As conversational AI systems increasingly engage users through multi-turn dialogue across diverse cultural contexts, cross-cultural competence, the ability to recognize, respect, and adapt to cultural differences, is essential for effective human-AI interaction. Existing evaluations largely rely on single-turn settings, overlooking the complexities of real-world, multi-turn conversations where cultural cues often emerge gradually. To address this gap, we introduce CultureTalk, a benchmark for evaluating multi-turn cultural reasoning in conversational AI. The dataset contains 1,741 culturally grounded dialogues derived from human-verified cultural knowledge resources and evaluates models’ ability to accumulate cultural context across turns and remain robust to topic shifts. Our evaluation reveals substantial performance gaps between single- and multi-turn settings, highlighting current models’ limitations in sustained culturally grounded context understanding. We further show that fine-tuning on augmented single-turn conversational data and inference scaling significantly improve multi-turn cultural reasoning, yielding a 64.7% relative gain in cultural knowledge entailment with a fine-tuned Qwen3-8B model.

Deploying large language models for specialized tasks requires lengthy prompts combining schema, demonstrations, and instructions — creating a fundamental tension between prompt richness and inference feasibility. We identify and characterize two distinct prompt regimes: pattern-dominant tasks, where demonstrations drive performance through implicit formatting and extraction conventions, and procedure-dominant tasks, where explicit instructions carry the critical signal. To surface this distinction, we design a two-inference diagnostic that compares model performance with instructions alone versus examples alone, revealing which component to target for compression. Guided by this diagnostic, we introduce PRISM, a self-distillation framework that generates silver training data from the model’s own full-prompt outputs and trains lightweight LoRA adapters to internalize the dominant component’s behavior. PRISM recovers 131% of example benefit on CUAD and 151% on MultiWOZ (pattern-dominant), and 39% of instruction benefit on GSM8K (procedure-dominant), eliminating 72–79% of prompt tokens at inference time. Silver data captures 80–94% of gold-labeled performance without human annotations. Beyond compression, PRISM frees context capacity: replacing a long prompt with LoRA weights enables 89.5% compliance with new runtime instructions versus 75.4% with recency prompting. The diagnostic suggests a practical pre-training heuristic to indicate when distillation is likely to succeed, helping practitioners avoid costly training on non-beneficial components.