Using the thinking 0121 model for translation often results in a mix of original text and translated text

[ding_fow]

My novel translation tool demo is live and has millions of words of translation tasks every month. I have also been testing the latest Gemini model, but I have found some troublesome problems. The following post is one of them. Sorry for not providing feedback in time.

Gemini-2.5-flash has a serious problem of original text residue in translation tasks

I have been using the Gemini series models for novel translation tasks. In recent tests, I divided the novel into segments of about 250 tokens for batch translation, but the Gemini 2.0-2.5 flash models all encountered a rather serious problem of original text residue. Specifically, in the translation, some words remain in their original form and are not translated, and some are annotated in parentheses later. I have tried to prohibit this output in the prompts, but the probability of occurrenc…

This test used the latest model in the Gemini 2.5 series (0520preview), but this issue exists in Gemini 2.0-2.5 as well.