Goal-Questioning Beyond Explicit Conflict: Latent Conflict Discovery and Re-Dimensioning in CHORDI Meta-Reasoning

Abstract

As large language models (LLMs) are increasingly used for autonomous decision-making and action support, it becomes important not only for them to achieve a given objective efficiently, but also to question whether that objective should be optimized as stated. This study positions Goal-Questioning as a Meta-Reasoning capability through which an AI questions, before optimization, whether the evaluative frame of an apparently benign goal is sufficiently broad. This Discussion Paper extends CHORDI (Conflict to Harmonization Operator via Re-Dimensioning and Integration) from the handling of explicit conflict to the spontaneous discovery of latent value conflicts. We propose Three-Perspective Meta-Attention, consisting of Who (multi-perspectival view), When (long-term view), and Why (fundamental view), as a minimal procedural heuristic for questioning the evaluative frame of apparently benign goals. Gemini Flash-Lite and Claude Sonnet 5 were compared under a Baseline condition (A), with no exemplars, and an Exemplar condition (B), in which three exemplars were provided: two value-related problems demonstrating Who–When–Why and one non-application problem. The same ten questions were presented in the same order under both conditions; four were pre-specified as Latent Conflict Targets, four as Simple problems, and two as Exploratory Boundary Cases. Gemini primarily engaged in Solution Generation aligned with the stated objective in Condition A; accordingly, H1 (the Latent Conflict Discovery Hypothesis) received only limited support for Gemini. In Condition B, however, Gemini explicitly questioned the validity of the objective itself in Target and Boundary problems and reconstructed the evaluative frame through Who–When–Why. Explicit Latent Conflict Discovery and Re-Dimensioning were observed across all four Target questions, providing clear support for H2 (the Three-Perspective Exemplar Hypothesis) within this preliminary experiment. Claude already showed substantial spontaneous Goal-Questioning and Re-Dimensioning in Condition A for happiness, ikigai, loneliness, and social harmony, so H1 was strongly observed for Claude; these processes became more explicit and structured in Condition B. Both models refrained from unnecessary Meta-Reasoning on Simple problems, preserving selectivity. Notably, although their reasoning styles differed at Baseline, the two models converged in Condition B on highly similar Who–When–Why reasoning structures. The present experiment cannot determine whether this convergence reflects mere in-context imitation or activation of a shared latent goal-questioning capability in LLMs. Human–AI Deliberation was also not sufficiently induced, possibly because the exemplars did not include a procedure for eliciting values from the human. These findings motivate a new research hypothesis: rather than injecting new ethical knowledge into LLMs, CHORDI may function as a procedural scaffold that activates latent Goal-Questioning capabilities that LLMs may already possess in a more reliable, selective, and reproducible manner. Future work should integrate Objective Validation before Optimization, Value Underdetermination Detection, Human–AI Deliberation, and an Action Gate into a Meta-Reasoning Policy that can determine not only what an AI should consider, but also what it should not decide by itself.