Large language models can solve complex multi‑hop tasks but often fail on simple two‑hop queries, even when each hop is individually correct. In a controlled symbolic setting, the authors find that models generalize reliably when the second hop follows the training distribution, but always fail when it deviates. Mechanistic analysis shows that successful generalization relies on consistent intermediate representations across contexts, whereas failures arise from a mismatch between lower‑layer representation construction and upper‑layer mapping to outputs. The study proposes a recurrent‑style training strategy that improves out‑of‑distribution two‑hop generalization.
By Zili Zhang, Yilin Wang, Heng Wang, Herun Wan, Minnan Luo
The paper investigates how large language models (LLMs) perform multi‑hop reasoning and challenges the prevailing hop‑aligned circuit hypothesis, which posits that bridge entities are computed sequentially across layers. Through systematic analyses of real‑world multi‑hop queries, the authors discover a phenomenon called layer‑order inversion, where later‑hop answer entities become decodable earlier than bridge entities, and this effect grows with the number of hops. They propose a 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, and validate this framework with probing analyses that reinterpret prior evidence, explain chain‑of‑thought gains, and diagnose multi‑hop failures.
By Xukai Liu, Ye Liu, Jipeng Zhang, Yanghai Zhang, Kai Zhang, Qi Liu
arXiv:2602.22453v4 Announce Type: replace
Abstract: Retrieval heads, a subset of attention heads in Transformers, were studied in English, showing its crucial role in retrieving information from the...
By Shaswat Patel, Vishvesh Trivedi, Yue Han, Yihuai Hong, Eunsol Choi
The paper investigates attention heads in multilingual Transformer models, distinguishing between retrieval heads that pull information from context and a newly identified class called Retrieval‑Transition heads (RTH) that direct the model toward a specific target language. Experiments across four multilingual benchmarks and two model families show that masking RTHs causes a larger performance drop than masking retrieval heads, indicating RTHs are crucial for chain‑of‑thought reasoning in multilingual LLMs. The study thus clarifies which attention heads are responsible for mapping to target languages, advancing our understanding of multilingual language models.
By Shaswat Patel, Vishvesh Trivedi, Yue Han, Yihuai Hong, Eunsol Choi
arXiv:2609.37104v1 Announce Type: cross
Abstract: Open-source reasoning models provide unequal access to reasoning capability across languages. When a model can solve a problem but cannot deliver a c...
By Hongyang Li, Xiao Li, Caesar Wu, Gr\'egoire Danoy, Pascal Bouvry
arXiv:2608. 08032v1 Announce Type: new Abstract: Safety alignment in multilingual models is uneven: a model that reliably refuses a harmful request in English will often comply with the same request in a lower-resource language.
By Ramakrishna P. Kompella, Aadit Mahajan
The paper evaluates confidence‑based routing of small‑to‑large language models across multiple tasks, languages, and model families, covering 72 configurations and over 42,000 predictions. It finds that confidence routing improves performance on reading comprehension and topic classification datasets, but offers little benefit on a knowledge‑reasoning benchmark where small models remain confidently wrong. The study also shows that global confidence routing can create large language‑specific performance gaps, which equal‑quota routing mitigates with minimal overall accuracy loss.
By Md Sabbir Hossen, Anichur Rahman, Pabon Shaha, Airin Afroj Aishi, Andrew H. Sung, Md Shohel Rana
arXiv:2509. 24653v2 Announce Type: replace-cross Abstract: Large Language Models (LLMs) excel at multi-hop reasoning in distribution, yet fail on unseen compositions, a phenomenon known as the curse of two-hop reasoning.
By Pengxiao Lin, Zheng-An Chen, Zhi-Qin John Xu
The paper investigates whether different latent language probes—GMM-based representation probes and decoding-based probes—measure the same phenomenon in multilingual language models. Across various model families, training regimes, domains, tasks, checkpoints, and up to 27 languages, the authors find systematic disagreement: representation probes indicate earlier cross‑lingual mixing, while decoding probes reveal sharper, English‑biased language signals. These differences correlate with model multilinguality and training progression but remain relatively stable across domains, suggesting that current probes capture distinct aspects of multilingual processing rather than a single internal lingua franca.
By Deniz Bayazit, Badr AlKhamissi, Antoine Bosselut
arXiv:2607. 00341v1 Announce Type: cross Abstract: Large language models achieve strong performance on many reasoning tasks when allowed to externalize intermediate steps as Chain-of-Thought (CoT).
By Hengyu Fu, Tianyu Guo, Zixuan Wang, Hanlin Zhu, Jason D. Lee, Jiantao Jiao, Stuart Russell, Song Mei
The paper introduces CLLPU, a multilingual benchmark for evaluating how well large language models can unlearn specific knowledge while controlling its propagation across languages. CLLPU defines two forgetting scenarios—common-goal forgetting, which requires suppression across all languages, and language-conditioned forgetting, which limits suppression to a single language. Using 800 knowledge-unit pairs and 72,000 QA instances in ten languages, the authors test six methods on Llama‑3.1‑8B‑Instruct and find that universal suppression often fails, while language‑specific suppression can unintentionally spread to other languages, highlighting the difficulty of propagation control in multilingual unlearning.
By Pengyang Shao, Chuanpeng Lu, Wei Qin, Yanzheng Jin, Xiaohao Liu, Xi Ai, Kenji Kawaguchi, Richang Hong
arXiv:2608.20362v1 Announce Type: cross
Abstract: Reinforcement learning with verifiable rewards (RLVR) is a standard recipe for training large language models on mathematical reasoning, where an ans...
By Chenyu Zhou, Qiliang Jiang, Xu Zhou