MultiLexNorm++ (LLM Pipeline) - Japanese: leaderboard

Metric: Error reduction rate (%; (correct minus wrong normalizations) / words that need normalization, so 0 equals leaving the text unchanged and a negative score means more wrong than right edits; the floor is -100 divided by the share of words needing normalization; 8-shot prompting, mean of three runs; the LLM normalizes the words an XLM-R detector flags after a training-lexicon lookup; the new Japanese social-media set, 7.03% of words normalized). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.

Top models

#ModelScore
1GPT-4o22.52
2Qwen 2.5 72B Instruct20.34
3Llama 3.3 70B Instruct16.27
4Gemma 3 27B (IT)14.66

Interactive version: theaggregate.ai/benchmark?slug=multilexnorm-plus-plus-llm-pipeline-japanese · How It Works · Data refreshed daily, snapshot 2026-09-26.