On PostTrainBench, AI post-training now beats the human-tuned SmolLM3-3B from the same base
maksym_andr · x · 2026-11-12
fullrank flagged a milestone: on PostTrainBench, starting from the same base model, AI-driven post-training now beats the officially human post-trained SmolLM3-3B on average.
Important caveat from the author: this doesn't mean AI is superhuman at post-training — the AI hill-climbs a single eval, while the human team built a general-purpose model, so the objectives differ.
A notable data point for automated model post-training.
More from Models
- Uncensored DeepSeek-V4-Pro-Qwen3.5-9B GGUF merge trends on Hugging Face — mradermacher · 2026-11-12
- Google launches Gemini Decisions API: 1,000 multimodal judgments in one pass — dejanseo · 2026-11-12
- User claims Gemini 4 Carbon clearly beats Fable 5.2 in practice — tokumin · 2026-11-12
- Frontier models race Tom Cruise's canyon in Top Gun — Claude Opus refused to fly — shekitup · 2026-11-12
- Liquid AI's 1.7B open model beats Baidu's ERNIE 5.0 on multi-turn instruction following — JosephJacks_ · 2026-11-12
- Reddit proposal: replace benchmark numbers with per-task red/yellow/green ratings — Toothpasteweiner · 2026-11-12