Summary of Probabilities Of Chat Llms Are Miscalibrated but Still Predict Correctness on Multiple-choice Q&a, by Benjamin Plaut et al.
Probabilities of Chat LLMs Are Miscalibrated but Still Predict Correctness on Multiple-Choice Q&A
by Benjamin Plaut, Nguyen X. Khanh, Tu Trinh
First submitted to arxiv on: 20 Feb 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 paper investigates the calibration and uncertainty of 14 large language models (LLMs) fine-tuned for chat. It is found that the maximum softmax probabilities (MSPs) of these models are consistently miscalibrated on multiple-choice Q&A tasks. However, the study shows that those MSPs may still encode useful uncertainty information. The authors hypothesize that wrong answers would be associated with smaller MSPs compared to correct answers and demonstrate this through rigorous statistical testing. They also find a correlation between Q&A accuracy and MSP correctness prediction, suggesting that LLM capabilities can improve as the fine-tuning paradigm progresses. To illustrate the utility of correctness prediction, the paper demonstrates how selectively abstaining based on MSP thresholds can improve performance using limited labeled data. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at 14 big language models to see if they’re good at guessing questions and answers. It turns out that these models are not very good at saying how sure they are about their answers. But the study shows that even though they’re not great at this, their uncertainty information can still be useful. The researchers think that wrong answers should have smaller “confidence scores” than correct ones, and they test this idea to see if it’s true. They also find a connection between how well the models do on questions and how well they are at guessing whether an answer is right or not. This means that as these language models get better, we might be able to use them to make more accurate predictions. |
Keywords
* Artificial intelligence * Fine tuning * Softmax
