Context-Aware Language Models: Assessing Inference and Reasoning for Incomplete Messages

Recent advancements in large language models (LLMs) have enhanced contextual reasoning in natural language processing (NLP), yet challenges remain in handling incomplete, ambiguous, or noisy input. This paper proposes GPT-4o Mini as an effective model for sentence reconstruction and evaluates it alongside Gemini 2.5 Pro, Phi-4, and DeepSeek-V2 on a benchmark dataset of manually curated sentences with structural errors, fragmentation, and ambiguity. Model outputs are assessed using BLEU, Edit Distance, ROUGE-2, and BERTScore. Results show that GPT-4o Mini consistently delivers high-quality reconstructions across most scenarios, achieving BLEU scores up to 0.6800, Edit Distances as low as 4, and BERTScore F1 values near 0.978–0.980. DeepSeek-V2 occasionally matches or slightly surpasses GPT-4o Mini in specific cases, while Gemini 2.5 Pro and Phi-4 exhibit lower robustness under structural degradation. These findings highlight GPT-4o Mini as an effective model for recovering meaning from noisy or incomplete text, establishing its suitability for real-world, context-aware NLP applications where maintaining coherence and semantic integrity is critical.