Language-model security
The evaluator changes the experiment.
I extended an existing red-teaming workflow and explored how the search and evaluator affected its results.
My role
I ran experiments, proposed a multi-evaluator extension, and presented the work in a UW–Madison course project.
Where it stands
Course research using a fork of TAP/PAIR. Results were weak or inconclusive; this is not a measured safety-breakthrough claim.
What changed.
Exploratory runs and a proposal for evaluating different aspects of an attack separately.
What I changed
I varied the search depth, width and branching in an existing Tree-of-Attacks-with-Pruning workflow and examined a multi-evaluator design. The framework was upstream work; my contribution was the experimentation and proposed extension.
What the result taught me
Evaluator capability and cost constrained the experiment. I keep the weak results visible because they change what the work supports: practical experience with model evaluation, rather than a universal jailbreak success rate.
The stack
- Python
- TAP / PAIR
- Tree search
- GPT APIs
- Weights & Biases