LLM compression harms knowledge retention and masks bias, study finds

A new research paper published on arXiv details the asymmetric harms that can arise from compressing large language models (LLMs). The study, which evaluated three LLMs across eleven compression methods, found that compression disproportionately reduces the retention of "head knowledge" compared to "tail knowledge." Additionally, compressed models often maintain high confidence in incorrect answers related to lost knowledge, and aggregate bias scores can mask significant, opposing shifts in stereotypical preferences across demographic subgroups. The findings underscore the necessity of granular evaluation for compressed models before deployment, as standard metrics like perplexity and accuracy fail to capture these nuanced behavioral changes. AI

IMPACT Highlights the need for granular evaluation of compressed LLMs to avoid hidden harms in knowledge retention and bias.

RANK_REASON Research paper published on arXiv detailing findings about LLM compression. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• accuracy
• arXiv
• head knowledge
• Hugging Face
• knowledge retention
• large language models
• LLMs
• perplexity
• social bias
• tail knowledge

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →