External context enables language models to accomplish tasks which rely on information beyond what is captured during pretraining. A standard way to improve contextual faithfulness in these settings is to fine-tune on synthetic, modified questions for which the model must adhere to provided contexts inconsistent with information that may have been seen during training. While this approach reliably improves contextual faithfulness, we find that adapting to a narrowly targeted counterfactual signal yields degradation on other faithfulness evaluations and regression on tasks that evaluate a model’s parametric world knowledge and reasoning. We find that data augmentation, in which a model is trained on both parametrically consistent and counterfactual examples improves performance on faithfulness tasks with reduced regression on parametric world knowledge and reasoning tasks. We find that these improvements are largely consistent across post-training objectives (i.e. SFT, DPO) and across ground-truth, counterfactual data augmentation ratios.

Large Language Models (LLMs) are increasingly deployed in multilingual settings, yet most bias evaluation remains English-centric and overlooks how bias manifests within reasoning. We present a systematic study of social bias in both predictions and chain-of-thought reasoning across English, Dutch, Spanish, and Turkish using the MBBQ benchmark. We evaluate instruction-tuned, CoT-prompted, and reasoning-native models under supervised fine-tuning and preference optimization, using accuracy, F1, bias metrics, and a novel reasoning-level language drift measure. We find that (1) bias varies substantially across languages, with consistent degradation in non-English settings, (2) reasoning traces often introduce additional stereotype-driven signals beyond final outputs, and (3) English-trained debiasing methods fail to generalize reliably, with preference optimization introducing cross-lingual trade-offs. We further show that performance gains in multilingual settings are frequently driven by implicit reliance on English-centric reasoning, revealed through increased language drift. Together, our results demonstrate that multilingual fairness cannot be inferred from English performance and requires reasoning-aware, language-specific evaluation and alignment.