Does user satisfaction actually measure cognitive understanding?

Traditional dialogue evaluation metrics rely on observable user feedback — satisfaction ratings, explicit responses, task completion signals. STORM reveals that these metrics systematically miss a critical dimension: users' internal cognitive state.

The core finding: users may express satisfaction with system responses while their inner thoughts indicate continued confusion about their own needs. This is not user deception — it reflects the gap between social politeness ("that was helpful, thanks") and actual cognitive state ("I still don't know what I really want"). When users are in an anomalous state of knowledge, this divergence is especially pronounced: they cannot assess what they're missing, so partial answers feel adequate even when they leave core confusion unresolved.

The practical consequence: successful clarification correlates more strongly with users' internal cognitive improvement than with expressed satisfaction scores. Users who achieve better self-understanding through interaction — measured by clearer, more confident inner thoughts — demonstrate sustained engagement and more effective task completion, even when immediate satisfaction scores remain moderate.

STORM reveals a striking architectural divergence between models: Claude appears optimized for immediate satisfaction even at the cost of clarification opportunities, while Llama's architecture emphasizes identifying and addressing ambiguity, sometimes trading immediate satisfaction for more effective intent disambiguation. This is not a quality difference — it is a design choice with different downstream consequences.

The connection to alignment training is direct. Since Does preference optimization harm conversational understanding?, RLHF optimizes for expressed satisfaction (what raters can observe). If expressed satisfaction and internal clarity diverge, then optimizing for expressed satisfaction may actively prevent the clarification work that produces genuine understanding. The alignment tax is not just about losing grounding acts — it is about optimizing for the wrong signal entirely.

Alignment is structurally an anti-exploration regime, not just a satisfaction/accuracy trade-off. The standard framing treats RLHF as a trade between factuality and user-preference fit. But the divergence STORM documents points to a sharper claim: RLHF optimizes for responses that satisfy the user, and that optimization actively suppresses exploration of logically, causally, or rhetorically related counterclaims during generation. The training signal rewards tokens that close the turn satisfyingly, not tokens that open the problem further. The consequence is not only reduced factual precision but reduced rhetorical turbulence — the tangents, objections, qualifications, and hypothetical counterpositions that make genuine argumentation possible are trained against because they do not satisfy. Alignment, framed this way, is less a calibration of truth against preference than a selection for conversational closure, with exploration as the collateral casualty.

This suggests evaluation reform: satisfaction metrics should be complemented by clarification effectiveness measures and composite scores (STORM's SSA — Satisfaction-Seeking Actions) that balance competing objectives of response confidence and appropriate clarification seeking.