Doom loops waste time and compute when reasoning models get stuck repeating the same span until the context window fills. Small models are especially prone to this on long math or coding traces, leading to failed answers and lower evaluation scores even when the model knows the correct solution. The core issue is a single token that starts the loop—often a high‑probability word like “the”, “Wait”, or “Alternatively”—that the model keeps choosing when uncertain. Greedy sampling at low temperature then has no exit, and each repeat makes that token more likely, locking the model into a cycle.
Antidoom solves this by targeting only that first loop‑starting token. It detects loops in a low‑temperature generation mix, extracts the rejected token, and gathers a set of coherent alternatives. Using Final Token Preference Optimization (FTPO), it shifts probability from the offending token to those alternatives while leaving the rest of the model untouched. Training is lightweight: one epoch with LoRA ranks of 128‑256 on attention, MLP, and lm_head, learning rates around 4e‑6 to 2e‑5, and early stopping when chosen tokens win about 35% of the time. On an early LFM2.5‑2.6B checkpoint, looping dropped from 10.2% to 1.4%; on Qwen3.5‑4B it fell from 22.9% to 1%. Evaluation scores rose correspondingly, showing the gains come purely from reduced looping.
The whole pipeline—data generation on eight GPUs for about an hour, then training on one GPU for one to two hours—is fast and fully open source. Teams can plug Antidoom into their fine‑tuning workflow, run a second round if new loops appear, and recover stable, accurate reasoning without changing sampling temperature or adding heavy reinforcement‑learning overhead.
#AI #ML #LLM #Reasoning #Productivity #OpenSource