Qwen3.6 hits infinite Yoshi reasoning loop

The "Yoshi incident" highlights a critical executive function gap in reasoning-native LLMs where internal logic becomes decoupled from output generation.

• –Native "thinking" toggles can create a "hallucination of progress," as the model endlessly self-corrects without delivering a final response.
• –Sparse MoE architecture (35B total / 3B active) allows for extreme speeds (200+ tok/s), making these recursive loops particularly aggressive in token consumption.
• –While the model hits a record 73.4% on SWE-bench Verified, its reliance on rigorous self-critique fails when there is no objective "correct" solution.
• –Community users suggest increasing repetition penalties or disabling thinking mode for non-logical tasks to avoid termination collapse.