Summary
AI systems that use large language models often fail when facing new situations and can't explain why they make certain choices. This paper explains that these problems come from how the data used to train and test the AI is collected and handled. The authors suggest a new method called Data-Centric Anchoring, which focuses on carefully managing and improving the data environment before training the AI. Their approach uses a four-step cycle—curate, augment, constrain, and attribute—that helps the AI learn from better data and make more reliable and understandable decisions. This method aims to fix common AI failures by improving the data itself rather than just tweaking the AI models.
agentic AIlarge language modelsdistribution shiftcounterfactual reasoningdata lifecyclerobustnessinterpretabilitydata augmentationinvarianceexplanation faithfulness
Abstract
Agentic AI systems built on large language models fail in two persistent ways that scaling does not fix: they break under distribution shift, and they cannot explain the decisions they make. We argue these are co-symptoms of one structural deficiency in the data lifecycle that governs how agents are trained, evaluated, and deployed. Observational interaction logs record what an agent did, not what it would have done otherwise. They encode spurious correlations without controlled variation, so they lack the counterfactual structure needed to separate causal signal from coincidence or to validate an explanation. No model-centric method can recover invariances the data never contained. We present Data-Centric Anchoring: robustness and interpretability should be engineered into the data environment, not extracted from models after training. Our central contribution is the Data-Centric Agentic Loop, a four-stage framework of Curate, Augment, Constrain, and Attribute. The ordering is structural, not stylistic. Curation precedes augmentation because generative models amplify whatever bias they are trained on. Augmentation precedes constraint because invariance objectives are vacuous without variation across environments to be invariant to. Attribution closes the loop, converting observed failures into targeted data interventions for the next iteration. Each stage manufactures the preconditions of the next, which makes the loop self-correcting rather than merely sequential. We ground the framework in a failure-driven taxonomy that links four core failure modes to the data lifecycle: spurious feature reliance, distribution-shift fragility, uncertainty miscalibration, and explanation unfaithfulness. We close with the limits of this approach and the open problems that stand between it and practical deployment at scale.