Even when explicit Personally Identifiable Information (PII) is removed from LLM queries, implicit contextual cues can still reveal sensitive user attributes. Yet current Privacy-Conscious Delegation frameworks overlook this threat entirely, and existing PII span-based evaluation metrics cannot address it. We address this gap on both fronts. First, we introduce Attribute Inference Leakage (AIL), a label-free metric that quantifies implicit privacy leakage through set-based semantic comparison of attributes inferred before and after sanitisation. Using AIL, we provide the first empirical evidence in Privacy-Conscious Delegation that even perfect PII removal eliminates only 18% of implicit leakage relative to direct delegation, leaking the vast majority of inferable attributes. Second, we propose GIST, a training-free approach for context-aware generalisation to protect both explicit and implicit privacy while preserving response quality. Experiments across six local models (1B to 8B parameters) show GIST consistently achieves a Pareto improvement over state-of-the-art PAPILLON on all three evaluation dimensions of quality, explicit and implicit leakage. GIST also consistently reduces implicit leakage below the explicit leakage lower bound (perfect PII removal), hence GIST addresses a privacy dimension that PII-centric methods cannot.

Named Entity Recognition (NER) for low-resource variants of English remains challenging, as most NER models are trained on datasets predominantly focused on American or British English. While recent work has shown that proprietary Large Language Models (LLMs) can perform NER effectively in low-resource settings through in-context learning, practical deployment is limited by their high computational costs and privacy concerns. Open-source Small Language Models (SLMs) offer promising alternatives, but the tendency of these Language Models (LM) to hallucinate poses challenges for production use. To address this, we introduce SLENDER, a novel output format for LM-based NER that achieves a three-fold reduction in inference time on average compared to JSON format, which is widely used for structured outputs. Our approach using Gemma-2-9B-it with the SLENDER output format and constrained decoding in zero-shot settings outperforms the en_core_web_trf model from SpaCy, an industry-standard NER tool, in all five regions of the Worldwide test set.