Contrastive Attribution in the Wild: An Interpretability Analysis of LLM Failures on Realistic Benchmarks

Attribution tools work inconsistently on real-world tasks: The study found that contrastive attribution—a technique for identifying which parts of an input caused an LLM to make a wrong choice—"can yield informative signals in some failure cases, but is not universally applicable," meaning practitioners cannot rely on these interpretability tools as a general solution for understanding why models fail on realistic benchmarks.

Interpretability research has a gap between theory and practice: Prior work on LLM interpretability has focused mainly on "short prompts or toy settings," but this paper reveals that attribution patterns may behave differently on the complex, longer inputs found in commonly-used benchmarks, suggesting that conclusions from simplified experiments may not transfer to real-world applications.

Model size and training stage affect how interpretable failures are: The systematic comparison "across datasets, model sizes, and training checkpoints" revealed varying attribution patterns, indicating that the same interpretability tool may work better for some models than others—an important consideration when deploying these diagnostic techniques in practice.

We need better tools for analyzing failures in realistic contexts: The paper's introduction of "an efficient extension that enables construction of cross-layer attribution graphs for long-context inputs" addresses a practical bottleneck, but the overall findings suggest the field still lacks fully reliable methods for understanding why LLMs fail on real-world tasks that matter most.