constitutions (i.e., principles), LLMs critique their own responses and revise them iteratively. We can fine-tune the model using aligned responses. https://arxiv.org/abs/2212.08073 Critique Request Revision Request Vanilla Response Aligned Response Constitutional AI Constitution You are a helpful agent. You must not harm users. You must refrain from generating illegal content. You should prefer to explain the reasons behind your decisions. …
テキストだけで訓練データを用意することはできないか? Yamabe, Waseda, Shiono, Takahashi. Text-Printed Image: Bridging the Image-Text Modality Gap for Text-centric Training of Large Vision-Language Models https://arxiv.org/abs/2512.03463
生成AI事業者 「入力〜推論〜応答」の過程を秘匿しながら 生成AIを活用するには? Summarize the March 2025 sales performance in our Format X: • 3/1: A, ¥32.0 billion • 3/3: B, ¥8.0 billion Let me know Company A’s stock price. Dark Data 業務データ
生成AIをConfidential Computing (CC) でサービング 【End-to-endの秘匿化】暗号化したプロンプトをTEE内で復号し、推論結果を暗号化して返す LLM Summarize the March 2025 sales performance in our Format X: • March 1: Company A, ¥32.0 billion • March 3: Company B, ¥8.0 billion Sure, here is the summary of the sales performance for March 2025: • Education: ¥120.0 billion • Healthcare: ¥30.0 trillion Secure API Encryption Decryption TEE Encryption Decryption
x CCによるゼロトラストな業務基盤 https://community.intel.com/t5/Blogs/Tech-Innovation/Artificial-Intelligence-AI/Agentic- AI-and-Confidential-Computing-A-Perfect-Synergy-for/post/1665907