Large language model (LLM) agents are typically optimized to achieve given objectives, but real-world objectives are often incomplete, ambiguous, or misaligned with underlying needs. This paper proposes a structured architecture for meta-objective reframing — the process of critically examining and reorganizing objectives themselves rather than simply pursuing them. Building on the RDAC (Reflective Difference-Driven Autonomous Cognition) framework, we introduce a four-role pipeline (Goal Expander, Goal Critic, Goal Synthesizer, Goal Diff) that induces structural reframing of objectives within a bounded constraint space. We further propose a five-dimensional Diff schema (Object, Metric, Scope, Premise, Relation) to characterize the nature of objective transformations in a structured and comparable form. Experiments across four objective conditions — ranging from high-constraint single-objective to multi-objective composite conditions — reveal that: (1) Relation Shift is consistently detected in composite objective conditions across repeated trials, with diverse restructuring patterns including conditionalization, staging, and parallelization; (2) a minimum-change constraint on the Synthesizer produces graded reframing outputs that enable differential pattern detection; (3) the Diff schema demonstrates adequate primary-dimension separation on controlled inputs, though complex generated inputs introduce attribution instability between Object, Scope, and Relation. We discuss limitations including the non-orthogonality of the five dimensions, the absence of a true baseline condition, and the need for refined boundary rules between dimensions. This work positions meta-objective reframing not as free objective generation but as structured reorganization within constraint spaces, functioning as a structural cognitive scaffold for human decision-makers with implications for autonomous agent design and human-AI collaborative deliberation.
Masanao Ishikawa (Sun,) studied this question.