Summary of Revisiting the Superficial Alignment Hypothesis, by Mohit Raghavendra et al.
Revisiting the Superficial Alignment Hypothesis
by Mohit Raghavendra, Vaskar Nath, Sean Hendryx
First submitted to arxiv on: 27 Sep 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 The Superficial Alignment Hypothesis suggests that language models learn most of their abilities and knowledge during pre-training, with post-training focusing on style and format. Researchers empirically study the scaling behavior of post-training as finetuning examples increase, evaluating performance using standardized benchmarks. Experiments with Llama-3, Mistral, and Llama-2 model families show that post-training task performance scales as a power law against the number of finetuning examples across various capabilities like mathematical reasoning, coding, instruction following, and multihop-reasoning. The results highlight the need for holistic evaluation programs using objective benchmarks and demonstrate that language models can integrate new knowledge during post-training. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Language models learn most of their abilities and knowledge before they’re fine-tuned to do a specific job. Researchers wanted to know if this is still true as they give the model more examples to work with. They tested three different models, Llama-3, Mistral, and Llama-2, in various tasks like math problems, coding, following instructions, and answering questions that require several steps. The results show that the better the model is at doing these tasks, the more examples it gets. This means that the model isn’t just changing its style or format, but it’s actually learning new things. |
Keywords
» Artificial intelligence » Alignment » Llama
