从「评估」到「照见自己」:一份偏好信号的四种产出

Crowdsourced pairwise "arena" evaluation has become the standard for ranking generative models, yet it treats the human evaluator as a mere labeling instrument and discards everything but the aggregate ranking. We present Aesthetic Arena, a deployed system in which a single stream of pairwise aesthetic judgments over AI-generated art is used simultaneously to (1) rank individual images and—novel to our setting—the generation methods (workflows/models) behind them, closing the loop back to production; (2) estimate each evaluator's discernment and weight their votes accordingly; and (3) infer a personal aesthetic profile that we reflect back to the evaluator as a "taste report," turning evaluation labor into an act of self-discovery that sustains participation. We aggregate votes with a shrinkage-regularized Bradley-Terry model reported with bootstrap confidence intervals and lower-bound ranking, decompose preference along interpretable axes (overall / "de-AI-ness" / creativity), and stratify pairings by subject to separate quality from subject preference. We detail the system and the bridge from collective evaluation to individual personalization, and report a deployment study on participation, the noise-robustness of discernment-weighting, and whether the self-discovery framing increases sustained engagement.