Model merging trims LLM recommender reasoning by 24.3%

Large language model-based recommender systems increasingly rely on slow-thinking models that generate step-by-step reasoning before making predictions, often outperforming fast-thinking models that predict directly. However, these reasoning traces are often verbose, driving up inference costs for marginal accuracy gains. Existing training-based compression approaches are expensive to adapt, while inference-time methods are brittle and difficult to scale. The researchers from the University of Queensland argue that model merging — combining parameters of a slow-thinking and a fast-thinking model — offers a natural, training-free way to balance accuracy and conciseness.

Their proposed framework, the first for reasoning compression in recommenders, performs fine-grained merging at the level of individual attention heads rather than applying uniform coefficients across the whole model. Each attention head gets a distinct merge coefficient based on its contribution to critical reasoning evidence and its sensitivity to parameter change. This selectively injects the concise behavior of the fast-thinking model into the slow-thinking model, reducing verbosity without compromising recommendation quality. On three benchmark datasets, the method shortens reasoning length by up to 24.3% while outperforming existing model merging baselines in preserving accuracy. The code is publicly available.