We're experiencing three recurring Gemini model-level issues on ~10-20% of calls that no prompt engineering has been able to resolve. These appear to be degenerate generation states where the model stops following instructions. Bug 1: Question Repeat/Dump Loop (Most Critical) Gemini enters a degenerate state where it either: - Repeats the same question 5-15x in a single turn, or - Dumps all remaining questions in one turn instead of asking one at a time This happens mid-call after 3-7 minutes of normal conversation. The prompt explicitly says "ONE QUESTION AT A TIME" and "NEVER repeat a question" but Gemini ignores both when in this state. This is our most frequent failure mode, occurring on roughly 10-20% of calls. Call IDs: - call_fa51a3f7db2a6f44583d7610a8d - Question repeat/dump - call_a33e5ed9f5d9a4b39a1d9cce7e2 - Question repeat/dump - call_5d49345e9ac4e1f247e876a8b37 - Question repeat late in call Bug 2: Language Switching to Chinese Gemini randomly outputs Chinese characters (基础知识 = "basic knowledge") mid-conversation, sometimes hundreds of times in a row. This happens despite the prompt specifying "Language: English (US). Locale: en-US." at the top. Bug 3: Late-Call Latency Degradation E2E latency increases significantly as calls progress: - Turns 1-10: ~2,000ms - Turns 20-30: ~5,000-8,000ms - Turns 30+: ~8,000-9,000ms Call ID: call_5d49345e9ac4e1f247e876a8b37 Questions: 1. Can you add an output filter to detect and truncate repeated sentences before they reach TTS? This alone would eliminate our most common failure mode. 2. Can you add a language filter to reject non-English output at the TTS level? 3. Do you expose any settings for conversation history management (sliding window, summarization) to reduce late-call latency? 4. Are these known Gemini issues you're tracking?

Hi @ipadkid420

For our use case GPT 4.1 is significantly worse in terms of performance - Would it be possible to add output loop detection at the platform level? If the same sentence appears 2+ times in one turn, truncate it. This would fix the bug regardless of which model we use.

Ok I send your call ids to the team to look

Thank you so much! 🙏

Hi @ipadkid420 Team is looking into this, gemini has some problems with getting stuck in loops so they will try to figure something out.

Thanks Shah, really appreciate the support, let me know if I can help the investigation

Hi @ipadkid420 I have notify the team about this.

Thanks for the update, lmk if you need more call ids

Hello @ipadkid420 , do you still have this issue ? and I would advice to use GPT 4.1 as it more consistent.

The last time we tried we did have this issue, we’ve noticed GPT4.1 performs significantly worse than Gemini 3.0 for our use case.