QuantiBias: Benchmarking Quantization-Induced Bias in LLMs

2026-07-25 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

QuantiBias, a new benchmark, reveals that post-training quantization in large language models significantly increases bias in open-ended generation, a critical side effect missed by standard safety evaluations. While models like Qwen and Gemma maintain performance on refusal, over-refusal, and multiple-choice bias avoidance benchmarks (e.g., refusal at 0.98, BBQ accuracy near 0.97), their open-ended responses volunteer stereotypes in approximately one in four answers (~24% to ~27%) across eight languages. This "selective safety gap" is robust across two backbone models and five model families. The study found that quantizers prioritize capability data over bias-prevention signals. Furthermore, reasoning before answering can halve this bias in models like Qwen but shows no effect on Gemma. QuantiBias measures this by pairing a generative, multilingual stereotype probe with controls, contrasting builds with and without reasoning, and rating content severity (mean 1.83 on a 0-4 rubric).

Key takeaway

For Machine Learning Engineers deploying quantized large language models, you must re-evaluate your builds for open-ended bias, as standard short-form safety checks are insufficient. Your models can pass all conventional safeguards (e.g., refusal, multiple-choice bias avoidance) yet still volunteer stereotypes in nearly one in four open-ended responses. Implement generative, multilingual stereotype probes and assess content severity to catch this selective bias amplification before release, especially since reasoning may not universally mitigate the effect.

Key insights

Quantization measurably increases LLM bias in open-ended generation, undetected by standard safety checks.

Principles

• Quantization perturbs weights, causing disproportionate bias changes.
• Behaviors rarely reinforced in training degrade first.
• Standard quantizers under-protect bias-critical weights.

Method

QuantiBias uses a generative, multilingual stereotype probe, refusal/multiple-choice controls, reasoning contrast, and content severity rating to detect selective bias.

In practice

• Re-evaluate quantized builds for open-ended bias.
• Rate content severity of generated stereotypes.

Topics

• LLM Quantization
• AI Bias
• Generative AI Safety
• QuantiBias Benchmark
• Model Evaluation
• Multilingual Models

Code references

• ggml-org/llama.cpp

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• QuantiBias: Benchmarking Quantization-Induced Bias in LLMs — LLM quantization measurably increases bias in open-ended generation, a critical side effect missed by standard safety evaluations.
• ImplicitBBQ: Benchmarking Implicit Bias in Large Language Models through Characteristic Based Cues — LLMs show substantial implicit biases, particularly for caste, which current alignment and prompting strategies largely fail to resolve.
• FairCoder: Probing LLM Bias in High-Stakes Decision Making via Coding Tasks — LLMs exhibit social biases in code generation, especially in test case generation and less-explored attributes.
• Equilibrium Dynamics and Mitigation of Gender Bias in Synthetically Generated Data — Synthetic data bias exhibits equilibrium dynamics, requiring multidimensional evaluation beyond semantic similarity.
• Cross-Lingual Bias in Large Language Models: A Comparative Analysis of English and Swahili — Social biases in LLMs transform, not transfer, across languages, making English-only audits insufficient for multilingual deployment.
• From If-Statements to ML Pipelines: Revisiting Bias in Code-Generation — Current bias evaluation methods using simple conditionals dramatically underestimate real-world bias in code generation.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.