LiquidAI releases Antidoom, Open-Source Fix for Doom Loops in Reasoning Models

Liquid AI released Antidoom, a new open-source method that eliminates "doom loops" — a common failure mode in small reasoning models where the model repeats a token span (e.g., "Wait, let me check") endlessly until the context window is exhausted, especially on hard math and coding tasks.

#Key Results

• Early LFM2.5-2.6B checkpoint: 10.2% → 1.4% doom-loop rate
• Qwen3.5-4B: 22.9% → 1% (greedy sampling)
• Evaluation scores improved across the board

#How it works

• Identifies the single overtrained trigger token (often interruptives like "Wait," "So")
• Uses Final Token Preference Optimization (FTPO) — a targeted DPO-family technique that retrains only the trailing token mid-generation
• Spreads probability to coherent alternatives with minimal disruption to the rest of the model

#Additional Insights

• After fixing loops, near-greedy sampling often performs best (high temperature may have masked the issue)
• Very efficient: ~2 hrs data generation on 8xH100 + 1-2 hrs training on a single H100 for 2-4B models
• Fully open source with blog post and code

sources: @liquidai (Liquid AI)