Authors: Pei Zhou, Jay Pujara, Xiang Ren, Xinyun Chen, Heng-Tze Cheng, Quoc V. Le, Ed H. Chi, Denny Zhou, Swaroop Mishra, Huaixiu Steven Zheng

Published on: February 06, 2024

Impact Score: 8.22

Arxiv code: Arxiv:2402.03620

Summary

• What is new: Introduces SELF-DISCOVER, a new framework enabling LLMs to self-discover and compose reasoning structures for complex problem-solving, significantly outperforming existing methods like CoT on tough benchmarks.
• Why this is important: Current LLM prompting methods struggle with complex reasoning tasks.
• What the research proposes: A self-discovery framework that helps LLMs identify and combine atomic reasoning modules into explicit structures to improve problem-solving.
• Results: Substantial performance improvements on difficult reasoning benchmarks, up to 32% over CoT and 20% over CoT-Self-Consistency, with significantly less computation required.

Technical Details

Technological frameworks used: SELF-DISCOVER

Models used: GPT-4, PaLM 2, Llama2

Data used: BigBench-Hard, grounded agent reasoning, MATH

Potential Impact

Educational technology, AI development platforms, companies focusing on AI-based problem solving and decision-making solutions

Want to implement this idea in a business?

We have generated a startup concept here: CogniStruct.