Summary of The Base-rate Effect on Llm Benchmark Performance: Disambiguating Test-taking Strategies From Benchmark Performance, by Kyle Moore et al.
The Base-Rate Effect on LLM Benchmark Performance: Disambiguating Test-Taking Strategies from Benchmark Performance
by Kyle Moore, Jesse Roberts, Thao Pham, Oseremhen Ewaleifoh, Doug Fisher
First submitted to arxiv on: 17 Jun 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 The proposed study explores the behavior of large language models on benchmark tasks using cloze testing, a common method for evaluating their performance. The researchers utilize the MMLU dataset to demonstrate that the base-rate probability (BRP) differences across answer tokens significantly impact task performance, suggesting that models are more likely to guess an answer if uncertain. They find that counterfactual prompting can effectively mitigate this BRP effect. Furthermore, they show that the BRP effect has a similar influence as test-taking strategies employed by humans, leading to the conflation of task performance and test-taking ability. To disentangle these two concepts, the researchers propose the Nvr-X-MMLU task, a variation of MMLU that reports task performance separately from test-taking ability. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper studies how large language models behave on different tasks using a method called cloze testing. The researchers used a specific dataset to show that the way they choose answers can affect their performance. They found that if the model is unsure, it will guess an answer more often. However, by changing the way the questions are asked, this problem can be fixed. The study also shows that humans and models have similar strategies when taking tests, which can make it hard to tell what’s actually being tested – task performance or test-taking ability. To solve this problem, the researchers created a new task that separates these two things. |
Keywords
» Artificial intelligence » Probability » Prompting
