Yuan Wu

Direct-decision models exhibit a learned "ordinal-scale-utilization bias," compressing the effective decision space and underutilizing available scale points, even when accuracy is high.

Proactive detection of flawed premises in recommendation requests is a major bottleneck for LLMs, with performance peaking at intermediate reasoning lengths.

MLLMs exhibit substantial reasoning deficits in text-rich image understanding, failing to apply visible rules and often providing incorrect answers despite compliant rationales.