5.11 Model Response Quality Issues

Typical symptoms: Off-topic responses / repetitive outputs / incorrect tool selection / raw <think>...</think> tags appearing in output / Chinese responses mixed with English (or vice versa).

30-Second Decision Guide

90% of quality issues aren't caused by the model itself—first rule out "environmental interference":

1. Has the Quick lane taken over a task that should be handled by the Thinking lane? Check the model tag at the bottom of AI Dock.
2. Is context usage exceeding 90%? Start a new session.
3. Are too many injected skills interfering with each other? Type /skills in AI Dock to review the list.
4. If it's truly a model capability issue → switch to a model in the Thinking lane.

Troubleshooting Checklist

1. Lane Mismatch — Using lite/mini models in the Quick lane for tasks requiring planning will inevitably result in poor quality. You can see the actual model being used at the bottom of AI Dock. If incorrect, go to Settings Panel → AI Engine and correctly assign the appropriate lane.

2. Exposed <think> Raw Tags — Some domestic models wrap their reasoning steps inside <think> tags. Studio’s InlineThinkingRouter automatically strips these tags. If you see raw tags:

   • Upgrade to the latest version (fixes various vendor-specific variants)
   • Or temporarily set Reasoning visibility to hidden

3. Mixed-Language Output — In Settings Panel → AI Engine → Advanced, set "Response Language" to "Follow System" or "Chinese". A few smaller models may still occasionally output English; switching to a more capable model is the most effective solution.

4. Incorrect Tool Selection — Type /skills in AI Dock to see which skills are currently injected. Too many redundant skills can interfere with decision-making; uninstall unused ones via Skill Workshop → Installed.

5. Degraded Performance from Long Conversations — Extended conversations accumulate excessive context, degrading model performance. Studio automatically compacts context at 70% usage, but if usage exceeds 90% and you're still asking follow-ups, start a new session.

Permanent Solutions

• Always select the strongest available model for the Thinking lane—don’t use mini models just to save tokens.
• For critical scenarios, go to Skill Workshop → Create New Skill and write a dedicated SKILL.md file so the Agent doesn’t have to "guess" which path to take.
• Pay attention to RDK Studio update notifications—the model understanding layer (including reasoning tag processors like InlineThinkingRouter) continuously improves with each version.