Legal case retrieval (LCR) defines relevance beyond text similarity, judgments may describe near-identical events yet be decided under different charges. Existing synthesis pipelines mine training pairs from text similarity and shared charges, which separates positives from negatives but leaves the difficulty of each pair implicit. We present Deka Retrieval, an LCR benchmark on Thai Supreme Court decisions with an LLM relevance cases mining pipeline for a civil-law jurisdiction. We decompose each judgment into case components (Charge, Fact, Reasoning, Decision) and split easy/hard level on exist binary signals and introduce an anomaly variant a conflict judgment that shares the anchor’s charge and facts but reaching a different reasoning and decision. Retriever performance follows query length: Fact > Lawyer expert > Lay people. Fine-tuning gains most on the weakest type, lifting Lay R@10 by 2.6x (7.6 -> 19.5) versus 1.5x on Lawyer and 1.3x on Fact. Training on the shortest query also transfers to the longer ones, but not the reverse. Adding anomaly cases as positives raises the chance of retrieving them at test time and degrades performance relative to training without them. Our dataset and models are available at https://hf.co/collections/VISAI-AI/deka-retrieval.

Large Language Models (LLMs) show impressive reasoning capabilities, yet their understanding of culturally specific social dynamics remains limited. We present KaiToM, a Theory of Mind (ToM) benchmark for Thai cultural contexts, focusing on Kreng-Jai, a uniquely Thai concept of social reluctance and consideration. Extending TactfulToM, we construct 882 conversations (39.2K questions) spanning three Kreng-Jai dimensions: White Lies, Avoidance of Confrontation, and Fear-Based Respect. To disentangle language proficiency from cultural knowledge, the dataset includes Thai scenarios at varying cultural depths, general English scenarios, and English translations of Thai content. We evaluate Thai-specialized (Typhoonv2.1, SEA-LION-v3), open-weight (Gemma-3, DeepSeek-V3.2, Kimi-K2, Qwen), and proprietary models (GPT-4o-mini, GPT-5.1, Gemini-3-Flash). Models achieve only moderate performance, struggle with Fear-Based Respect, and cannot reliably identify specific Kreng-Jai types, a failure shared by Thai-specialized models. Human evaluation shows that native speakers significantly outperform models on culturally loaded tasks. Performance also degrades in deep relative to surface-level cultural contexts, and generalist models often perform better on English translations than on the original Thai, suggesting that models capture surface cues but lack deeper cultural alignment. KaiToM supports the development of culturally aware systems for low-resource languages. Code is available at https://github.com/Augustus2011/KAI-ToM.