HorizonBench: Long-Horizon Personalization with Evolving Preferences

Long-horizon personalization requires tracking how user preferences evolve through life events, yet existing benchmarks conflate two distinct failure modes, retrieval of a stated preference and belief update when a life event renders it obsolete, making it impossible to identify which is the bottleneck. We introduce HorizonBench, a generation framework and benchmark that isolates belief update under controlled preference evolution. A mental state graph with typed dependency edges maintains preferences and their provenance across 6-month, ~130K-token timelines; an expression-tracking system separates when preferences change from when they are discussed; and pre-evolution values serve as hard-negative distractors that require update rather than retrieval. Evaluating 25 frontier models on 4,245 items from 350 users, the best model achieves only 53% on this 5-option task and most score at or below the 20% chance baseline. When models err on evolved preferences, they select the pre-evolution distractor on over a third of wrong answers --- far exceeding the 25% uniform-error baseline (p < 0.001, all 25 models) --- confirming that models retrieve originally stated preferences rather than integrating the life events that changed them. This failure is independent of context length and expression mode, establishing that the bottleneck for long-horizon personalization is not finding preferences but updating them.