Summary of Adversarial Moment-matching Distillation Of Large Language Models, by Chen Jia
Adversarial Moment-Matching Distillation of Large Language Models
by Chen Jia
First submitted to arxiv on: 5 Jun 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Machine Learning (cs.LG)
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 This paper presents an innovative approach to knowledge distillation (KD) for large language models (LLMs). Traditional KD methods rely on minimizing explicit distribution distance between teacher and student probability predictions. In contrast, this work explores imitation learning strategies for KD of LLMs by matching the action-value moments of the teacher’s behavior from both on- and off-policy perspectives. The authors propose an adversarial training algorithm to jointly estimate the moment-matching distance and optimize the student policy to minimize it. Experimental results demonstrate the effectiveness of this approach, achieving new state-of-the-art performance in task-agnostic instruction-following experiments and task-specific tasks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about teaching a smaller language model (student) how to behave like a larger, more powerful language model (teacher). Instead of trying to make the student models predict the same things as the teacher, this paper tries something new. It matches what the teacher does with what the student does, kind of like imitating behavior. The authors tested their idea and found it works really well, even better than other methods that tried to make the student models predict the same things as the teacher. This could help us build more efficient language models in the future. |
Keywords
» Artificial intelligence » Knowledge distillation » Language model » Probability
