Chain-of-thought (CoT) reasoning is a key tool for explaining and monitoring large language models, but research on its faithfulness has been mostly English-centric. A new arXiv preprint introduces the "Reliance Hypothesis," which proposes that CoT faithfulness actually rises as the language resource level falls—that is, when a language has less text available for training, the model's reasoning trace may be more reliable.
The abstract notes that languages differ in the amount of text available to learn from, but the full details are not yet available. The authors argue that this pattern matters for monitoring, explaining, and safely deploying LLMs, especially in lower-resource languages that are often overlooked in safety research.
If the hypothesis holds, it would challenge the assumption that English-language evaluations of CoT faithfulness generalize to other languages. The study is preliminary, and the abstract does not provide experimental results, but it points to a promising direction for multilingual AI safety.