and the Weak-Model Compensation Effect — Structured intent representations significantly enhance goal alignment and reduce variance across diverse AI models and languages.
• Endogenous Resistance to Activation Steering in Language Models — Large language models can internally monitor and resist task-misaligned activation steering, recovering mid-generation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.