Reproduce “Back Attention: Understanding and Enhancing Multi-Hop Reasoning in Large Language Models” 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
Zeping Yu, Yonatan Belinkov, Sophia Ananiadou
Researchers at the University of Manchester and Technion – Israel Institute of Technology developed "Logit Flow" to analyze information flow in LLMs, revealing a four-stage single-hop reasoning process and identifying conflicting logit interference as a failure point in multi-hop reasoning. Informed by these insights, they propose "Back Attention," a lightweight architectural modification that improved LLM accuracy on various reasoning datasets and enabled smaller models to achieve performance comparable to larger counterparts.
Key insights
- Single-hop knowledge prediction in LLMs follows a four-stage process where initial layers extract general features, and deeper layers at relation/last positions refine and differentiate the correct answer.
- Multi-hop reasoning failures stem from conflicting logits, where the first-hop's "relation attribute extraction" disproportionately reinforces an intermediate answer, interfering with the integration of subsequent hop information.
- Giving lower layers access to the more abstract representations of higher layers through "Back Attention" can improve reasoning performance and increase model efficiency by facilitating better information integration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.