Liquid AI Open Sources Antidoom Method to Cut Doom Loops in Reasoning Models

Liquid AI released Antidoom, an open-source method that uses Final Token Preference Optimization to reduce repetitive loops in reasoning models. The approach identifies the first token where a loop begins and trains the model to pick a coherent alternative instead.

The package includes code and the LiquidAI/antidoom-mix-v1.0 dataset. Tests on LFM2.5-2.6B showed loop rates on hard math and coding tasks falling from 10.2 percent to 1.4 percent. On Qwen3.5-4B the same tasks dropped from 22.9 percent to 1 percent. The entire training run finishes in a few hours on standard hardware.

Reasoning models often enter doom loops when they repeat the same token sequence without progress. Earlier fixes either changed the full output distribution or required heavy reinforcement learning from human feedback. Antidoom avoids both problems by focusing only on the token that starts the cycle.

Liquid AI states the method keeps the rest of the model behavior intact. That claim separates it from broader alignment techniques that can shift model style or capability. The company published the full pipeline so other labs can test the same token-level intervention on their own models.

Developers working with long chain-of-thought outputs now have a lightweight option to reduce wasted tokens. The change matters most in math, code, and tool-use workflows where a single loop can double inference cost. Early adopters report the fix integrates into existing fine-tuning scripts without new infrastructure.

Antidoom also supplies a mixed dataset of loop examples and clean continuations. Training uses a simple preference loss only on the flagged first token, leaving later tokens under the original next-token objective. The result preserves benchmark scores while cutting unproductive repetition.

The release comes at a time when multiple labs are scaling test-time compute for reasoning. Longer internal traces raise the chance of self-reinforcing loops. A targeted patch that works in hours rather than days of reinforcement learning gives smaller teams a practical defense.

Several open questions remain. The published numbers cover only two model sizes, and it is unclear how the same token-level fix performs on models above 30 billion parameters. Liquid AI has not released ablations that isolate whether the dataset or the loss formulation drives most of the gain.

Independent labs will need to verify whether the method transfers to agentic settings where loops can involve tool calls rather than pure text. If the first-token signal stays reliable across those domains, Antidoom could become a standard post-training step for any reasoning pipeline.

The code and data are available now on Hugging Face under the LiquidAI organization. Teams can reproduce the reported numbers or adapt the pipeline to their own loop detection heuristics. Further updates from Liquid AI or early users will show whether the approach scales beyond the initial benchmarks.