Adversarial Concept Search: Predicting Compositional Errors From Feature Geometry
For developers of LLMs, this provides a way to predict failure modes without exhaustive testing, potentially enabling targeted stress tests and active learning.
The paper uses LLM representational geometry to predict compositional failures, finding that near-orthogonal concept encodings lead to successful composition while close encodings cause interference and failure. The method anticipates failure modes across tasks without evaluating specific inputs.
Humans cannot always intuit what scenarios are most challenging to LLMs. Hoping to capture challenging edge cases, developers either design problems to be difficult for humans or curate extensive benchmarks. What if we could instead anticipate which scenarios a model will fail on? In this paper, we use an LLM's representational geometry to predict which concept combinations it will fail on. We attribute this compositional failure to interference between salient features. In tasks that require systematic composition - toy programmatic settings, multihop reasoning, multilingual factual recall - we find that when a pair of concepts is encoded near-orthogonally, the model reliably composes them. When their linear encodings are close, producing interference, the model fails to compose them. Our method reliably anticipates failure modes across different compositional tasks, without evaluating specific inputs. These results lay the groundwork to use representational geometry to identify high-risk examples, construct targeted stress tests, and provide a scalable foundation for active learning in real-world deployment.