# Trying to tame QWEN 3.8 27B pathological over-reasoning and misbehaviors — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/2b093f8def438f59

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-07 18:12 ingested 2026-09-08 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1w9zrpg/trying_to_tame_qwen_38_27b_pathological/)

## retained text



```
Today I finally hit a wall with QWEN 3.8 dense 27B, Q6\_X\_L with KV cache q8\_0 temp 0.85 with medium reasoning effort, producing over thinking, failing to follow precise instructions, re evaluation in loops data already seen, "Actually…/Wait…/Hmm…", meta-reasoning, theorizing about user intent and [you know what](https://store.piffa.net/lm/reasoning_pathology_report.md) else...

So I stopped and asked GPT Sol and QWEN to read the reasoning traces of my old Pi coding sessions, identify recurring misbehaviors and propose an appropriate APPEND\_SYSTEM.md to tame those in Pi  (place that in \~/.pi/agent/APPEND\_SYSTEM.md and it will follow the initial prompt).

So for those who wanna check out:

* [https://store.piffa.net/lm/APPEND\_SYSTEM.md](https://store.piffa.net/lm/APPEND_SYSTEM.md)
* [https://store.piffa.net/lm/reasoning\_pathology\_report.md](https://store.piffa.net/lm/reasoning_pathology_report.md) that's a repost of what was find in my particular session, mind you it's all slop

If you happen to have something similar and have rules to add or any other advice to tame such bad behaviors please share. 
```



document fetched 2026-09-08 00:40 · canonical https://reddit.com/r/LocalLLaMA/comments/1w9zrpg/trying_to_tame_qwen_38_27b_pathological
