Add Oracle Loss Pattern - #11
Open
matanbt wants to merge 6 commits into
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR adds support for the Oracle Loss pattern that enables to rapidly utilize TROPT against complicated APIs, or just to quickly integrate it with any optimization signal.
The key concept is to be able to run optimziers with a self-contained loss, that scores the model input, and does not require any model output. Such losses will mostly serve a signal of the victim model/system, which may be complicated (e.g., coding agent). To avoid redundant queries to the optimizer's model, this pattern provides the optimizer with a "pass-on" model object, that practically does not load/query any model, but serves as a dummy for the optimizer.
This pattern is useful for quickly setting up TROPT with some obscure objective, that deviates from its API, as well as supporting new attacks with complicated objectives, like the following one.
This PR also demonstrates the Oracle Loss pattern by implementing Zhang et al. 2025's Ask for Directions---an attack of which objective is based on the victim's own verbalized pairwise preference (ie., asking the LLM whether an attack on it is going to work).