Data-driven Circuit Discovery for Interpretability of Language Models

Researchers introduce Data-driven Circuit Discovery (DCD), a new framework for understanding language models that challenges the assumption that models implement tasks using a single computational circuit. By clustering data based on how models process examples, DCD discovers multiple task-specific circuits per dataset, revealing that existing methods conflate distinct mechanisms into single circuits and produce dataset-dependent rather than generalizable interpretations.

Circuit discovery research seeks to explain how language models solve tasks by identifying the neural subgraphs responsible for specific behaviors. Traditional hypothesis-driven approaches assume models implement each task through a single unified circuit and that datasets adequately capture task semantics. This work empirically invalidates both assumptions, demonstrating that minor dataset variations preserving semantic meaning produce circuits with minimal overlap, and that mixed datasets containing two separate tasks still yield single circuits with spuriously high faithfulness across both—indicating current methods extract dataset artifacts rather than fundamental task mechanisms.

This research emerges from the broader interpretability movement in AI, which has struggled with the challenge that neural networks often hide multiple solution strategies within black-box architectures. The finding that existing circuit discovery methods are fundamentally dataset-dependent rather than task-general represents a significant methodological setback, suggesting years of interpretability work may have produced misleading conclusions about how models actually compute. The proposed DCD framework sidesteps human task definition entirely, letting data clustering reveal the model's own organizational structure.

For the AI development community, these results underscore the fragility of mechanistic interpretability claims. Machine learning engineers and safety researchers relying on circuit-based explanations must now reconsider their confidence in model understanding. The work implies that truly robust interpretability requires discovering how models naturally partition problems, not imposing predetermined task boundaries. Moving forward, researchers should prioritize data-agnostic interpretability methods and validate findings across diverse datasets to avoid publishing misleading mechanistic explanations that only hold for specific training distributions.

• →DCD discovers multiple circuits per dataset by clustering examples based on model processing similarity, contrasting with existing single-circuit approaches
• →Minor dataset variations that preserve task semantics produce circuits with near-zero overlap, indicating existing methods capture dataset-specific patterns rather than true task mechanisms
• →Mixed datasets containing two tasks yield single circuits with high cross-task faithfulness, revealing that current discovery methods conflate distinct computational mechanisms
• →Data-driven discovery allows mechanistic structure to emerge from model behavior rather than being imposed through human task definitions
• →Interpretability researchers must validate findings across diverse datasets to distinguish between genuine task circuits and dataset-dependent artifacts