What is evaluation of token-level vs sequence-level metrics in LLMs?

Token-level metrics evaluate individual predictions, while sequence-level metrics evaluate full outputs.

Token-level metrics (like perplexity) measure likelihood accuracy per token but ignore global structure. Sequence-level metrics (BLEU, ROUGE, BERTScore) evaluate full generated outputs and semantic similarity. In LLM evaluation, token-level metrics fail to capture reasoning quality, coherence, and factual correctness, making sequence-level and human-aligned metrics essential.