Recent large language models (LLMs) demonstrate impressive capabilities in handling long contexts, some exhibiting near-perfect recall on synthetic retrieval tasks. However, these evaluations have mainly focused on English text and involved a single target sentence within lengthy contexts. Our work investigates how LLM performance generalizes to multilingual settings with multiple hidden target sentences. We create a new dataset -- mLongRR -- to comprehensively evaluate several multilingual long-context LLMs on retrieval and reasoning tasks across five languages: English, Vietnamese, Indonesian, Swahili, and Somali. These languages share the Latin script but belong to distinct language families and resource levels. Our analysis reveals a significant performance gap between languages. The best-performing models such as Gemini-1.5 and GPT-4o, achieve around 96% accuracy in English to around 36% in Somali with a single target sentence. However, this accuracy drops to 40% in English and 0% in Somali when dealing with three target sentences. Our findings highlight the challenges long-context LLMs face when processing longer contexts, an increase in the number of target sentences, or languages of lower resource levels.
Evaluating Multilingual Long-Context Models for Retrieval and Reasoning
LLMs exhibit varying performance across multilingual settings and multiple target sentences in long-context retrieval and reasoning tasks.
- Year
- 2024
- Venue
- arXiv 2024
- Authors
- 5
- Hosting
- Abstract onlyARXIV-DEFAULT
Cite
Notes
Only stored in your browser.
Attribution
- Abstract & full text
- arxiv.org/abs/2409.18006v3ARXIV-DEFAULT
- TL;DR
- Semantic Scholar