GPT-4.5 and GPT-4 both face challenges when dealing with low-resource languages, but there are some differences in their approaches and performance.
GPT-4 Performance on Low-Resource Languages
GPT-4, like other large language models, has shown impressive capabilities in high-resource languages but struggles with low-resource languages. Studies have indicated that GPT-4's performance in these languages is not as robust as in English or other well-represented languages[1][3]. This is partly due to the limited training data available for these languages, which results in less effective tokenization and understanding of linguistic nuances[1]. Additionally, GPT-4's safety filters have been found to be less effective when dealing with inputs translated into low-resource languages, making it easier to bypass safeguards[5].
GPT-4.5 Improvements for Low-Resource Languages
GPT-4.5 aims to improve upon GPT-4's capabilities, including its handling of low-resource languages. While specific improvements for low-resource languages are not extensively detailed, GPT-4.5 is noted to outperform GPT-4 in multilingual evaluations. For instance, in evaluations using the MMLU test set translated into 14 languages, including low-resource languages like Yoruba, GPT-4.5 showed better performance compared to GPT-4[9]. This suggests that GPT-4.5 might have enhanced multilingual support and potentially better handling of linguistic nuances in low-resource languages.
However, the improvements in GPT-4.5 are more about overall multilingual performance rather than specific enhancements for low-resource languages. The use of human translators for evaluating multilingual capabilities indicates a focus on ensuring accurate translations, which could indirectly benefit low-resource languages by providing more reliable data for future improvements[9].
Challenges and Future Directions
Despite these improvements, both GPT-4 and GPT-4.5 still face significant challenges with low-resource languages. Fine-tuning and specialized prompting techniques are often recommended to enhance performance in these languages[7]. The disparity in safety and performance between high-resource and low-resource languages highlights the need for more inclusive training data and safety protocols that account for linguistic diversity[5].
In summary, while GPT-4.5 offers some improvements over GPT-4 in multilingual contexts, specific enhancements for low-resource languages are not extensively detailed. Further research and development are necessary to address the persistent challenges in these languages.
Citations:
[1] https://aclanthology.org/2024.findings-emnlp.920.pdf
[2] https://topmostads.com/gpt-4-5-vs-gpt-5-release/
[3] https://www.asianlp.sg/conferences/ialp2024/proceedings/papers/IALP2024_P027.pdf
[4] https://www.techtarget.com/searchenterpriseai/tip/GPT-35-vs-GPT-4-Biggest-differences-to-consider
[5] https://arxiv.org/pdf/2310.02446.pdf
[6] https://teamai.com/blog/large-language-models-llms/understanding-different-chatgpt-models/
[7] https://aclanthology.org/2025.coling-main.559.pdf
[8] https://www.reddit.com/r/ClaudeAI/comments/1dqj1lg/claude_35_sonnet_vs_gpt4_a_programmers/
[9] https://cdn.openai.com/gpt-4-5-system-card.pdf