Humanitarian organizations must adapt text classifiers to new crises with small annotation budgets, yet conventional active learning optimizes model uncertainty without representing operational criticality or coverage across disaster events. We study an impact-aware batch acquisition policy that combines expected criticality, uncertainty, representation-space diversity, and event balance. Our evaluation crosses three recent open embedding backbones, three representation treatments, five acquisition strategies, five paired seeds, and total budgets of 50–500 labels. We separate exploratory development, event-held-out robustness on HumAID, and an untouched cross-dataset confirmation on CrisisMMD. On CrisisMMD, impact-aware selection with a conservative residual adapter improves Qwen’s area under the learning curve over random sampling by 0.089 for critical-category recall, 0.026 for macro-F1, and 0.032 for worst-event macro-F1. It raises the critical share of the 500-example labeled set from 3.4% to 26.8% and allocates annotations more evenly across events. Critical-recall gains also hold for Harrier and EmbeddingGemma on CrisisMMD and transfer to unseen HumAID events for Harrier, but not Qwen, exposing a consequential representation–dataset interaction. No configuration dominates every operational objective. Finally, 512-dimensional Qwen halves embedding storage while remaining within 0.009 AULC of its full representation across primary outcomes and both test settings. These results support impact-aware acquisition as a useful, but not always-on, policy and motivate reporting Pareto trade-offs rather than a single aggregate score.

Romanized writing is a common interface for speakers of Indic languages, but it is intrinsically many-to-one: several informal spellings express the same underlying utterance. Existing evaluations typically use one canonical transliteration and may therefore overestimate robustness. We introduce a 4,000-set convention-profile benchmark alongside native-script, human-Romanized, and human-attested equivalence views in Bengali, Hindi, Tamil, and Telugu. Across twelve pretrained and instruction-tuned models from three families, we measure token fragmentation, layer-wise representation dispersion, language-modeling loss, and cross-script retrieval instability. We formalize the target property asemphselective stability—equivalent spellings must contract while distinct meanings stay separated—operationalize it with a normalized within-equivalence distance, and test whether it predicts retrieval failure beyond token fertility. We find that cross-script invariance is aemphtransient phase in depth whose location is architecture-dependent, that representation distance predicts inconsistency better than fertility (ROC-AUC .774 vs. .634 on the small instruction model), and that a lightweight low-rank repair trained only on canonical and synthetic variants transfers to unseen natural spellings, held-out convention families, and unseen languages without improving Romanized reading comprehension. A per-language decomposition shows that the phase is also largely language-conditional: one Gemma 9B layer retrieves 90.4

Shortcut learning remains a major obstacle to robust NLP systems: models can achieve high in-distribution accuracy by relying on surface cues that fail under distribution shift. We study whether shortcut reliance can be diagnosed and mitigated in small instruction-tuned language models using a simple representation-level quantity. We fine-tune Gemma 3 1B Instruct and Llama 3.2 1B on two synthetic sentiment shortcuts in SST-2 and one natural shortcut in MNLI based on lexical overlap. During training, we fit linear probes for the task label and the shortcut attribute at every layer and define CDRE as the absolute cosine similarity between the two probe directions. Across settings, increasing shortcut prevalence produces a sharp rise in the robustness gap between shortcut-aligned and shortcut-free test sets, and higher deep-layer CDRE tracks this degradation. At a 99% shortcut ratio, Llama’s clean accuracy on capitalization-biased SST-2 drops from 93.2% at 0% bias to 49.0%, while Gemma drops from 91.8% to 60.2%. A CDRE-regularized objective substantially improves robustness for capitalization and lexical-overlap shortcuts, but offers little benefit for a speaker-prefix shortcut whose learned directions are already nearly orthogonal. These results show that probe-derived representation entanglement provides a reliable signal of harmful shortcut reliance and offers a practical criterion for determining when shortcut mitigation is likely to be effective.