GPT-5 vs Other LLMs in Long Short-Context Performance

TL;DR

This paper addresses the gap between theoretical long-context capacity and actual performance on complex, detail-rich tasks by empirically evaluating four state-of-the-art LLMs—Grok-4, GPT-4, Gemini 2.5, and GPT-5—across three datasets that vary in length and content. Using prompts designed to test retrieval and classification over long inputs, the study reveals that accuracy can drop sharply when inputs exceed roughly 70K tokens, yet GPT-5 maintains high precision (~95%) in sensitive tasks like depression detection, indicating a favorable precision-robustness trade-off even as overall accuracy declines. The work also finds that the long-standing “lost in the middle” problem is largely mitigated in newer models, underscoring advances in extended-context reasoning, while also demonstrating the importance of evaluating precision and stability across input scales, not just accuracy. These findings inform practical deployment of long-context LLMs in real-world domains and guide future architecture design toward improved information retention and multi-source coherence for large-scale, heterogeneous data analyses.

Abstract

With the significant expansion of the context window in Large Language Models (LLMs), these models are theoretically capable of processing millions of tokens in a single pass. However, research indicates a significant gap between this theoretical capacity and the practical ability of models to robustly utilize information within long contexts, especially in tasks that require a comprehensive understanding of numerous details. This paper evaluates the performance of four state-of-the-art models (Grok-4, GPT-4, Gemini 2.5, and GPT-5) on long short-context tasks. For this purpose, three datasets were used: two supplementary datasets for retrieving culinary recipes and math problems, and a primary dataset of 20K social media posts for depression detection. The results show that as the input volume on the social media dataset exceeds 5K posts (70K tokens), the performance of all models degrades significantly, with accuracy dropping to around 50-53% for 20K posts. Notably, in the GPT-5 model, despite the sharp decline in accuracy, its precision remained high at approximately 95%, a feature that could be highly effective for sensitive applications like depression detection. This research also indicates that the "lost in the middle" problem has been largely resolved in newer models. This study emphasizes the gap between the theoretical capacity and the actual performance of models on complex, high-volume data tasks and highlights the importance of metrics beyond simple accuracy for practical applications.

Figures (7)

• Figure 1: Growth in the number of parameters of large language models surpassing the trillion mark, alongside the expansion of their context windows..
• Figure 2: The workflow for our experiment, detailing the four primary phases: Dataset Preparation, Prompt Injection, Model Execution, and Performance Evaluation.
• Figure 3: Evaluation results for the Culinary Recipes dataset across four subsets: 125, 250, 500, and 1K recipes.
• Figure 4: Evaluation results for the Math–Problem dataset across four subsets: 125, 250, 500, and 1K problems. The purple columns represent accuracy, while the orange columns represent precision.
• Figure 5: Evaluation results for the Depress.-Twitter dataset across four subsets: 5K, 10K, 15K, and 20K Tweets. The purple columns represent accuracy, while the orange columns represent precision.