Summary of Listening to the Wise Few: Select-and-copy Attention Heads For Multiple-choice Qa, by Eduard Tulchinskii et al.
Listening to the Wise Few: Select-and-Copy Attention Heads for Multiple-Choice QA
by Eduard Tulchinskii, Laida Kushnareva, Kristian Kuznetsov, Anastasia Voznyuk, Andrei Andriiainen, Irina Piontkovskaya, Evgeny Burnaev, Serguei Barannikov
First submitted to arxiv on: 3 Oct 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 A novel evaluation approach for Large Language Models (LLMs) is proposed to address the limitations of traditional multiple-choice question answering formats. The authors introduce two new scores, the Query-Key Score (QK-score) and the Attention Score, which capture a model’s underlying knowledge more accurately than previous methods. These scores are extracted from specific attention heads in LLMs, demonstrating consistent performance across popular Multi-Choice Question Answering (MCQA) datasets. The proposed method improves knowledge extraction, achieving up to 16% gain for smaller models and up to 10% for larger models on MCQA benchmarks. Additionally, the accuracy on a simple synthetic dataset increases by almost 60%, indicating the method’s efficiency in mitigating format limitations. Experiments are conducted on models ranging from 7 billion to 70 billion parameters in both zero- and few-shot setups. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models are getting better at answering questions, but how do we know they’re really understanding what they’re saying? One way we test them is by giving them multiple-choice questions, but this format has some big limitations. Even if the model knows the right answer, it might struggle to choose the correct option because of the way the question is phrased. To fix this, researchers have come up with two new scores that show how well a language model really understands what it’s saying. These scores are based on how the model processes the words in a sentence and can tell us if it’s just guessing or actually knows the answer. By using these new scores, we can get a better sense of how well language models are doing at understanding natural language. |
Keywords
» Artificial intelligence » Attention » Few shot » Language model » Question answering
