aiqbench analysis
LLMs' inability to perform implicit multi-hop reasoning stems from a lack of compositional contexts during pretraining rather than a deficit in individual fact retrieval. The study confirms that exposure to such contexts is a necessary condition for this capability, as data-centric augmentations fail to transfer reasoning skills to individuals absent from these contexts. This finding highlights a fundamental architectural limitation in current models that must be addressed through specific pretraining strategies to achieve robust compositional reasoning.
Multi-Hop Knowledge Composition is Bound by Pretraining Exposure
arXiv:2606.09338v1 Announce Type: new Abstract: Large Language Models fail at implicit multi-hop reasoning: a model answers "When was $X$ born?" and "Who is $Y$'s closest friend?" correctly but fails on "When was $Y$'s closest friend born?" in a single forward pass, even when both facts are perfectly memorized and individually retrievable. We study this failure in a controlled natural language setting with a strict separation between individuals exposed to compositional contexts during pretraining and those that never appear in any such context. We confirm that compositional failure persists even at 97% 1-hop accuracy, establishing the gap as a pretraining failure rather than a knowledge absence. We propose and test nine data-centric augmentation formats and find that compositional pretraining transfers to unseen questions for exposed individuals, but never to individuals absent from compositional pretraining, suggesting that exposure to compositional contexts during pretraining is a necessary condition for implicit multi-hop reasoning.
Abstract:Large Language Models fail at implicit multi-hop reasoning: a model answers "When was $X$ born?" and "Who is $Y$'s closest friend?" correctly but fails on "When was $Y$'s closest friend born?" in a single forward pass, even when both facts are perfectly memorized and individually retrievable. We study this failure in a controlled natural language setting with a strict separation between individuals exposed to compositional contexts during pretraining and those that never appear in any such context. We confirm that compositional failure persists even at 97% 1-hop accuracy, establishing the gap as a pretraining failure rather than a knowledge absence. We propose and test nine data-centric augmentation formats and find that compositional pretraining transfers to unseen questions for exposed individuals, but never to individuals absent from compositional pretraining, suggesting that exposure to compositional contexts during pretraining is a necessary condition for implicit multi-hop reasoning.
Computation and Language (cs.CL)
arXiv:2606.09338 [cs.CL]
(or arXiv:2606.09338v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.09338 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 239 words
Related coverage
Multi-Hop Knowledge Composition is Bound by Pretraining Exposure
arXiv · Sep 17, 2026
What's Continual Learning, and Why Might We Expect To See It In Advanced LLM Agents?
LessWrong · Jun 12, 2026
Reflection’s Beam model hits 501 billion parameters
Quantum Zeitgeist · Oct 5, 2026
To Learn is to Wander: Learning Across Graphs and Tasks with Random Walks
arXiv · Oct 6, 2026