Liquid AI's "Antidoom Training" cuts model doom loops from 10% to 1%
helloiamleonie · x · 2026-09-25
A quick refresher on RL with verifiable rewards (RLVR): extract the final answer from the model's response, verify it against ground truth, and grade the trajectory accordingly. This works well but the reward signal is sparse, making long multi-turn generations inefficient.
To address this, the team developed antidoom training: detect a doom loop, find the first token that triggers it, reject and downsample it, and upsample alternative tokens. On LFM2.5-2.6B this reduced doom loops from 10% to 1%.
More from Models
- Gemini 4 Carbon matches predecessor's score but explores less, gets distracted — patience_cave · 2026-11-11
- Fable 5.2 takes MazeBench frontier, matches GPT-6 Astra puzzle-solving — patience_cave · 2026-11-11
- MazeBench: Fable 5.2 scores 15%, near-frontier at half price; Gemini 4 Carbon 11% — patience_cave · 2026-11-11
- ERNIE 5.5 Flash Preview enters LMArena Battle Mode for crowdsourced ranking — arena · 2026-11-11
- Indie dev flips from $20 Pro back to $200 Gemini Ultra after two days of deep use — yihui_indie · 2026-11-11
- OpenAI retiring ChatGPT's Standard Voice with 9 original voices on Dec 11, petition launched — ChaosGremlinOG · 2026-11-11