Adversarial Concept Search Predicts LLM Failures from Representation Geometry
A new paper introduces Adversarial Concept Search (ACS), which uses LLM representation geometry to predict which concept combinations a model will fail on without running evaluations; the authors found the method works better on stronger models across all scales examined.
2026-11-07 ~ 2026-11-07 · 3 related posts
- Paper: Predicting LLM Compositional Failures from Feature Geometry, No Eval Inputs Needed — nsaphra · 2026-11-07
- Paper Author: The Stronger the Model, the Better Geometry-Based Failure Prediction Works — nsaphra · 2026-11-07
- Kempner Institute's Adversarial Concept Search predicts LLM failure modes from representational geometry — nsaphra · 2026-11-07