Summary of Learning From Committee: Reasoning Distillation From a Mixture Of Teachers with Peer-review, by Zhuochun Li et al.
Learning from Committee: Reasoning Distillation from a Mixture of Teachers with Peer-Review
by Zhuochun Li, Yuelyu Ji, Rui Meng, Daqing He
First submitted to arxiv on: 4 Oct 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 Medium Difficulty summary: While advancements in large language models (LLMs) have been impressive, researchers are now exploring ways to improve smaller open-source models through knowledge distillation (KD) from commercial LLMs. However, most studies rely on a single LLM’s responses as the gold standard, whereas human learning involves understanding both correct answers and incorrect reasoning. To address this limitation, we introduce Fault-Aware DistIllation via Peer-Review (FAIR), which asks teachers to identify and explain student mistakes, providing customized instructional data. Our method also employs a simulated peer-review process between teacher LLMs, selecting only rationales above an acceptance threshold to improve data quality. We demonstrate the effectiveness of FAIR on mathematical, commonsense, and logical reasoning tasks through comprehensive experiments and analysis. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Low Difficulty summary: This paper is about improving smaller language models by learning from bigger ones. Right now, most studies use a single model as the perfect answer, but humans learn from mistakes too! To fix this, we created a new way to train models called Fault-Aware Distillation via Peer-Review (FAIR). It asks teachers to explain why students got something wrong, which helps create better instructional data. We also simulate how teacher models review each other’s work, choosing only the best explanations. This new approach works well for math, common sense, and logical thinking problems. |
Keywords
» Artificial intelligence » Distillation » Knowledge distillation
