Reproduce “Towards Efficient Reasoning in LLM-Based Recommender Systems via Model Merging” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Linh Dieu Le, Tong Chen, Shazia Sadiq, Hongzhi Yin, Ming Jin, Junliang Yu

The University of Queensland researchers developed REAM, a model merging framework that reduces reasoning length in LLM-based recommender systems by selectively transferring fast-thinking behaviors to slow-thinking models. This approach leverages head-level signals like retrieval criticality and decision faithfulness to achieve up to a 24.3% reduction in reasoning tokens while improving recommendation accuracy, for example, a 4.7% lower MAE on Amazon Book.

Key insights

  • Reasoning-relevant behaviors within LLMs are localized to a sparse subset of attention heads, implying that fine-grained merging is more effective than uniform or layer-wise approaches.
  • "Retrieval criticality" and "decision faithfulness" are complementary and causally important head-level signals for identifying critical reasoning components, enabling targeted compression.
  • Excluding the final few FFN layers from merging is critical, as these layers are highly sensitive and directly influence the reasoning-to-rating mapping, impacting accuracy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.