Chain of thoughts, Tree of thoughts and Retrieval Augmented thoughts - Laboratoire d'informatique de l'X (LIX)
Reports (Technical Report) Year : 2024

Chain of thoughts, Tree of thoughts and Retrieval Augmented thoughts

Davide Buscaldi
  • Function : Author
  • PersonId : 1430844
Leo Liberti

Abstract

This paper presents novel contributions to improving large language models (LLMs) through the integration and enhancement of three reasoning frameworks: Chain-of-Thoughts (CoT), Tree-of-Thoughts (ToT), and Retrieval-Augmented Thoughts (RAT). CoT generates linear intermediate reasoning steps, ToT introduces a branching structure for exploring multiple reasoning pathways, and RAT incorporates real-time external knowledge retrieval. A key innovation discussed is the use of decision trees, both handcrafted and automatically generated, to guide and validate the reasoning processes of CoT and ToT. This approach enhances the interpretability, accuracy, and adaptability of LLMs, particularly in complex problem-solving tasks like medical diagnostics and financial forecasting.
Fichier principal
Vignette du fichier
Tech_CoT_ToT_RAT.pdf (331.03 Ko) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04748901 , version 1 (22-10-2024)

Licence

Copyright

Identifiers

  • HAL Id : hal-04748901 , version 1

Cite

Nabil Moncef Boukhatem, Davide Buscaldi, Leo Liberti. Chain of thoughts, Tree of thoughts and Retrieval Augmented thoughts. LIX, École Polytechnique. 2024. ⟨hal-04748901⟩
0 View
0 Download

Share

More