Liquid AI's Maxime Labonne on post-training: SFT, RL, MOPD and regression
ziv_ravid · x · 2026-11-18
The Information Bottleneck podcast hosts Maxime Labonne, who runs post-training at Liquid AI, on how post-pretraining model training has changed over two years.
Key points:
- You still can't skip SFT: DeepSeek tried with R1-Zero and it was worse;
- After SFT, labs either run RL on a large mix of environments, or train one teacher per domain (code, math, etc.) and distill them into a single student — the latter is MOPD;
- RL vs SFT/distillation: SFT and distillation get the model to land on the right answer in some of its tries; RL makes it land there on the first one, despite giving roughly one bit of signal per long rollout;
- Labonne also thinks models write worse than a year ago because labs only chase agentic benchmarks now.
Related event: Liquid AI Post-Training Lead Details Two Years of SFT/RL/MOPD Shifts(2 posts)→
More from Models
- MachinePulseAI releases karpo-decision-31B decision model on Hugging Face — rohanpaul_ai · 2026-11-18
- OpenAI ships GPT-6.2 Astra with stronger engineering accuracy and website visual reasoning — KarelDoostrlnck · 2026-11-18
- Dev take: closed-weight flagship models are fine, but small models should be open-weight — gnukeith · 2026-11-18
- FutureSim Q3 Results: Gemini 4 Carbon Wins, Builds Its Own World Cup Simulator — SamuelAlbanie · 2026-11-18
- Microsoft team member hails in-house model trained to understand work, now in core products — jteevan · 2026-11-18
- SID-2 Search Model Uses Compute-Scaling Swarms, Claims Better Results Than Opus at 8x Speed — KarelDoostrlnck · 2026-11-18