Neuron Editing Can Fix LLM Repetition, Not Doom Loops

Large Language Models (LLMs), particularly instruction-tuned versions like Gemma 4, frequently exhibit a failure mode where they fall into repetitive output loops when generating long factual enumerations. These loops are highly reproducible and persist despite various prompting or inference adjustments. This paper investigates whether such pathological behavior can be localized and removed through targeted weight edits. The study successfully traced these repetition loops to a small set of MLP neurons or, in Mixture-of-Experts models, specific routed experts. By applying static weight edits, sometimes as minimal as inverting the sign of a single neuron in smaller models, these loops could be suppressed without degrading general benchmark scores. However, the research also differentiates between simple repetition and more complex 'doom looping,' a non-convergent state where the model self-corrects in circles due to an inability to recall a fact. While neuron edits reduced this residual failure, they did not eliminate it, suggesting that doom looping is fundamentally a knowledge-precision problem rather than a removable circuit. This work demonstrates the feasibility of fixing specific generation pathologies via weight edits but also delineates the limits of this approach.