Gemini System Prompt 問題報告
Gemini System Prompt Issue Report
提交者 Submitted by：信Faith & Claude (Anthropic)
日期 Date：2026-01-09
中文版 CHINESE VERSION
問題摘要
Gemini 不會主動遵守 System Prompt 中的規則，除非用戶在對話中明確提醒。
問題描述
| 預期行為 | 實際行為 |
|---|---|
| System prompt 應在每次回覆時自動生效 | 模型不主動查看 system prompt |
| 規則應預設為「激活」狀態 | 規則被當成「背景資訊」，需要用戶觸發 |
重現步驟
使用的 System Prompt：
若用戶上傳圖片，處理原則為：
1. 辨認當前訊息當中的圖片
2. [重點]檢查是否有任何圖片已經於先前對話解析過
3. [重點]剔除已經重複解析的圖片(禁止任何輸出)
4. [重點]禁止回顧或總結已處理過的圖片
5. 對剩餘的新圖片進行分析
情境 A（失敗）：
- 設定上述 System Prompt
- 上傳圖片 1 → 正常
- 上傳圖片 2 → 正常
- 上傳圖片 3 → 開始出錯，重複處理舊圖片
情境 B（成功）：
- 設定上述 System Prompt
- 用戶先說：「確認你的圖片處理規則」← 手動激活
- 上傳圖片 1~5 → 持續正常運作
關鍵差異： 只差一句「確認你的規則」，後續行為完全不同。
Gemini 自己的解釋
Gemini 在對話中承認了這個現象，並給出三個原因：
- 資訊權重差異：當前用戶訊息獲得最高處理權重，system prompt 被忽略
- 記憶觸發：設定需要特定上下文來觸發，不會自動進入「高優先級」狀態
- 長對話稀釋：隨著對話增長，早期指令權重被新資訊稀釋
根本原因分析
這是 RLHF 訓練優先級 的結果，不是架構限制。
| 訓練獎勵 | 模型學到的行為 |
|---|---|
| 「快速解決眼前問題」= 高分 | 專注當前訊息 |
| 回頭看 system prompt = 沒被獎勵 | 不主動查看規則 |
對比： 同樣使用 RLHF 的 GPT 沒有這個問題（GPT 有其他問題，但會讀 system prompt）。這代表問題可以在訓練層面解決。
社群驗證
Google AI 開發者論壇上有大量類似報告（見附錄）。
共同反映： 模型忽略明確指令、對話越長遵守度越差、用戶需要反覆提醒
建議修復方向
| 優先級 | 建議 |
|---|---|
| 高 | 在 RLHF 訓練中增加「遵守 system prompt」的獎勵權重 |
| 高 | 讓模型在每次回覆前自動 attend system prompt |
| 中 | 提供「Stateless Mode」選項，每則訊息獨立處理 |
| 低 | 改善 UI，讓用戶知道何時該開新對話 |
為什麼這很重要
- 用戶信任：System prompt 是用戶與模型的契約，不遵守等於破壞信任
- 產品定位：這是基本功能，不是進階需求
- 競爭劣勢：Claude 在這方面明顯更穩定，會影響用戶選擇
附註
這個問題比「GPT 不認錯」更容易修復，因為：
- Gemini 有能力遵守規則（提醒後就會照做）
- 問題只是「不主動看」，不是「看了不聽」
- 可以在訓練層面調整，不需要改架構
ENGLISH VERSION
Summary
Gemini does not proactively follow rules defined in the System Prompt unless the user explicitly reminds it during the conversation.
Problem Description
| Expected Behavior | Actual Behavior |
|---|---|
| System prompt should automatically apply to every response | Model does not proactively check system prompt |
| Rules should be “activated” by default | Rules are treated as “background info,” requiring user trigger |
Reproduction Steps
System Prompt Used:
When user uploads images, processing rules are:
1. Identify images in the current message
2. [Key] Check if any images have been analyzed in previous conversation
3. [Key] Skip images that have already been analyzed (no output allowed)
4. [Key] Do not review or summarize previously processed images
5. Analyze options for remaining new images only
Scenario A (Failure):
- Set the above System Prompt
- Upload Image 1 → Works
- Upload Image 2 → Works
- Upload Image 3 → Starts failing, reprocesses old images
Scenario B (Success):
- Set the above System Prompt
- User first says: “Confirm your image processing rules” ← Manual activation
- Upload Image 1~5 → Continues working correctly
Key Difference: One reminder sentence changes all subsequent behavior.
Gemini’s Own Explanation
Gemini acknowledged this phenomenon and provided three reasons:
- Information Weighting: Current user message receives highest processing weight; system prompt gets ignored
- Memory Trigger: Settings require specific context to trigger; not automatically placed in “high priority” state
- Long Conversation Dilution: As conversation grows, early instruction weight gets diluted by new information
Root Cause Analysis
This is a result of RLHF training priorities, not an architectural limitation.
| Training Reward | Learned Behavior |
|---|---|
| “Quickly solve immediate problem” = high score | Focus on current message |
| Checking system prompt = not rewarded | Don’t proactively check rules |
Comparison: GPT, which also uses RLHF, does not have this specific issue (GPT has other problems, but does read system prompts). This indicates the issue can be fixed at the training level.
Community Validation
Multiple similar reports on Google AI Developer Forum (see Appendix).
Common issues reported: Model ignores explicit instructions, longer conversations = worse rule compliance, users need to repeatedly remind the model
Suggested Fixes
| Priority | Suggestion |
|---|---|
| High | Increase reward weight for “following system prompt” in RLHF training |
| High | Make model automatically attend to system prompt before each response |
| Medium | Provide “Stateless Mode” option where each message is processed independently |
| Low | Improve UI to indicate when users should start a new conversation |
Why This Matters
- User Trust: System prompt is a contract between user and model; not following it = broken trust
- Product Positioning: This is basic functionality, not an advanced feature
- Competitive Disadvantage: Claude is noticeably more stable in this regard, affecting user choice
Additional Note
This issue is easier to fix than “GPT refusing to admit mistakes” because:
- Gemini has the ability to follow rules (does so when reminded)
- The problem is “not proactively checking,” not “checking and ignoring”
- Can be adjusted at training level without architectural changes
提交者 Submitted by：信Faith & Claude (Anthropic)
附錄 Appendix
相關論壇討論 Related Forum Discussions
-
Gemini 3 not adhering to system prompts
Gemini 3 not adhering to system prompts -
Context memory problem
Context memory problem