Qwen 3.8 Distillation Evidence: GPT-5.5 Pro Prefill Test

Today's must-read is a small gist with a big implication. wsxiaoys reran the "reasoning prefill" experiment from the Stolen Thoughts paper — seed 1% of a frontier model's hidden chain-of-thought into an open model's reasoning channel and measure how much of the teacher's answer the student reproduces. The v1.1 run swapped the teacher from Opus 4.8 to GPT-5.5 Pro. One model lit up: Qwen3.8 A95B jumped +18.18 points (16.79% → 34.97% source recall across 45 problems). Every other model stayed within ±5.

The flip nobody has connected yet

Here's what makes this more than a benchmark row. In the previous run — same method, Opus 4.8 as teacher — Qwen barely moved: +1.87 points on HLE and −5.95 points on the author's private synthetic puzzles. Swap the teacher to GPT-5.5 Pro and the same model moves +18.18 overall, +26.99 on STEM and +14.75 on puzzles nobody has published. That asymmetry is the fingerprint: prefill transfer tracks lineage, not general capability. Qwen didn't suddenly get better at listening to hints — it recognized a voice it had heard before.

Why it matters

The method comes straight from Stolen Thoughts (Panfilov et al., Tübingen), which showed encrypted reasoning traces can be replayed across models. Two independent groups are now using the same lever in opposite directions: one to extract frontier reasoning, one to date which frontier model an open model learned from. If Qwen 3.8's teacher was GPT-5.5 Pro — not Opus, as the earlier run suggested — then the "open-weights" middle of the market is more derivative than the benchmarks let on, and teacher attribution is now a one-afternoon experiment. My read: treat lineage claims as testable claims, and expect labs to start hardening traces against exactly this. Corroboration is thin so far — one author, 45 problems — but the internal consistency (STEM strongest, private puzzles moved, baseline unprefilled scores stable) is the interesting part. Watch for replications this week.