A huge thank you to co-authors @fredhohman.bsky.social, @domoritz.de, @jeffreybigham.com, @kenholstein.bsky.social, and Mary Beth Kery! This work was done during my summer internship w/ Apple AIML, and I’m thankful to work with this wonderful team :)

arxiv.org/abs/2409.18203
#UIST25 talk: Wed 11am!

We can extend policy maps to enable Git-style collaboration and forking, aid live deliberation, and support longitudinal policy test suites & third-party audits. Policy maps can transform a nebulous space of model possibilities to an explicit specification of model behavior.

With our system, LLM safety experts rapidly discovered policy gaps and crafted new policies around problematic model behavior (e.g., incorrectly assuming genders; repeating hurtful names in summaries; blocking physical safety threats that a user needs to be able to monitor).

Given the unbounded space of LLM behaviors, developers need tools that concretize the subjective decisionmaking inherent to policy design. They should have a visual space to systematically explore, with explicit conceptual links between lofty principles and grounded examples.

Our system creates linked map layers of cases, concepts, & policies: so an AI developer can author a policy that blocks model responses involving violence, visually notice a gap of physical threats that a user ought to be aware of, and test a revised policy to address this gap.

We made updates to LLooM after the CHI publication to support local models (and non-OpenAI models)! More info here, though we haven't run evals across open-source models: stanfordhci.github.io/lloom/about/...

Qualitatively, I found that the BERTopic groupings were still rather large, so I anticipate the GPT labels would still be quite generic (as opposed to specific/targeted concepts).

That's a good point! In the technical evaluations, we used GPT to automatically find matches between the methods (including a GPT-only condition), but it could have evened the playing field even more to generate GPT-style labels for BERTopic before the matching step.

Thanks so much for sharing our work! :)