Reasoning language models (RLMs) increasingly tackle complex problems through long reasoning traces, but identifying when and where failures arise remains difficult. In this work, we study whether reasoning errors are signaled in model representations before they manifest as failed traces. We first show that step-level errors are locally predictable within a short window of neighboring steps, suggesting that models encode early warning signals. However, trace-level failures often result not from isolated errors, but from persistent ones originating earlier in the trace. To study this, we distinguish *recoverable* from *unrecoverable* traces via counterfactual resampling. We find that unrecoverable traces induce final-step representations that are clearly separable from those of recoverable traces. Moreover, the persistent error can be retrospectively decoded from the final-step representation, indicating that error signals propagate and accumulate through the reasoning trace.

Multilingual neural machine translation systems learn to map sentences of different languages into a common representation space. Intuitively, with a growing number of seen languages the encoder sentence representation grows more flexible and easily adaptable to new languages. In this work, we test this hypothesis by zero-shot translating from unseen languages. To deal with unknown vocabularies from unknown languages we propose a setup where we decouple learning of vocabulary and syntax, i.e. for each language we learn word representations in a separate step (using cross-lingual word embeddings), and then train to translate while keeping those word representations frozen. We demonstrate that this setup enables zero-shot translation from entirely unseen languages. Zero-shot translating with a model trained on Germanic and Romance languages we achieve scores of 42.6 BLEU for Portuguese-English and 20.7 BLEU for Russian-English on TED domain. We explore how this zero-shot translation capability develops with varying number of languages seen by the encoder. Lastly, we explore the effectiveness of our decoupled learning strategy for unsupervised machine translation. By exploiting our model’s zero-shot translation capability for iterative back-translation we attain near parity with a supervised setting.