Summary of System-2 Mathematical Reasoning Via Enriched Instruction Tuning, by Huanqia Cai et al.
System-2 Mathematical Reasoning via Enriched Instruction Tuning
by Huanqia Cai, Yijun Yang, Zhifeng Li
First submitted to arxiv on: 22 Dec 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: Computation and Language (cs.CL)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper introduces Enriched Instruction Tuning (EIT), a method to enhance large language models’ (LLMs) mathematical reasoning abilities. EIT synergizes human and AI feedback to create fine-grained reasoning trajectories for LLMs, addressing the scarcity of deliberate multi-step reasoning data. The method consists of two steps: Enriching with Reasoning Plan (ERP) and Enriching with Reasoning Step (ERS). ERP generates a high-level plan breaking down complex instructions into simpler objectives, while ERS fills in reasoning contexts often overlooked by human annotators. Unlike CoT prompting methods relying on LLM’s internal knowledge, EIT leverages human-annotated initial answers as “meta-knowledge” to generate more detailed and precise reasoning processes. Experimental results show EIT achieves state-of-the-art accuracy on GSM8K (84.1%) and MATH (32.5%), surpassing fine-tuning and prompting methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper helps computers solve complex math problems better by using a new way to train language models. Currently, these models struggle with multi-step reasoning, but the authors came up with a solution called Enriched Instruction Tuning (EIT). This method combines human and AI feedback to create more detailed reasons for the computer’s math answers. The process involves two steps: planning what to do first and filling in gaps where humans might not have written enough explanations. By using this approach, the language models can provide more accurate answers to complex math problems. |
Keywords
» Artificial intelligence » Fine tuning » Instruction tuning » Prompting
