Summary of Nudging: Inference-time Alignment Via Model Collaboration, by Yu Fei et al.
Nudging: Inference-time Alignment via Model Collaboration
by Yu Fei, Yasaman Razeghi, Sameer Singh
First submitted to arxiv on: 11 Oct 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); 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 Large language models (LLMs) require alignment to effectively and safely follow user instructions. This process necessitates training aligned versions for every model size in each model family, resulting in significant computational overhead. Our proposed nudging algorithm aligns any base model at inference time using a small aligned model, eliminating the need for additional training. We find that nudging tokens steer the large base model’s output toward desired directions when its uncertainty is high. Evaluating nudging across 3 model families and 13 tasks, we achieve zero-shot performance comparable to or surpassing that of large aligned models. For example, nudging OLMo-7b with OLMo-1b-instruct achieves a 10% absolute improvement on GSM8K. Our approach enables off-the-shelf collaboration between model families, such as nudging Gemma-2-27b with Llama-2-7b-chat, outperforming Llama-2-70b-chat on various tasks. Nudging offers a simple yet powerful approach to token-level model collaboration, providing a modular solution to LLM alignment. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models need help following instructions. To do this, they must be “aligned” with human feedback. This takes a lot of computer power and training data. Our new method, called nudging, makes it easier by using a small aligned model to guide the big base model when it’s unsure. We tested nudging on many different models and tasks, and it worked just as well as more complex methods that required extra training. |
Keywords
* Artificial intelligence * Alignment * Inference * Llama * Token * Zero shot
