CoT similarity test suggests Qwen3.8 may have been trained on GPT 5.5 reasoning traces

Chromix_ · reddit · 2026-09-10

Community analysis (speculative, unofficial) presents evidence that Qwen3.8 may have been post-trained on GPT 5.5 chain-of-thought data:

Method

Building on a technique to extract hidden CoT from API-only models, researchers compared answer similarity across open models on a small, diverse benchmark
In a second run, Qwen3.8's reasoning was prefilled with a small amount of GPT 5.5 CoT; the model picked it up and its visible answers became more like GPT 5.5's — a signal not expected in models that never saw that CoT during training
A private benchmark guarded against false positives; the effect rules out training on HLE results alone

Results (n=45)

Qwen3.8 A95B: 16.79% → 34.97%, a +18.18 pp jump, far above others
Kimi K3: +4.54 pp — the author suspects it saw some Claude traces
DeepSeek V4 Flash: −1.17 pp; Inkling: +0.46 pp

Implication: if true, OpenAI could ship a small model with full CoT distillation that beats similar-sized Qwen models, though it would compete with its own Luna line.

Original post →

CoT similarity test suggests Qwen3.8 may have been trained on GPT 5.5 reasoning traces

More from Models