Type: url
Original: https://arxiv.org/pdf/2601.04556
Status: generated
Summary
A preprint from two Tencent authors proposing 4D-ARE (4-Dimensional Attribution-Driven Agent Requirements Engineering), a design-time methodology for specifying what an LLM decision-support agent should reason about — as distinct from runtime reasoning frameworks such as ReAct and Chain-of-Thought, which govern how it reasons. Its premise is that a decision-maker who asks about an outcome is seeking attribution rather than an answer, and that attribution concerns organize into four dimensions (Results, Process, Support, Long-term) forming a causal chain traced backwards from an observed gap; the methodology operationalizes this through five specification layers whose YAML artifacts compile into system prompts. It is demonstrated through a single industrial pilot at a commercial bank, and the authors are explicit throughout that the four-dimensional structure is motivated by theory rather than derived from it, and that their deployment observations are preliminary hypotheses rather than validated findings.