GPT-5 vs Other LLMs in Long Short-Context Performance
Open accessAbstract
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.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
Every retrieved artifact competes for limited context budget, and excessive context loading can degrade performance through information overload or the 'lost in the middle' phenomenon.
“Every retrieved artifact competes for limited context budget, and excessive context loading can degrade performance through information overload or the 'lost in the middle' phenomenon [Corallo and Papotti, 2026, Mishra et al., 2026, Esmi et al., 2025].”
From Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering (2026) · cites this paper as Esmi et al., 2025
The Abstract (p. 1) and Conclusion (p. 9) report that performance degraded as input volume increased, and Prior Literature (p. 2) describes the ‘lost in the middle’ phenomenon. The paper does not discuss retrieved artifacts competing for a context budget, and its Conclusion says newer models have largely mitigated ‘lost in the middle.’ (LLM, conf=0.86)
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
Lost in the Middle: How Language Models Use Long Contexts
Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00638 · Cited by 1,612 · cited 1× · see citing works
“…highlights how models exhibit primacy and recency biases, performing best when relevant information is at the beginning or end of the input context, while struggling to access and utilize data in the middle, as evidenced in tasks like multi-document question answering and key-value retrieval [1].”
Large language models to identify social determinants of health in electronic health records
Guevara, M., Chen, S., Thomas, S., Chaunzwa, T. L., Franco, I., Kann, B. H., Moningi, S., Qian, J. M., Goldstein, M., Harper, S., et al. (2024) · npj Digital Medicine · doi:10.1038/s41746-023-00970-0 · Cited by 278 · see citing works
LoongServe: Efficiently Serving Long-Context Large Language Models with Elastic Sequence Parallelism
Wu, B., Liu, S., Zhong, Y., Sun, P., Liu, X., & Jin, X. (2024) · Proceedings of the ACM SIGOPS 30th Symposium on Operating Systems Principles · doi:10.1145/3694715.3695948 · Cited by 62 · see citing works
Yang, X., Wang, Z., Wang, Q., Ke, W., Zhang, K., & Shi, J. (2024) · International Journal of Web Information Systems · doi:10.1108/ijwis-12-2023-0256 · Cited by 31 · see citing works
Long-context LLMs Struggle with Long In-context Learning
Li, T., Zhang, G., Do, Q. D., Yue, X., & Chen, W. (2024) · arXiv · doi:10.48550/arxiv.2404.02060 · Cited by 31 · see citing works
Ecosystem of Large Language Models for Code
Yang, Z., Shi, J., Dévanbu, P., & Lo, D. (2025) · ACM Transactions on Software Engineering and Methodology · doi:10.1145/3731753 · Cited by 18 · see citing works
XBRL Agent: Leveraging Large Language Models for Financial Report Analysis
Han, S., Kang, H., Jin, B., Liu, X.-Y., & Yang, S. Y. (2024) · ACM International Conference on AI in Finance (ICAIF) · doi:10.1145/3677052.3698614 · Cited by 12 · see citing works
Stress detection through prompt engineering with a general-purpose LLM
Esmi, N., Shahbahrami, A., Nabati, Y., Rezaei, B., Gaydadjiev, G. N., & de Jonge, P. (2025) · Acta Psychologica · doi:10.1016/j.actpsy.2025.105462 · Cited by 4 · see citing works
Self-citationHosseini, P., Castro, I., Ghinassi, I., & Purver, M. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2408.01866 · Cited by 1 · see citing works
Too Long, Didn't Model: Decomposing LLM Long-Context Understanding With NovelsPreprint
Hamilton, S., Hicke, R. M. M., Wilkens, M. J., & Mimno, D. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2505.14925
Exploring Context Window of Large Language Models via Decomposed Positional Vectors
Dong, Z., Li, J., Men, X., Zhao, W., Wang, B., Tian, Z., Chen, W., & Wen, J.-R. (2024) · Advances in Neural Information Processing Systems 37 · doi:10.52202/079017-0330
Flagged first, then by citations
Cited by
Sorted by citations
Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering
Zhou, C., Chai, H., Chen, W., Guo, Z., Shan, R., Song, Y., Xu, T., Yang, Y., Yu, A., Zhang, W., et al. (2026) · arXiv · doi:10.48550/arxiv.2604.08224 · Cited by 1 · see citing works
HBM-HBF-Centric Memory Pooling Architecture With Custom Base Die for Terabyte-Scale LLM Inference
Park, J., An, H., Suh, H., Yoon, Y., LEE, H., & Kim, J. (2026) · IEEE Computer Architecture Letters · doi:10.1109/lca.2026.3703982 · Cited by 2 · see citing works
- Authors
- Esmi, N., Nezhad-Moghaddam, M., Borhani, F., Shahbahrami, A., Daemdoost, A., & Gaydadjiev, G.
- Affiliations
- University of Groningen (NL)
- University of Guilan (IR)
- Khazar University (AZ)
- Delft University of Technology (NL)
- Venue
- 2025 3rd International Conference on Foundation and Large Language Models (FLLM)
- Publisher
- IEEE
- Published
- 2025-11-25
- Pages
- 129-133
- OpenAlex
- W7130584504
- Cited by
- 1
- References
- 7
- Citations by year
- 2026: 1
- FWCI
- 0.95
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- International Conference on Foundation and Large Language Models (FLLM) (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Social Sciences › Psychology › Social Psychology › Mental Health via Writing
- Keywords
- large language models, GPT-5, GPT-4, Grok-4, Gemini 2.5, context window, depression detection, social media posts, recipe retrieval, mathematical problem solving, precision and accuracy, model evaluation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
