Nemotron 3 Superは自分の推論をユーザーのメッセージとして読み取ってしまうのか？

そう... Nemotron 3 Superを試してみたいと思いました（llama-serverをバックエンドとして、Aiderをクライアントとして使用します）。そして、それは自分自身の推論をユーザーメッセージとして解釈し、それをループしてしまうようです... これは興味深い問題で、他のモデルでは一度も経験したことがありません。多分、私が何か間違っているのでしょうか？何かアイデアはありますか？

重要な点として、llama-serverを以下のフラグで実行しています。ハードウェアにのみ関連するフラグはスキップします： --special（このフラグなしでも同じ動作をしました）、 --jinja

参考として、彼の推論の冒頭部分（および8kトークン分の思考の末の最終回答 xd）:

Okay, the user has shared a query that seems to be a mix of instructions and possibly a test or example related to text processing. The main content appears to be a series of instructions about how to handle a query, followed by some repetitive and possibly corrupted text. Let me break this down. The user starts with: "Okay, the user has shared a query that seems to be a mix of instructions and possibly a test or example related to text processing." But then they provide a block of text that looks like it's meant to be processed. Looking at the provided text, it begins with: "Okay, the user has shared a query that seems to be a mix of instructions and possibly a test or example related to text processing. The main content appears to be a series of instructions about how to handle a query, followed by some repetitive and possibly corrupted text." Then it goes into a block that says: "Okay, the user has shared a query that seems to be a mix of instructions and possibly a test or example related to text processing. The main content appears to be a series of instructions about how to handle a query, followed by some repetitive and possibly corrupted text." This seems to be repeating. Then it has: "Let me break this down. The user starts with: "Okay, the user has shared a query that seems to be a mix of instructions and possibly a test or example related to text processing." But then they provide a block of text that looks like it's meant to be processed." This is getting meta and confusing. It appears the user might have pasted some instructions multiple times or there's a loop in the text. The key part I need to focus on is the actual instruction at the very beginning: "Okay, the user has shared a query that seems to be a mix of instructions and possibly a test or example related to text processing. The main content appears to be a series of instructions about how to handle a query, followed by some repetitive and possibly corrupted text." But then the user says: "Let me break this down." and proceeds to describe what they think the user shared. However, the core instruction I should follow is given at the very start of the user's message: "Okay, the user has shared a query that seems to be a mix of instructions and possibly a test or example related to text processing. The main content appears to be a series of instructions about how to handle a query, followed by some repetitive and possibly corrupted text." ... ... <a couple thousands of tokens later> I apologize, but I'm unable to identify a clear query or specific text to process in your input due to repetitive, meta-commentary content and potential formatting issues. Your message appears to contain multiple instances of self-referential instructions about handling queries (e.g., "Okay, the user has shared a query that seems to be a mix of instructions and possibly a test or example related to text processing...") without a distinct, actionable request or data to work on. To assist you effectively, please: • Resend your actual question, text for processing, or task in a clear, concise format. • Avoid including meta-instructions or repetitive commentary in your resubmission. • Example of a clear query: "Summarize the key points of this article: [paste text]" or "What is the capital of Japan?" Once you provide a straightforward input, I’ll be happy to help with text processing, analysis, or any other task you need. Thank you for your patience!