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.
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).
arXiv:2604. 07822v2 Announce Type: replace-cross Abstract: We study implicit reasoning, i.
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.
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.
Compositional generalization remains a persistent limitation of transformer language models: even when a model learns to combine known facts correctly in-distribution (ID), the same compositional mech...
Large Language Models struggle with implicit multi‑hop reasoning, correctly answering individual facts but failing to combine them in a single pass. In a controlled setting, the authors show that this failure persists even with high 1‑hop accuracy, indicating it is due to pretraining exposure rather than missing knowledge. They test nine data‑centric augmentation formats and find that only individuals seen in compositional contexts during pretraining enable transfer to unseen questions, proving exposure to such contexts is necessary for implicit multi‑hop reasoning.
The paper examines the compositionality gap in GPT‑2‑like transformers, where models can combine known facts in‑distribution but fail to generalize to out‑of‑distribution knowledge. By aligning the training distribution so that both ID and OOD atomic facts appear in the same positional format as the second‑hop retrieval step of a known compositional circuit, the authors show that the model can transfer its learned compositional behavior to unseen OOD facts. This demonstrates that the gap is not an inherent architectural limitation but can be addressed through targeted training distribution design.
The paper investigates whether different forms of intermediate computation in large language models—such as token-based traces, pause tokens, and latent reasoning—rely on the same underlying mechanism. By training five variants of GPTNeoX on an extended multi-hop reasoning task, the authors find that while vanilla, Chain-of-Thought, and Pause Token models perform well on in-distribution data, they fail to generalize to longer-hop out-of-distribution problems. In contrast, latent-reasoning models exhibit better depth generalization, with causal analysis revealing a sparse recurrent search circuit that implements forward reachability propagation across the graph.
arXiv:2603.21676v2 Announce Type: replace-cross Abstract: Standard Transformers have a fixed computational depth, limiting their ability to generalize to tasks that require variable-depth reasoning....
Large Language Models fail at implicit multi-hop reasoning: a model answers "When was $X$ born? " and "Who is $Y$'s closest friend?
arXiv:2609.38764v1 Announce Type: new Abstract: Language models are typically pretrained from random initialization. Recent work challenges this convention, showing that a brief warm-up on abstract,...
The paper argues that pattern recognition and step‑by‑step reasoning lie on a spectrum, with large language models (LLMs) learning the latter when the next token depends on only a few preceding tokens. It formalises reasoning traces as paths on a De Bruijn graph, showing that the number of edges is far smaller than the number of possible traces, making step‑by‑step reasoning sample‑efficient. Experiments fine‑tuning Qwen2.5‑1.5B‑Instruct demonstrate that a moderate density of states balances accuracy and robustness, and that real‑world models like Qwen3 retain most of their performance even when attention is limited to a small sliding window.