はじめに

この記事では、Azure、AWS、Google Cloudの3大クラウドサービスのAIサービスの新規機能リリース履歴をまとめています。

主に以下のURLの情報をもとに新機能のキャッチアップを行っています。


Azure


AWS


Google Cloud

2025年06月02日: Vertex AIでGeminiのThinkingモデルで思考プロセス「Thought signatures」の応答が可能に

GeminiシリーズのモデルでThinkingモデルを利用する際に、その生成過程や内部的判断の痕跡に相当するThought signaturesを付与できるようになりました。

Thought singnaturesには、例えばモデルがどのルールや知識源を参照したか、どのステップでどんな計算を行ったかなどのメタ情報が含まれます。

モデルの推論過程を可視化できるため、応答の妥当性やバイアスのチェックがしやすくなり、モデルがどういう判断を経てその答えに至ったかを細かく解析可能になります。

サポートモデル

  • Gemini 2.5 Pro preview
  • Gemini 2.5 Flash preview

APIで利用するとき

REST APIでGeminiモデルを呼び出す際は、APIのリクエストボディに、includeThoughtsTRUEに設定して、リクエストすることでThought singnatures が有効になります。

{
  "contents": [
    {
      "role": "user",
      "parts": [
        {
          "text": "こんにちは、元気ですか?"
        }
      ]
    }
  ],
  "system_instruction": {
    "parts": [
      {
        "text": "You are a helpful assistant."
      }
    ],
    "role": "model"
  },
  "generation_config": {
    "maxOutputTokens": 4096,
    "temperature": 0.1,
    "frequencyPenalty": 0.1,
    "presencePenalty": 0.1,
    "topP": 0.1,
    "thinkingConfig": {
      "thinkingBudget": 4,
      "includeThoughts": true
    }
  }
}

レスポンスボディのうち、thoughtTrueになっているメッセージがThought singnaturesになります。

{
  "candidates": [
    {
      "content": {
        "role": "model",
        "parts": [
          {
            "text": "Alright, let's see... I'm about to greet someone in Japanese. Now, the core phrase I need is \"Konnichiwa.\" That's the standard, polite greeting for \"Hello\" during the day. Simple enough. Just gotta make sure my pronunciation is clear. I should add a slight upward inflection at the end, to make it sound like a genuine greeting and not a declarative statement. Focus on the \"chi\" and the \"wa\", a bit more than the \"konni\". Got it. Now, I'm saying it, *internally* of course.  Let's roll! \"Konnichiwa.\"  Yeah, that's pretty solid.  Good start.\n",
            "thought": true
          },
          {
            "text": "こんにちは!はい、元気ですよ。ありがとうございます。"
          }
        ]
      },
      "finishReason": "STOP",
      "avgLogprobs": -0.37282960755484446
    }
  ],
  "usageMetadata": {
    "promptTokenCount": 11,
    "candidatesTokenCount": 14,
    "totalTokenCount": 41,
    "trafficType": "ON_DEMAND",
    "promptTokensDetails": [
      {
        "modality": "TEXT",
        "tokenCount": 11
      }
    ],
    "candidatesTokensDetails": [
      {
        "modality": "TEXT",
        "tokenCount": 14
      }
    ],
    "thoughtsTokenCount": 16
  },
  "modelVersion": "gemini-2.5-flash-preview-05-20",
  "createTime": "2025-06-05T13:52:51.369624Z",
  "responseId": "M6FBaNjHFo23k7QP7a2AqAk"
}