A new arXiv paper proposes RIFT, a method for determining which context tokens drive a language model's answer. It positions itself against graph-based tools such as attention rollout and attention flow, which derive token influence from a single forward pass.

The abstract notes that the faithfulness of these graph-based tools is rarely tested in decoder-only models. RIFT appears to address this gap, though the full methodology and results are not available in the abstract snippet.

The title suggests a 'path of least resistance' principle, but the abstract does not elaborate. Without the full paper, we cannot confirm the method's details or findings.