Add PerspectiveGap to agent-specific evaluation - #68
Open
WhymustIhaveaname wants to merge 1 commit 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Resubmitting PerspectiveGap on its own, per your note on #43. One resource, §9, placed after the MAST entry it references.
It benchmarks the orchestration prompt itself. 110 scenarios across 10 topologies, each posed twice: assign information fragments to sub-agent roles, then write those sub-agents' prompts. Every scenario carries one planted distractor. Scoring is rule-only and deterministic, and the containment detector was validated against a 716-row hand-audited test set (99.44% agreement).
Why §9: the section already covers multi-agent evaluation through MAST, which classifies failures after a run. PerspectiveGap measures the step before it, whether a model can decide what each sub-agent needs to know. 33 commercial models average 17.2% combined pass, so there is a lot of headroom left.
Two things changed since #43. The magic-words entry is gone; an embedding-model attack isn't agent evaluation, and I should have caught that. The leakage figure now carries the abstract's parenthetical, so 217.9% reads as a per-scenario leak-event count rather than a percentage.
Disclosure: I am first author on this paper.
URLs verified. Dataset on HuggingFace (
sun1245/PerspectiveGap), code MIT.