A new research paper published on arXiv explores the limitations of large language models (LLMs) in performing multi-hop reasoning. The study, "Multi-Hop Knowledge Composition is Bound by Pretraining Exposure," demonstrates that LLMs struggle with complex reasoning tasks even when individual facts are known and retrievable. Researchers found that this failure is rooted in the pretraining phase, as models exposed to compositional contexts during training performed better on unseen compositional questions, while those not exposed continued to fail. AI
IMPACT Highlights a fundamental limitation in LLM reasoning capabilities, suggesting pretraining data composition is key for complex problem-solving.
RANK_REASON Research paper detailing limitations of LLMs in multi-hop reasoning. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Gotit.pub
- Hugging Face
- large-language models
- ScienceCast
- Yannis Karmim
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
