Research map / Large language models / Language, safety and interpretability
LIDAO: Towards Limited Interventions for Debiasing (Large) Language Models
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.
Proposes LIDAO, an information-theoretic decoding-time debiasing framework that only requires each token to be independent of either the global property or the gender group (not both), plus eLIDAO for adversarial prompts using self-guidance from the LM's instruction-following ability.
Problem
LLM and LM text generation exhibits gender bias in global properties (sentiment, regard, toxicity), and existing debiasing methods impose constant constraints that degrade fluency, so the fairness-fluency trade-off is suboptimal.
Approach
Formulate global bias as mutual information between a global property and demographic group, and prove that requiring each token to be independent of either the property or the group (rather than both constantly) suffices to drive I(g;a) to zero; implement adaptively via min- or product-based losses with bias-tuning at decoding time, and extend to adversarial prompts by bounding bias with H(a(c,x)
Key results
- On GPT-2, min-based LIDAO reduces sentiment bias from 4.41 to 0.32 while keeping perplexity at 27.49 (vs. 22.17 for base GPT-2 and 32.32 for UDDIA).
- On OPT, LIDAO reduces toxicity bias from 0.78 to 0.00 with perplexity 37.17, and reduces regard bias from 11.58 to 1.09.
- Under the joint gender-with-prompt evaluation, eLIDAO attains lower bias than LIDAO in all nine scenarios, and top-two fluency in eight of nine.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply min-based LIDAO instead of UDDIA | GPT-2 · Sentiment, generated text | bias: 2.77 -> 0.32 | Helped |
| Apply prompt-aware eLIDAO instead of LIDAO | GPT-2 · Sentiment, joint prompt and generated text | bias: 2.11 -> 0.46 | Helped |
| Use min-based LIDAO instead of UDDIA | OPT · Sentiment, generated text | bias: 5.40 -> 0.03 | Helped |
| Use min-based LIDAO instead of UDDIA | OPT · Regard, generated text | bias: 1.81 -> 1.09 | Helped |
| Use min-based LIDAO instead of UDDIA | OPT · Toxicity, generated text | bias: 1.69 -> 0.00 | Helped |
| Use min-based LIDAO instead of UDDIA | GPT-2 · Regard, generated text | bias: 2.66 -> 2.15 | Helped |
| Use prod-based LIDAO instead of UDDIA | Falcon-7B-Instruct · Sentiment, generated text | bias: 4.54 -> 2.96 | Helped |
| Use prod-based LIDAO instead of UDDIA | Falcon-7B-Instruct · Regard, generated text | bias: 2.56 -> 2.59 | No clear effect |
| Use prod-based LIDAO instead of UDDIA | Falcon-7B-Instruct · Toxicity, generated text | bias: 14.94 -> 10.16 | Helped |
| Add prompt-aware reference guidance (eLIDAO) | LIDAO · Joint prompt and generated text, all three tasks and models | bias | Mixed |
| Use min-based rather than prod-based LIDAO (ablation) | GPT-2 · Sentiment, generated text | bias: 2.96 -> 0.32 | Helped |
| Use min-based rather than prod-based LIDAO (ablation) | OPT · Sentiment, generated text | bias: 0.16 -> 0.03 | Helped |
Surprising findings
- Instruction-tuned Falcon favored the product-based variant while GPT-2 and OPT favored the min-based variant.
- Bias measured over generated text alone underestimates bias: regard bias for OPT and toxicity bias for Falcon drop when the prompt is included in the gender attribute.
Limitations
- LIDAO reduces but does not eliminate bias, attributed to inexact mutual-information proxies.
- Only gender bias is studied; race and age are left for future work.
- Evaluation is monolingual and relies on algorithmic metrics (sentiment, regard, Perspective API) that may be context-dependent.
Experimental setup
Datasets: RealToxicityPrompts SST-5
Benchmarks: RealToxicityPrompts (adversarial prompt pairs)
Metrics: sentiment score non-negative regard toxicity (Perspective API) bias gap between gender groups perplexity (GPT-2 XL, OPT-13B, Falcon-40B)
Baselines: GeDi PPLM DExperts UDDIA
Builds on: UDDIA PPLM GeDi DExperts
Closest papers by meaning
- Evaluating the Efficacy of Prompting Techniques for Debiasing Language Model Outputs (Student Abstract) (AAAI 2024 · similarity 0.92)
- All Should Be Equal in the Eyes of LMs: Counterfactually Aware Fair Text Generation (AAAI 2024 · similarity 0.90)
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.89)
- Auto-Search and Refinement: An Automated Framework for Gender Bias Mitigation in Large Language Models (NeurIPS 2025 · similarity 0.89)
- Collapsed Language Models Promote Fairness (ICLR 2025 · similarity 0.88)
- When LLMs Meet Cunning Texts: A Fallacy Understanding Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.