Researchers have found that compressing large language models (LLMs) can lead to uneven effects on their performance. Specifically, compression tends to reduce the retention of 'head knowledge' more severely than 'tail knowledge'. This means that compressed models may become overconfident in their incorrect answers and exhibit shifts in social bias. The study used 11 different compression methods to evaluate three LLMs and found that aggregate metrics like perplexity and accuracy can mask these underlying issues. --- Why it matters: This research matters because it highlights the importance of granular evaluation of compressed models before deployment, which could have significant consequences for applications relying on AI. Source: https://arxiv.org/abs/2608.19670