With the growing adoption of machine learning models in critical domains, techniques for explaining differences between models have become essential for trust, debugging, and informed deployment. Previous approaches address this by identifying input transformations that cause divergent predictions or by learning joint surrogate models to align and contrast behaviors. These methods often require access to training data and do not produce natural language explanations. In this paper, we introduce SLED, a framework that generates faithful natural language explanations of when and how two ML models converge or diverge in their predictions. SLED first uses gradient-based optimization to synthesize input samples that highlight divergence and convergence patterns, and then leverages a large language model (LLM) to generate explanations grounded in these synthetic samples. Across both text-based (3 tasks, 7 models) and structured (10 tasks, 4 models) classification tasks, we show that SLED explanations are 18–24% more faithful than the strongest baselines. User studies also indicate that SLED explanations achieve a real-world simulatability of 63.5%. Importantly, SLED requires minimal access to training data and generalizes well to real-world samples, enabling transparent and data-efficient model comparison.

Machine Translation (MT) is an essential tool for communication amongst people across different cultures, yet Named Entity (NE) translation remains a major challenge due to its rarity in occurrence and ambiguity. Traditional approaches, like using lexicons or parallel corpora, often fail to generalize to unseen entities, and hence do not perform well. To address this, we create a silver dataset using the Google Translate API and fine-tune the facebook/nllb200-distilled-600M model with LoRA (LowRank Adaptation) to enhance translation accuracy while also maintaining efficient memory use. Evaluated with metrics such as BLEU, COMET, and M-ETA, our results show that fine-tuning a specialized MT model improves NE translation without having to rely on largescale general-purpose models.

In recent studies, the extensive utilization oflarge language models has underscored the importance of robust evaluation methodologiesfor assessing text generation quality and relevance to specific tasks. This has revealeda prevalent issue known as hallucination, anemergent condition in the model where generated text lacks faithfulness to the source anddeviates from the evaluation criteria. In thisstudy, we formally define hallucination and propose a framework for its quantitative detectionin a zero-shot setting, leveraging our definitionand the assumption that model outputs entailtask and sample specific inputs. In detectinghallucinations, our solution achieves an accuracy of 0.78 in a model-aware setting and 0.61in a model-agnostic setting. Notably, our solution maintains computational efficiency, requiring far less computational resources than other SOTA approaches, aligning with the trendtowards lightweight and compressed models.