Skip to content

Latest commit

 

History

History
20 lines (15 loc) · 2.27 KB

2404.05966.md

File metadata and controls

20 lines (15 loc) · 2.27 KB
Raw

Background

  • Background The paper discusses how Large Language Models (LLMs) like GPT, LLaMA, and Claude improve their reasoning task capabilities through distinct prompting strategies and instructional guidance.

  • Existing Work Previous methodologies like Chain-of-Thought (CoT) and its extensions, such as Tree-of-Thoughts and Graph-of-Thoughts, only enable exploration within a fixed set of candidates without supporting continuous revisions of original answers in later steps. This limits the capability of LLMs to handle problems necessitating frequent revisions and modifications.

Core Contributions

  • Introduced a novel graph-based framework

    • Challenge 1: Creating a reasoning network that allows continuous self-revision Existing LLM reasoning strategies do not support continuous revision of previous outputs during the reasoning process. The paper presents THOUGHTSCULPT, a new graph-based framework that uses a self-revision mechanism to enable LLMs to refine and improve upon their previous outputs iteratively.

    • Challenge 2: Addressing the vast search space in text generation For cases where all possible paths cannot be exhaustively explored with conventional graph search algorithms like DFS, BFS, and A* Search, researchers employ Monte Carlo Tree Search (MCTS), a powerful heuristic technique that navigates the search space effectively and provides high-quality solutions.

Implementation and Deployment

THOUGHTSCULPT consists of three core modules: thought evaluator, thought generator, and decision simulator. It has been tested across three challenging tasks: Story Outline Improvement, Mini-Crosswords Solving, and Constrained Generation. Compared to existing state-of-the-art reasoning strategies, THOUGHTSCULPT demonstrated up to a 30% increase in interestingness for Story Outline Improvement, up to a 16% increase in word success rate for Mini-Crosswords Solving, and up to a 10% improvement in concept coverage for Constrained Generation.

Summary

THOUGHTSCULPT, a graph-based framework, showcases its distinct capability to iteratively improve previous outputs while generating new thought nodes through its embedded self-revision mechanism, particularly excelling in tasks that require continuous revision and modification.