The Special Token `<Think>` Problem/Bug of Latest DeepSeek LLM
The latest version of the DeepSeek LLM has been found to produce unstable responses when given specific incomplete or special tokens such as <think> or <think. The issue was discovered by users who reported abnormal behavior when interacting with the model, including unrelated responses and hidden reasoning traces. The cause of the problem is still being investigated, with possible explanations including tokenizer or prompt parser bugs, special token parsing issues, and GPU cache or load balancer cache leakage.
- ▪The DeepSeek LLM model may produce highly unstable responses when given specific incomplete or special tokens.
- ▪The issue can be reproduced by inputting the text <think> into the model with the Deep Think Function enabled.
- ▪The model's behavior can vary, with some tests producing normal explanations and others generating unrelated responses or attempting to answer unseen questions.
Opening excerpt (first ~120 words) tap to expand
Recently the users observed an issue/bug in the latest version of the DeepSeek LLM: during testing, it was found that when specific incomplete or special tokens such as <think> or <think are included in the user prompt, the model may produce highly unstable responses, severe hallucinations, abnormal reasoning outputs, or unexpected behavior.(adsbygoogle = window.adsbygoogle || []).push({}); The purpose of this article is to document the issue, demonstrate how the behavior can be reproduced, and discuss the verification experiments conducted to better understand the possible cause of the problem.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at PixelsTech.
Discussion
0 commentsMore from PixelsTech
