The field of language models is moving towards improving implicit multi-hop reasoning capabilities, which enable models to solve complex tasks in a single forward pass without explicitly verbalizing intermediate steps. Researchers are exploring the limits of this capability, including the amount of training data required and the necessary depth of transformer layers. Recent studies have shown that implicit multi-hop reasoning can be achieved with large amounts of training data, but the required data grows exponentially with the complexity of the task. Furthermore, studies have found that pretraining on procedural data can instil modular structures for algorithmic reasoning in language models, and that these structures can be composed to jointly reinforce multiple capabilities. The development of diagnostic tools has also provided new insights into the interpretability of implicit multi-hop reasoning in language models, helping to clarify how complex reasoning processes unfold internally. Notable papers include:

• A study that identifies the benefits of curriculum learning in mitigating the data requirement for implicit multi-hop reasoning.
• Research that discovers modular structures for algorithmic reasoning in small transformers pretrained on procedural data, and shows that these structures can be composed to improve robustness and data efficiency.