Liquid AI open-sources 'antidoom' FTPO training to fix small-model doom loops
helloiamleonie · x · 2026-09-15
Leonie of Liquid AI explains that small models with thinking capabilities get stuck in "doom loops" on complex tasks. Liquid AI's fix: "antidoom" training via Final Token Preference Optimization (FTPO).
They open-sourced a full tutorial notebook showing what doom loops are, how FTPO reduces them, and how to implement a custom DPOTrainer in TRL — available in the Liquid4All/cookbook repo (2.5k stars).
Related event: Liquid AI Open-Sources FTPO to Fix Small Models' Doom Loops(2 posts)→
More from coding & agent
- Atria Dawn Preview open-sourced: 744B MoE agent model tops BrowseComp at 92.5 — alexcovo_eth · 2026-09-16
- LangChain makes every Managed Deep Agent an MCP server callable by coding assistants — LangChain · 2026-09-16
- LlamaIndex on Building SDKs with Stainless, Whose Team Joined Anthropic — llama_index · 2026-09-16
- HappyRobot CEO on what it takes for an SE to become an FDE at his $1.2B startup — ivory_tang · 2026-09-16
- Andrew Ng: Prompting dies in 6 months, replaced by loops and graphs — Roger_M_Taylor · 2026-09-15
- Same Test: 0 of 21 Runs Produced a T-Shirt Design I Would Wear — hazelcough · 2026-09-15