Layer-Order Inversion: Rethinking Latent Multi-Hop Reasoning in Large Language Models
This work addresses a fundamental gap in mechanistic interpretability for LLMs, specifically for researchers and practitioners in AI and NLP, by challenging existing assumptions and providing a new framework, though it is incremental in building on prior hypotheses.
The paper tackles the problem of understanding how large language models internally compose multiple facts for multi-hop reasoning, showing that the hop-aligned circuit hypothesis does not generalize, with later-hop answer entities becoming decodable earlier than bridge entities, a phenomenon called layer-order inversion that strengthens with total hops. It proposes a probabilistic recall-and-extract framework, validated through empirical analyses, to explain this behavior and diagnose multi-hop failures.
Large language models (LLMs) perform well on multi-hop reasoning, yet how they internally compose multiple facts remains unclear. Recent work proposes \emph{hop-aligned circuit hypothesis}, suggesting that bridge entities are computed sequentially across layers before later-hop answers. Through systematic analyses on real-world multi-hop queries, we show that this hop-aligned assumption does not generalize: later-hop answer entities can become decodable earlier than bridge entities, a phenomenon we call \emph{layer-order inversion}, which strengthens with total hops. To explain this behavior, we propose a \emph{probabilistic recall-and-extract} framework that models multi-hop reasoning as broad probabilistic recall in shallow MLP layers followed by selective extraction in deeper attention layers. This framework is empirically validated through systematic probing analyses, reinterpreting prior layer-wise decoding evidence, explaining chain-of-thought gains, and providing a mechanistic diagnosis of multi-hop failures despite correct single-hop knowledge. Code is available at https://github.com/laquabe/Layer-Order-Inversion.