QuantiBias: Benchmarking Quantization-Induced Bias in LLMs
Almost every large language model that reaches a broad audience is quantized: trained in full precision, then compressed for efficiency. This step is assumed harmless and its safety is rarely re-checked. We find its principal side effect is increased bias that standard safety evaluation misses. Holding the model, its training, and the prompts fixed, a quantized model still refuses harmful requests, still avoids over-refusing benign prompts, and still selects the unbiased multiple-choice answer.
Researchers found that quantized large language models exhibit increased bias, with around 24-27% of open-ended answers containing stereotypes. QuantiBias, a benchmark, was developed to evaluate this bias. It tested two models, Qwen and Gemma, across eight benchmarks and found that quantization can introduce bias not detected by standard safety checks.
Summarised by netranta from News. Open the original for the full story.
Observations (0)
Log in to add an observation.
No observations yet — add the first.