💬 NLP

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

This paper challenges the prevailing "hop-aligned" hypothesis of multi-hop reasoning in LLMs by demonstrating "layer-order inversion," where later-hop answers can be decoded before intermediate bridge entities, and proposes a "probabilistic recall-and-extract" framework to explain this phenomenon through broad shallow recall followed by selective deep extraction.

Original authors: Xukai Liu, Ye Liu, Jipeng Zhang, Yanghai Zhang, Kai Zhang, Qi Liu

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xukai Liu, Ye Liu, Jipeng Zhang, Yanghai Zhang, Kai Zhang, Qi Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Large language models are the digital minds behind the most advanced artificial intelligence systems today. They are trained on vast amounts of text, learning to predict the next word in a sentence with remarkable accuracy. This ability allows them to answer questions, write stories, and solve problems that require connecting different pieces of information. One of the most impressive feats these models perform is multi-hop reasoning. This is the process of answering a question that cannot be solved with a single fact, but instead requires chaining together several pieces of knowledge. For instance, to answer "What country is the sport that Stephen Curry plays originally from?", a model must first identify the sport, then find the origin of that sport, and finally combine those two facts to reach the answer. For years, scientists believed these models solved such puzzles by following a strict, step-by-step internal path, much like a human solving a math problem one equation at a time. They thought the model would first figure out the intermediate fact, store it, and then use it to find the final answer. Understanding exactly how these machines perform this mental gymnastics is crucial, not just for improving their intelligence, but for knowing when and why they might fail.

A team of researchers from the University of Science and Technology of China and the Hong Kong University of Science and Technology has now challenged this long-held belief. By peering inside the internal workings of large language models, they discovered that these systems do not always follow the orderly, step-by-step path that was previously assumed. Instead, they found a phenomenon they call layer-order inversion. In a standard, step-by-step process, you would expect the model to figure out the middle piece of information before it can figure out the final answer. However, the researchers observed that in many cases, the model seems to "know" the final answer before it has fully figured out the intermediate steps. This happens as the questions become more complex, involving more links in the chain of reasoning. The deeper the question goes, the more likely it is that the model will have the final answer ready in its internal memory before it has explicitly processed the bridge facts that connect the dots.

To uncover this, the researchers used a technique called Patchscopes, which acts like a microscope for the model's internal thoughts. They fed the models complex questions with up to four steps of reasoning and watched how the information emerged as the data passed through the model's layers. They looked at two specific points: the beginning of the question, where the subject is introduced, and the very end of the question, where the model prepares to generate an answer. They found that while the model often processed the first step of the reasoning near the beginning, the final answer often appeared in the model's internal state much earlier than expected, sometimes even before the intermediate facts were fully formed. This contradicts the idea that the model builds a rigid circuit where each step must be completed before the next one begins. Instead, it suggests a more fluid process where the model can recall the final answer directly, almost like a gut feeling, before it has fully constructed the logical path to get there.

The researchers propose a new way to understand this behavior, which they call a probabilistic recall-and-extract framework. Rather than a rigid machine following a fixed set of rules, they suggest the model works more like a vast library where information is recalled based on probability. In this view, the model does not just move vertically through its layers, solving one step at a time. It also moves horizontally across the words in the sentence. The researchers found that in the shallow layers of the model, near the end of the sentence, there is a broad, intuitive recall of the final answer. This recall acts like a signal or a hunch that points toward the correct solution. This signal is not the result of a completed logical chain, but rather a direct association that the model picks up on. In weaker models or for very difficult questions, this intuitive signal becomes even more dominant, sometimes leading the model to the right answer even if it hasn't fully reasoned through the steps, or conversely, leading it astray if the signal is misleading.

This new perspective helps explain why models sometimes succeed or fail in ways that were previously confusing. When a model gets a multi-hop question right, it might be because this intuitive signal was strong enough to guide the final answer, even if the step-by-step reasoning was incomplete. Conversely, when a model fails, it is often not because it lacks the facts, but because it failed to retrieve the final answer through this intuitive signal or failed to extract the correct information from the accumulated knowledge. The study shows that the model's ability to answer complex questions relies on a balance between this broad, intuitive recall and a more focused, deep extraction of the specific answer. This discovery suggests that the internal logic of artificial intelligence is less like a rigid assembly line and more like a dynamic process of memory and association, where the destination can sometimes be seen before the path is fully walked.

Technical Summary: Layer-Order Inversion in Large Language Models

Problem Definition

Despite the demonstrated success of Large Language Models (LLMs) in multi-hop reasoning, the internal mechanisms by which they compose multiple factual knowledge steps remain poorly understood. Recent literature has proposed the hop-aligned circuit hypothesis, suggesting that LLMs resolve multi-hop queries sequentially across layers: bridge entities (intermediate facts) are computed in earlier layers and propagated forward to yield final answers in deeper layers.

However, this hypothesis relies heavily on analyses of two-hop queries and synthetic datasets. It remains unclear whether this strict, sequential layer-wise ordering generalizes to higher-hop queries (3-hop, 4-hop) in real-world settings. Furthermore, existing circuit-based explanations struggle to account for phenomena such as "shortcuts" (answering correctly despite failing intermediate single-hop checks) and the specific failure modes of multi-hop reasoning when single-hop knowledge is intact.

Methodology

The authors investigate latent multi-hop reasoning using the MQuAKE benchmark, which contains multi-hop queries with up to four hops and annotated intermediate facts. The study focuses on two decoder-only models: GPT-J-6B and Llama 3-8B.

The methodology employs three primary analytical tools:

Patchscopes: An unsupervised probing technique used to extract hidden states from specific tokens (the subject token and the last token) at various layers. These states are patched into a target prompt to determine when specific entities (bridge entities vs. final answer entities) become decodable. The authors apply global and local filtering to mitigate noise from shallow layers.
Causal Intervention: A method to assess the causal contribution of specific activations. The authors compare a "Clean Run" against a "Patched Run" where Gaussian noise is added to specific token-layer activations (e.g., at the subject or last token) to measure the resulting drop in the probability of the correct answer.
Hidden-State Similarity Analysis: Used to track the evolution of internal representations across layers, comparing multi-hop queries to their corresponding single-hop counterparts.

Key Findings

1. Discovery of Layer-Order Inversion

Contrary to the hop-aligned circuit hypothesis, the authors observe a phenomenon they term layer-order inversion. In higher-hop queries (3-hop and 4-hop), the final answer entity often becomes decodable in the hidden states of the last token at layers earlier than the corresponding bridge entities become decodable at the subject token.

Evidence: In 4-hop queries on Llama 3, the final entity ( $e_4$ ) emerges earlier than the first bridge entity ( $e_1$ ) across all settings.
Implication: This ordering contradicts the assumption that bridge entities must be explicitly computed and decoded before later-hop entities can be accessed.

2. The Probabilistic Recall-and-Extract Framework

To explain layer-order inversion, the authors propose a probabilistic recall-and-extract framework. This framework posits that multi-hop reasoning is not a deterministic, step-by-step computation but a superposition of probabilistic paths involving:

Vertical Recall: Knowledge is recalled across layers (conditioned on current representations), consistent with prior work on MLPs.
Horizontal Recall: Knowledge is recalled across tokens. Accumulated context allows deeper-hop knowledge to become available at the last token before all bridge entities are explicitly decoded.
Intuition-like Signal: The shallow horizontal recall of deeper-hop knowledge at the last token acts as an "intuition-like" signal for answer selection. Causal intervention shows that perturbing these shallow activations at the last token significantly impacts prediction accuracy, particularly in weaker models (GPT-J) and higher-hop queries.
Selective Extraction: Deeper attention layers perform a selective extraction of the accumulated, answer-relevant knowledge to generate the final output.

3. Reinterpretation of Failure Modes

The framework provides mechanistic diagnoses for multi-hop failures:

Insufficient Shallow Recall: In "Incorrect" cases (where single-hop facts are correct but the multi-hop answer is wrong), the model fails to sufficiently recall deeper-hop knowledge in the shallow layers of the last token.
Insufficient Extraction: Even if recall occurs, deeper attention layers may fail to sufficiently amplify the answer-relevant knowledge to dominate the decoding distribution.

Key Contributions

Identification of Layer-Order Inversion: The paper is the first to systematically demonstrate that later-hop entities can become decodable earlier than bridge entities in higher-hop queries, challenging the universality of hop-aligned circuits.
Probabilistic Framework Proposal: The authors introduce a framework modeling multi-hop reasoning as probabilistic recall (both vertical and horizontal) followed by selective extraction. This explains why "intuition-like" shallow signals can influence final answers without a strict sequential derivation.
Unified Explanation of Phenomena: The framework coherently explains prior observations, including hop-aligned circuits (as dominant recall trajectories in restricted settings), shortcuts (direct recall paths bypassing intermediate steps), and chain-of-thought prompting (contextual enrichment strengthening horizontal recall).
Mechanistic Diagnosis of Errors: The study attributes multi-hop failures to insufficient recall or extraction under multi-hop contexts, even when the underlying single-hop knowledge is present.

Significance and Claims

The paper argues that the hop-aligned circuit hypothesis is insufficient as a general explanation for latent multi-hop reasoning, particularly as the number of hops increases. Instead, the authors claim that LLMs utilize a probabilistic mechanism where knowledge is recalled horizontally across tokens and vertically across layers.

The significance of this work lies in:

Correcting Misconceptions: It challenges the view that LLMs strictly simulate step-by-step reasoning circuits, suggesting instead a more fluid, probabilistic process.
Explaining Robustness and Fragility: It offers a new perspective on why models can sometimes "guess" correctly via shortcuts (relying on shallow intuition-like signals) and why they fail when these signals are weak or when extraction mechanisms falter.
Guiding Future Research: The authors suggest that future work should focus on understanding and potentially strengthening these "intuition-like" shallow recall signals to improve reasoning performance, rather than solely focusing on enforcing strict sequential circuits.

The authors remain modest regarding their claims, noting that their experiments are limited to relatively small-scale models (6B–8B parameters) and that the "intuition-like" nature of shallow recall remains a hypothesis driven by their empirical observations, requiring further theoretical analysis and controlled experiments to fully verify and leverage.