Qwen Reasoning Traces Show Strange Refusals and Self-Commands
wombweed · reddit · 2026-09-02
A Reddit user reported observing strange refusal-related content and self-commands in the reasoning traces of Qwen3.8-Flash-Next during routine coding sessions.
- Anomaly: During standard Python/Go web development tasks with no actual safety or IP issues, the model's reasoning traces began centering around alignment/refusal, repeating assurances that the task is safe to proceed and ignoring non-existent reminders.
- Self-Directing: The traces later shifted to imperative self-instructions, such as "Please edit the file to make it more testable."
- Context: The issue occurred using llamacpp with bf16 kv and q80 weights. The user speculates the Unsloth gguf version might be triggering refusal activations that persist via the prefix cache.
- Outcome: Despite the corrupted thought process, the model's final output and tool calls remained functional and correct.
More from coding & agent
- Pi Durable Agent harness on Cloudflare powers fully autonomous apps with zero human input — MicahBerkley · 2026-10-18
- Dev: the right prompts can make your codebase 2x faster every 10 minutes for an hour straight — cephaloform · 2026-10-18
- LLM one-shots a Minecraft 3x3 piston door by building its own Linux test infra — Vjeux · 2026-10-18
- Dev builds Mumbai-inspired 3D manga background generator from OpenStreetMap data — CurieuxExplorer · 2026-10-18
- Solo founder builds a money-printing data service on Claude's Pi Durable harness — MicahBerkley · 2026-10-18
- Dev builds Arch-based OS with custom desktop and browser using Claude Code — Snoo-10493 · 2026-10-18