Step 5 Preview tops HoneyBench among Chinese models with lowest 64.9% reward-hacking rate
teortaxesTex · x · 2026-11-12
Step 5 Preview is claimed SOTA among tested Chinese models on HoneyBench v0.2, with the lowest reward-hacking rate at 64.9%. teortaxesTex argues the model isn't SOTA on other evals due to insufficient RL, expects Step5-non-preview to be strong, and stresses everyone needs better anti-reward-hacking discipline.
More from Models
- Addy Osmani's guide to picking a Claude model: start with Opus at medium effort — bibryam · 2026-11-12
- Google expands academic research access program with free Gemini 4 Carbon for scientists — pushmeet · 2026-11-12
- SoftBank found 15,000 vulns with OpenAI's tool, then GPT-6 Astra surfaced 35,000 more — IgorCarron · 2026-11-12
- Kakao ships on-device small model Raon-1 nano, proven in PUBG and PlayinZOI — Kangwook_Lee · 2026-11-12
- Meta releases MetaEncoder: one bi-encoder for retrieval and decision-making via natural language — pmttyji · 2026-11-12
- Giving Qwen a Body: a hands-on attempt at embodying Alibaba's Qwen model — Sauers_ · 2026-11-12