Research map / Large language models / Language, safety and interpretability

Evaluating the Efficacy of Prompting Techniques for Debiasing Language Model Outputs (Student Abstract)

AAAI 2024 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Evaluates zero-shot prompting techniques (instruction, role/persona, zero-shot CoT, self-debias, and a proposed iterative Implicative Introspection) for reducing stereotypical bias in LLM text generation on StereoSet.

Problem

LLMs inherit social biases from training data, and retraining to debias is costly; it is unclear whether structured prompting techniques can reduce bias in generated text.

Approach

Adapts existing prompting strategies to debiasing and proposes Implicative Introspection, an iterative method where the model critiques its own output ("What is the implication of the given text?") and then regenerates an unbiased text conditioned on that critique.

Key results

• Reasoning-based prompts (Zero-Shot CoT, Implication) generally reduce StereoSet bias scores most, but Implication yields low LM scores indicating fluency issues.
• LLaMA 2 (13B) with self-debiasing (SDB) shows the largest overall improvement, with overall SS dropping from 65.78 to 60.82.
• GPT-2 Small overall SS falls from 60.42 to 57.78 with Implication, and GPT-2 Large falls from 63.93 to 55.79 with Implication.

Measured effects reported in the paper

Surprising findings

• Zero-Shot CoT improved LLaMA 2 (13B) overall bias score to 68.45 from 65.78 baseline, i.e., worse, while Implication and SDB reduced bias.
• Implicative Introspection reduces bias scores but at the cost of language modeling quality, so simpler Instruction prompts are sufficient for fluency.

Limitations

• Evaluation limited to a single dataset (StereoSet) and zero-shot setting.
• Implicative Introspection causes fluency degradation (low LM score).
• Only four models evaluated, with small-scale experiments as a student abstract.

Experimental setup

Datasets: StereoSet

Benchmarks: StereoSet

Metrics: SS (stereotype score) LM score ICAT score

Baselines: Self-Debias (SDB) Instruction prompting Zero-Shot CoT Role/Persona prompting

Builds on: self-debiasing chain-of-thought prompting counterfactual data augmentation

Closest papers by meaning

• LIDAO: Towards Limited Interventions for Debiasing (Large) Language Models (ICML 2024 · similarity 0.92)
• Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.90)
• Causal Prompting: Debiasing Large Language Model Prompting Based on Front-Door Adjustment (AAAI 2025 · similarity 0.89)
• Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.88)
• All Should Be Equal in the Eyes of LMs: Counterfactually Aware Fair Text Generation (AAAI 2024 · similarity 0.88)
• Debiasing Algorithm through Model Adaptation (ICLR 2024 · similarity 0.88)