Language Models That Walk the Talk: A Framework for Formal Fairness Certificates

2025/05/19 by D.-J. Chen, Tobias Ladner, Chen, Danqing +5
Computer Science · Social Sciences · #Adversarial Robustness in Machine Learning #Artificial Intelligence (cs.AI) #Ethics and Social Impacts of AI #FOS: Computer and information sciences #Hate Speech and Cyberbullying Detection

paper · pdf · doi:10.48550/arxiv.2505.12767

openalex publication_date 2025/05/19 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

As large language models become integral to high-stakes applications, ensuring their robustness and fairness is critical. Despite their success, large language models remain vulnerable to adversarial attacks, where small perturbations, such as synonym substitutions, can alter model predictions, posing risks in fairness-critical areas, such as gender bias mitigation, and safety-critical areas, such as toxicity detection. While formal verification has been explored for neural networks, its application to large language models remains limited. This work presents a holistic verification framework to certify the robustness of transformer-based language models, with a focus on ensuring gender fairness and consistent outputs across different gender-related terms. Furthermore, we extend this methodology to toxicity detection, offering formal guarantees that adversarially manipulated toxic inputs are consistently detected and appropriately censored, thereby ensuring the reliability of moderation systems. By formalizing robustness within the embedding space, this work strengthens the reliability of language models in ethical AI deployment and content moderation.

Citations

• The Fifth International Verification of Neural Networks Competition (VNN-COMP 2024): Summary and Results
• First-Person Fairness in Chatbots
• Neural Network Verification with Branch-and-Bound for General Nonlinearities
• Towards Building a Robust Toxicity Predictor
• Harnessing Neuron Stability to Improve DNN Verification
• The Fourth International Verification of Neural Networks Competition (VNN-COMP 2023): Summary and Results
• Certifying LLM Safety against Adversarial Prompting
• Bias and Fairness in Large Language Models: A Survey
• Bias and Fairness in Large Language Models: A Survey
• A Survey on Fairness in Large Language Models
• Learning to Generate Equitable Text in Dialogue from Biased Training Data
• Large Language Models are not Fair Evaluators
• Marked Personas: Using Natural Language Prompts to Measure Stereotypes in Language Models
• A Survey of Safety and Trustworthiness of Large Language Models through the Lens of Verification and Validation
• Hybrid Zonotopes Exactly Represent ReLU Neural Networks
• POLAR-Express: Efficient and Precise Formal Reachability Analysis of Neural-Network Controlled Systems
• GPT-4 Technical Report
• Convex Bounds on the Softmax Function with Applications to Robustness Verification
• First Three Years of the International Verification of Neural Networks Competition (VNN-COMP)
• Provable Fairness for Neural Network Models using Formal Verification
• BERN-NN: Tight Bound Propagation For Neural Networks Using Bernstein Polynomial Interval Arithmetic
• CertiFair: A Framework for Certified Global Fairness of Neural Networks
• Algorithmic Fairness Verification with Graphical Models
• Probabilistic Verification of Neural Networks Against Group Fairness
• Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Loss
• Database Reasoning Over Text
• Unmasking the Mask -- Evaluating Social Biases in Masked Language Models
• Incorporating External Knowledge to Enhance Tabular Reasoning
• Beta-CROWN: Efficient Bound Propagation with Per-neuron Split Constraints for Complete and Incomplete Neural Network Robustness Verification
• Recent Advances in Adversarial Training for Adversarial Robustness
• Persistent Anti-Muslim Bias in Large Language Models
• Understanding the Behaviour of Contrastive Loss
• Fast and Complete: Enabling Complete Neural Network Verification with Rapid and Massively Parallel Incomplete Verifiers
• A Survey on Contrastive Self-supervised Learning
• Measuring and Reducing Gendered Correlations in Pre-trained Models
• RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models
• Justicia: A Stochastic SAT Approach to Formally Verify Fairness
• Verifying Individual Fairness in Machine Learning Models
• Unsupervised Learning of Visual Features by Contrasting Cluster Assignments
• Big Self-Supervised Models are Strong Semi-Supervised Learners
• Fairness without Demographics through Adversarially Reweighted Learning
• Toxicity Detection: Does Context Really Matter?
• Language Models are Few-Shot Learners
• Supervised Contrastive Learning
• StereoSet: Measuring stereotypical bias in pretrained language models
• NNV: The Neural Network Verification Tool for Deep Neural Networks and Learning-Enabled Cyber-Physical Systems
• Logical Natural Language Generation from Open-Domain Tables
• Improved Baselines with Momentum Contrastive Learning
• Learning Certified Individually Fair Representations
• Robustness Verification for Transformers
• A Simple Framework for Contrastive Learning of Visual Representations
• Perfectly Parallel Fairness Certification of Neural Networks
• Momentum Contrast for Unsupervised Visual Representation Learning
• Momentum Contrast for Unsupervised Visual Representation Learning
• A Comprehensive Survey on Transfer Learning
• A Comprehensive Survey on Transfer Learning
• Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
• The Impact of Data Preparation on the Fairness of Software Systems
• TabFact: A Large-scale Dataset for Table-based Fact Verification
• Achieving Verified Robustness to Symbol Substitutions via Interval Bound Propagation
• On Measuring and Mitigating Biased Inferences of Word Embeddings
• A Survey on Bias and Fairness in Machine Learning
• Is BERT Really Robust? A Strong Baseline for Natural Language Attack on Text Classification and Entailment
• Measuring Bias in Contextualized Word Representations
• Identifying and Reducing Gender Bias in Word-Level Language Models
• A Machine Learning Approach to Comment Toxicity Classification
• A Theoretical Analysis of Contrastive Unsupervised Representation Learning
• BERT has a Mouth, and It Must Speak: BERT as a Markov Random Field Language Model
• EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks
• Probabilistic Verification of Fairness Properties via Concentration
• Efficient Neural Network Robustness Certification with General Activation Functions
• Formal Verification of Neural Network Controlled Autonomous Systems
• Understanding the Origins of Bias in Word Embeddings
• Learning Gender-Neutral Word Embeddings
• Contextual Augmentation: Data Augmentation by Words with Paradigmatic Relations
• Generating Natural Language Adversarial Examples
• Deep contextualized word representations
• A comparative study of fairness-enhancing interventions in machine learning
• Attention Is All You Need
• Automated Hate Speech Detection and the Problem of Offensive Language
• Reluplex: An Efficient SMT Solver for Verifying Deep Neural Networks
• Ex Machina: Personal Attacks Seen at Scale
• Safety Verification of Deep Neural Networks
• Equality of Opportunity in Supervised Learning
• Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings
• SQuAD: 100,000+ Questions for Machine Comprehension of Text
• Character-level Convolutional Networks for Text Classification
• Explaining and Harnessing Adversarial Examples
• Going Deeper with Convolutions
• Intriguing properties of neural networks
• Distributed Representations of Words and Phrases and their Compositionality
• Efficient Estimation of Word Representations in Vector Space
• ImageNet classification with deep convolutional neural networks

Related

• A Survey of Safety and Trustworthiness of Large Language Models through the Lens of Verification and Validation
• Fortifying Ethical Boundaries in AI: Advanced Strategies for Enhancing Security in Large Language Models
• Towards Safer AI Moderation: Evaluating LLM Moderators Through a Unified Benchmark Dataset and Advocating a Human-First Approach
• Safeguarding Large Language Models: A Survey
• Towards Trustworthy AI: A Review of Ethical and Robust Large Language Models