Layer-Order Inversion: Rethinking Latent Multi-Hop Reasoning in Large Language Models

TL;DR

This paper challenges existing assumptions about how large language models perform multi-hop reasoning, revealing that answer entities can be decoded earlier than bridge entities, and proposes a probabilistic framework to explain this behavior.

Contribution

It introduces the layer-order inversion phenomenon and a probabilistic recall-and-extract model to better understand multi-hop reasoning in LLMs.

Findings

Later-hop answer entities can be decoded before bridge entities.

The proposed framework explains multi-hop reasoning behaviors and failures.

Systematic probing validates the probabilistic recall-and-extract model.

Abstract

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.