Pseudo-Deliberation in Language Models: When Reasoning Fails to Align Values and Actions

2026-05-10

AI Summary

This work addresses the systematic inconsistency between value expression and actual behavior in large language models, a gap that persists even when explicit reasoning is employed and often manifests as “pseudo-deliberation”—superficially coherent reasoning that fails to align actions with stated values. The paper introduces the concept of pseudo-deliberation for the first time and proposes VALDI, a comprehensive evaluation framework encompassing 4,941 human-centric scenarios, three distinct tasks, and five alignment metrics. Furthermore, it designs VIVALDI, a multi-agent intervention mechanism, and demonstrates its efficacy across both open- and closed-source models. Experimental results show that VIVALDI significantly narrows the value-action gap and enhances models’ value alignment capabilities.

Abstract

Large language models (LLMs) are often evaluated based on their stated values, yet these do not reliably translate into their actions, a discrepancy termed "value-action gap." In this work, we argue that this gap persists even under explicit reasoning, revealing a deeper failure mode we call "Pseudo-Deliberation": the appearance of principled reasoning without corresponding behavioral alignment. To study this systematically, we introduce VALDI, a framework for measuring alignment between stated values and generated dialogue. VALDI includes 4,941 human-centered scenarios across five domains, three tasks that elicit value articulation, reasoning, and action, and five metrics for quantifying value adherence. Across both proprietary and open-source LLMs, we observe consistent misalignment between expressed values and downstream dialogues. To investigate intervention strategies, we propose VIVALDI, a multi-agent value auditor that intervenes at different stages of generation.

Problem

• value-action gap
• Pseudo-Deliberation
• alignment
• language models
• reasoning

Innovation

• Pseudo-Deliberation
• value-action gap
• VALDI
• VIVALDI
• alignment evaluation