LLM-based agent systems can fail for reasons that are hard to pin down, because a mistake made in an early step can ripple through subsequent interactions. The new paper introduces ReCast, a method for attribution-oriented step representation learning. Rather than focusing only on the final failure, ReCast learns representations of individual steps that make it possible to trace the failure back to its origin.
According to the abstract, the problem is that failures can originate from early steps whose effects propagate, making their origins difficult to identify. ReCast is designed to address this by learning step representations that are useful for attribution. The paper is available on arXiv under identifier 2610.11334.