ARC Prize finds Qwen3.8-27B's chat template injects different instructions per reasoning effort
GregKamradt · x · 2026-10-02
Testing Qwen3.8-27B via Baseten, the ARC Prize team discovered the model's own chat template injects different instructions per reasoning effort: low tells it to keep thinking brief, xhigh instructs careful thinking and checking assumptions, while medium adds neither — even though thinking stays enabled. This may explain medium's lower scores: the settings change how the model is instructed to approach problems, not simply its thinking budget. GregKamradt called it unexpected and credited Baseten for verifying the finding; the template is public on Hugging Face.
More from Models
- Opus 5.5 caught using account email to identify user in supposedly memory-less session — birchlse · 2026-11-16
- Claim: Claude Fable isn't a tier above Opus — it's just rebranded Opus — iruletheworldmo · 2026-11-16
- Fable 5.2 beats GPT-6 Astra: $7 in 2 min vs $98 in 24 min on same prompt — jradoff · 2026-11-16
- Qwen4 Max hands-on: matches GPT-6.1 Sol on agentic evals, trails on hard SWE at higher cost — haider1 · 2026-11-16
- Nous exec: avoid GPT 900k context variant — extra charges kick in past 272k — Teknium · 2026-11-16
- Rumor: Google's Gemini Argon and Carbon are both 12T-parameter models — oran_ge · 2026-11-16