ACL Findings2025
Do Large Language Models Perform Latent Multi-Hop Reasoning without Exploiting Shortcuts?
Sohee Yang, Nora Kassner, Elena Gribovskaya, Sebastian Riedel, Mor Geva
Abstract
We evaluate how well Large Language Models (LLMs) latently recall and compose facts to answer multi-hop queries like "In the year Scarlett Johansson was born, the Summer Olympics were hosted in the country of". One major challenge in such evaluation is that LLMs may have developed shortcuts by encountering the head entity "Scarlett Johansson" and the answer entity "United States" in the same training sequences or merely guess the answer based on frequency-based priors. To prevent shortcuts, we exclude test queries where the head and answer entities might have co-appeared during training. Through careful selection of relations and facts and systematic removal of cases where models might guess answers or exploit partial matches, we construct an evaluation dataset SOCRATES (SHORTCUT-FREE LATENT REASONING). We observe that LLMs demonstrate promising latent multi-hop reasoning abilities without exploiting shortcuts, but only for certain types of queries. For queries requiring latent recall of countries as the intermediate answer, the best models achieve 80% latent composability, but this drops to just 5% for the recall of years. Comparisons with Chainof-Thought highlight a significant gap between the ability of models to reason latently versus explicitly. Analysis reveals that latent representations of the intermediate answer are constructed more often in queries with higher latent composability, and shows the emergence of latent multi-hop reasoning during pretraining. 1
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
7 top-tier papers have cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- Measuring and Narrowing the Compositionality Gap in Language Models cites Large Language Models are Zero-Shot Reasoners
- Deep sequence models tend to memorize geometrically; it is unclear why cites Measuring and Narrowing the Compositionality Gap in Language Models
- Deep sequence models tend to memorize geometrically; it is unclear why cites Locating and Editing Factual Associations in GPT
- Deep sequence models tend to memorize geometrically; it is unclear why cites How do Transformers Learn Implicit Reasoning?
- Deep sequence models tend to memorize geometrically; it is unclear why cites Roll the dice & look before you leap: Going beyond the creative limits of next-token prediction
- Deep sequence models tend to memorize geometrically; it is unclear why cites Faith and Fate: Limits of Transformers on Compositionality
- Bilinear representation mitigates reversal curse and enables consistent model editing cites Locating and Editing Factual Associations in GPT
Cited by top-tier papers7
- How do Transformers Learn Implicit Reasoning?Jiaran Ye, Zijun Yao, Zhidian Huang, Liangming Pan et al.NeurIPS 2025 · 17 citations
- Deep sequence models tend to memorize geometrically; it is unclear whyShahriar Noroozizadeh, Vaishnavh Nagarajan, Elan Rosenfeld, Sanjiv KumarICML 2026 · 11 citations
- Programming by Backprop: An Instruction is Worth 100 Examples When Finetuning LLMsJonathan Cook, Silvia Sapora, Arash Ahmadian, Akbir Khan et al.ICLR 2026 · 2 citations
- Bilinear representation mitigates reversal curse and enables consistent model editingDong-Kyum Kim, Minsung Kim, Jea Kwon, Nakyeong Yang et al.ICLR 2026 · 1 citation
- Roll the dice & look before you leap: Going beyond the creative limits of next-token predictionVaishnavh Nagarajan, Chen Henry Wu, Charles Ding, Aditi RaghunathanICML 2025
Builds on32
- Large Language Models are Zero-Shot ReasonersTakeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo et al.NeurIPS 2022 · 8,168 citations
- Locating and Editing Factual Associations in GPTKevin Meng, David Bau, Alex Andonian, Yonatan BelinkovNeurIPS 2022 · 3,415 citations
- Efficient Memory Management for Large Language Model Serving with PagedAttentionWoosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng et al.SOSP 2023 · 1,016 citations
- Faith and Fate: Limits of Transformers on CompositionalityNouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li et al.NeurIPS 2023 · 728 citations
- Measuring and Narrowing the Compositionality Gap in Language ModelsOfir Press, Muru Zhang, Sewon Min, Ludwig Schmidt et al.EMNLP Findings 2023 · 296 citations
Related papers
- Investigating Multi-Hop Factual Shortcuts in Knowledge Editing of Large Language ModelsTianjie Ju, Yijin Chen, Xinwei Yuan, Zhuosheng Zhang et al.ACL 2024 · 2 citations
- Seemingly Plausible Distractors in Multi-Hop Reasoning: Are Large Language Models Attentive Readers?Neeladri Bhuiya, Viktor Schlegel, Stefan WinklerEMNLP 2024 · 2 citations
- CofCA: A STEP-WISE Counterfactual Multi-hop QA benchmarkJian Wu, Linyi Yang, Zhen Wang, Manabu Okumura et al.ICLR 2025
- Do Large Language Models Latently Perform Multi-Hop Reasoning?Sohee Yang, Elena Gribovskaya, Nora Kassner, Mor Geva et al.ACL 2024
- Unveiling Internal Reasoning Modes in LLMs: A Deep Dive into Latent Reasoning vs. Factual Shortcuts with Attribute Rate RatioYiran Yang, Haifeng Sun, Jingyu Wang, Qi Qi et al.EMNLP 2025
