Why does chain-of-thought reasoning fail for personalization?

PRIME documents a two-layer failure in applying reasoning to personalization:

Layer 1: Generic CoT fails. Enabling standard chain-of-thought often underperforms the non-thinking baseline for personalization tasks. The uncustomized reasoning trace "merely scratches the surface, seeking broad answers rather than to-the-point, user-specific responses." Generic reasoning explores the problem space without being anchored to the specific user's preferences, values, or communication style — producing reasoning that is logically sound but personally irrelevant.

Layer 2: Fine-tuning destroys thinking capacity. The "fast thinking" training paradigm (direct input→output mapping) turns fine-tuned LLMs into specialist models overfitted to the target space. They lose the generalist capability of generating meaningful intermediate thoughts when prompted. A common error is token repetition — the model has been trained to shortcut directly to outputs and can no longer produce coherent intermediate reasoning. This is not a minor degradation — the model structurally cannot think anymore.

The fix: personalized self-distillation. The model generates its own personalized thinking traces (using its pre-fine-tuning generalist capability), then trains on those traces alongside the standard fine-tuning objective. This produces reasoning that is both user-specific (anchored to the individual's preferences) and deep (maintaining the capacity for intermediate thought). The self-distillation approach leverages the model's own capabilities rather than requiring external reasoning trace data.

This finding extends the reasoning/judgment split documented elsewhere. Since When does explicit reasoning actually help model performance?, personalization is a clear case of "continuous nuanced judgment" — matching preferences, style, and implicit expectations cannot be reduced to logical derivation steps. But PRIME shows the split is not absolute: personalized reasoning can help, provided the reasoning traces themselves are customized to the user.

The connection to Why does asking models to think first hurt performance? is structural: both findings demonstrate that thinking initially hurts but becomes helpful after the thinking process is adapted to the domain. In PRIME's case, self-distillation is the adaptation mechanism; in the TPO case, RL training is. The shared principle: raw thinking capability must be tuned to the domain before it adds value.