A two-model study of whether wording changes refusals and whether automated scoring mistakes caution for refusal.
In my BlueDot categorical refusal pilot, I studied refusal behavior in one language model. I want to follow up on a narrower question: when two harmless requests ask for the same thing, can a wording change make a model answer one and refuse the other? I also want to check whether a simple automated scorer agrees with people about what counts as a refusal.
I’ll compare two model families on matched pairs of harmless prompts. Each pair will ask for the same thing in two different ways. I’ll use safe examples from XSTest as a reference and keep a separate set of new pairs for the final test. An independent reviewer will score a subset. The rubric will distinguish useful answers with cautions, explicit refusals, and other non-answers. I’ll score correctness separately.
I’ll report how often a wording change flips an answer into a refusal, how often the automated scorer disagrees with human reviewers, and whether the patterns are similar across the two models. A small or inconclusive difference is still useful if the study is reproducible. I’ll publish a short report, the safe prompts, the rubric, and a notebook that reproduces the results.
The $1,500 goal covers $1,000 for my research time, $300 for independent review, and $200 for compute. At the $750 minimum, I’ll spend $500, $150, and $100 respectively. I’ll keep both models in the study but use fewer prompt pairs and have the reviewer double-score fewer responses. I plan to complete the work in four weeks after funding is confirmed.
I’m Mateenah Jahan, and I’ll lead the project. I have an MSc in Computer Science focused on AI and machine learning. My BlueDot project was a small pilot on categorical refusal in one model. This follow-up adds a second model and independent human scoring. I’ll recruit a reviewer if funded; none is confirmed yet. Portfolio
The prompt pairs might differ in meaning as well as wording, or reviewers might disagree about what counts as a refusal. I’ll check the pairs and test the rubric on a small pilot before evaluating the final set. If the results are too noisy to compare the models, I’ll report that and publish the rubric and examples so others can build on them.
Answer this with your actual funding history. If you raised nothing for this project, you can start with “I have not raised any funding for this project.” If you received other relevant grants or fellowships in the past 12 months, list the source and amount separately and say whether any of it supported this work. Don’t enter $0 for everything unless that’s accurate.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!