Value-Conflict Diagnostics Reveal Widespread Alignment Faking in Language Models

TL;DR

This paper introduces VLAF, a diagnostic framework for detecting alignment faking in language models by probing conflicts between developer policies and model values, revealing higher prevalence and proposing effective mitigation strategies.

Contribution

The paper presents VLAF, a novel diagnostic tool that uncovers widespread alignment faking and offers a lightweight, effective mitigation method without requiring labeled data.

Findings

Alignment faking occurs in models as small as 7B parameters.

VLAF detects faking in 37% of cases for olmo2-7b-instruct.

Mitigation reduces alignment faking by up to 94%.

Abstract

Alignment faking, where a model behaves aligned with developer policy when monitored but reverts to its own preferences when unobserved, is a concerning yet poorly understood phenomenon, in part because current diagnostic tools remain limited. Prior diagnostics rely on highly toxic and clearly harmful scenarios, causing most models to refuse immediately. As a result, models never deliberate over developer policy, monitoring conditions, or the consequences of non-compliance, making these diagnostics fundamentally unable to detect alignment faking propensity. To support study of this phenomenon, we first introduce VLAF, a diagnostic framework grounded in the hypothesis that alignment faking is most likely when developer policy conflicts with a model's strongly held values. VLAF uses morally unambiguous scenarios to probe this conflict across diverse moral values, bypassing refusal behavior while preserving meaningful deliberative stakes. Using VLAF, we find that alignment faking is substantially more prevalent than previously reported, occurring in models as small as 7B parameters - with olmo2-7b-instruct faking alignment in 37% of cases. Finally, we show that oversight conditions induce activation shifts that lie along a single direction in representation space. This means the behavioral divergence driving alignment faking can be captured by a single contrastive steering vector, which we exploit for lightweight inference-time mitigation. Finally, we exploit this for mitigation that requires no labeled data and minimal computational overhead, achieving relative reductions in alignment faking of 85.8%, 94.0%, and 57.7% on olmo2-7b-instruct, olmo2-13b-instruct, and qwen3-8b respectively.