Kaggle LLM attribution contest: GLM text is flagged as Claude 44% of the time
HarveenChadha · x · 2026-10-18
While exploring a new Kaggle competition where you must predict which LLM produced a given text, the author found telling failure modes: GLM's text is predicted as Claude 44% of the time, dragging its recall down to 0.26–0.30. GLM shares Claude's punctuation fingerprint — about 1.4–1.5 tight em-dashes per 100 words, straight quotes and no curly quotes (Gemini uses no em-dashes, GPT uses curly quotes). GLM-5.3-Flash's answers are more similar to Claude Fable-5's answers to the same prompt than to GLM-5.3's own, suggesting GLM was heavily trained on Claude output. The author speculates this may relate to distillation being banned, possibly explaining why new Chinese models are taking longer to ship (unverified).
More from Models
- Elie Bakouch: classic SFT distillation should matter less without logits access — eliebakouch · 2026-10-18
- Neither LLMs nor most humans can say "I don't know", observer notes — StewartalsopIII · 2026-10-18
- Dean Ball: distillation likely explains ~50% of Chinese labs' close-behind pace — deanwball · 2026-10-18
- Gemini's Video Limits: One 10s Clip Eats 32% of Your Hourly Quota — nikitasius · 2026-10-18
- OpenAI staffer: distillation plays bigger role in China's benchmark catch-up than realized — deanwball · 2026-10-18
- Users find GPT-6 lacks GPT-4's magic: forced jokes, hedging, answer dodging — Emergency-Ad-8185 · 2026-10-18