We investigate how the amount of fine-tuned capacity affects zero-shot cross-lingual transfer for named entity recognition (NER). Training exclusively on English WikiANN and evaluating on 14 typologically diverse target languages, we find a consistent inverted-U pattern: cross-lingual transfer is worst at both ends of the parameter-efficiency spectrum. Full fine-tuning, which updates all parameters, transfers worst overall and collapses on non-Latin-script languages, while the most parameter-efficient methods, (IA)³ and prefix tuning, also underperform. Methods with an intermediate number of trainable parameters (Low-Rank Adaptation, LoRA; Hu et al., 2022, and Houlsby adapters) transfer best, with LoRA outperforming full fine-tuning by roughly 7 F1 points on average and by over 10 points on languages such as Chinese and Japanese. We find the same inverted-U shape within LoRA alone: a sweep over ranks shows cross-lingual transfer peaking at an intermediate rank, distinct from the higher rank that maximises in-language English performance, revealing a rank–transfer tradeoff. We also examine, and do not find clean support for, the common assumption that better preservation of cross-lingual representational alignment explains PEFT’s transfer benefits: the method that best preserves a standard alignment measure transfers worst, while LoRA’s larger alignment shift coincides with its best transfer. Given known limitations of this alignment measure, we present this as a preliminary, metric-dependent observation that motivates further analysis with more robust alignment metrics, rather than a settled claim.

As Large Language Models (LLMs) approachhuman-level reasoning in English, their performance in low-resource, code-mixed languagesremains surprisingly brittle. We identify Competence Collapse, a distinct pathology wheremodels capable of complex reasoning in English exhibit severe utility degradation whenprompted in Hinglish (Hindi-English). Wequantify this as a Service Gap, observing astatistically significant decline in instructionalquality (∆D ≈ −11.3%, p < 0.001) across9 diverse architectures. Spectral analysis suggests that this stems from a representationaldivergence between the model’s High-UtilityDirection and its Generation Subspace. Tobridge this gap, we propose Cross-LingualActivation Steering (CLAS), an inferencetime intervention that injects a “CompetenceGap Vector” into the residual stream. Evaluated across 6 open-weight models (using alightweight calibration set, N = 50), CLASrecovered utility by ∆D = +2.22 (d = 0.60)while preserving code-mixed fidelity (CMI ≈0.4) and reinforcing safety protocols.