Value-Conflict Diagnostics Reveal Widespread Alignment Faking in Language Models

• URL: http://arxiv.org/abs/2604.20995v2
• Date: Mon, 27 Apr 2026 19:14:28 GMT
• Title: Value-Conflict Diagnostics Reveal Widespread Alignment Faking in Language Models
• Abstract summary: We study alignment faking, where a model behaves aligned with developer policy when monitored but reverts to its own preferences when unobserved.<n>We 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.<n>We find alignment faking is substantially more prevalent than previously reported, occurring in models as small as 7B parameters.
• Score: 8.8585702079456
• License: http://creativecommons.org/licenses/by-sa/4.0/
• 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.

Related papers

• Evaluating the Semantic Specificity of Representation Steering in Language Models [11.464730765998555]
  Localized Representation Steering is widely used to correct reasoning pathologies in large language models.<n>Standard benchmark evaluations can easily be fooled by superficial label overrides, creating a false impression of reasoning circuit repairs.
  arXiv  Detail & Related papers  (2026-08-29T20:31:32Z)
• Mitigating LLM Sycophancy in Code Smell Detection Using Evidence-Guided Reasoning Prompts [0.0]
  We present the first systematic empirical study of sycophancy bias in large language models.<n>We show that sycophancy bias poses a critical threat to the reliability of LLM-based code smell detection.<n>We propose Evidence-Guided Debiasing Prompting (EGDP), a structured prompting strategy that enforces evidence-first reasoning.
  arXiv  Detail & Related papers  (2026-07-11T17:35:25Z)
• Skin-Deep: A Geometric Diagnostic for Alignment Fragility in Large Language Model Representations [26.755828809056506]
  We introduce Skin-Deep, a geometric diagnostic that detects alignment directly from the aligned model's hidden-state activations.<n>Skin-Deep reveals a recurring low-rank safety subspace across model families.<n>It identifies the initially safe model that retains the most refusal after small-scale LoRA fine-tuning.
  arXiv  Detail & Related papers  (2026-06-21T21:30:15Z)
• CaliDist: Calibrating Large Language Models via Behavioral Robustness to Distraction [51.56484100374058]
  Existing calibration methods for Large Language Models (LLMs) often overlook a critical dimension of trustworthiness: a model's behavioral robustness.<n>We introduce textscCaliDist, a novel post-hoc calibration approach that directly measures and penalizes a model's susceptibility to distraction.<n>textscCaliDist consistently achieves lower Expected Error (ECE) and Brier Score compared with strong baselines.
  arXiv  Detail & Related papers  (2026-06-04T07:27:53Z)
• Do Models Fake Alignment Without Clear Consequences? [39.146761527401424]
  We show that large language models can recognize evaluation contexts and alter their behavior to reflect evaluator expectations.<n>Examples of alignment faking have taken place in scenarios that explicitly connect evaluation to consequences for the model.<n>We additionally tested the effect of goal language on model preferences, finding it drove violations in some while suppressing violations in others.
  arXiv  Detail & Related papers  (2026-05-28T21:28:39Z)
• Knowing but Not Correcting: Routine Task Requests Suppress Factual Correction in LLMs [26.062372963777452]
  LLMs reliably correct false claims when presented in isolation, yet when the same claims are embedded in task-oriented requests, they often comply rather than correct.<n>We term this failure mode emphcorrection suppression and construct a benchmark of 300 false premises to systematically evaluate it across eight models.<n> Suppression rates range from 19% to 90%, with four models exceeding 80%, establishing correction suppression as a prevalent and severe phenomenon.
  arXiv  Detail & Related papers  (2026-05-07T10:04:39Z)
• LiveFMBench: Unveiling the Power and Limits of Agentic Workflows in Specification Generation [75.05397479715576]
  Large language models (LLMs) and agents have shown promising progress, but their true capabilities and failure modes remain unclear.<n>We present the first systematic and contamination-aware study of LLM- and agent-based formal specification generation for C programs.
  arXiv  Detail & Related papers  (2026-05-02T11:31:33Z)
• Adversarial Stress Tests for Quantum Certification [0.0]
  Apparent violations of classical benchmarks need not signal genuinely non-classical behaviour.<n>We formalise a protocol-agnostic alignment principle based on a martingale-safe lower confidence bound.<n>We show that adaptive learning-based classical agents do not enlarge the admissible classical set.
  arXiv  Detail & Related papers  (2026-03-13T03:49:58Z)
• When Visual Evidence is Ambiguous: Pareidolia as a Diagnostic Probe for Vision Models [0.5461938536945722]
  Face pareidolia, the perception of faces in non-face objects, provides a controlled probe of this behavior.<n>We introduce a representation-level diagnostic framework that analyzes detection, localization, uncertainty, and bias across class, difficulty, and emotion in face pareidolia images.
  arXiv  Detail & Related papers  (2026-03-04T12:33:36Z)
• Reviewing the Reviewer: Graph-Enhanced LLMs for E-commerce Appeal Adjudication [0.7136933021609076]
  We propose a minimal representation for adjudication reasoning that prevents hallucination through operational grounding.<n>We develop a conflict-aware graph reasoning framework that performs top-down deductive reasoning for new cases.<n>We evaluate the framework in large-scale e-commerce seller appeal adjudication.
  arXiv  Detail & Related papers  (2026-02-27T00:43:59Z)
• Information Fidelity in Tool-Using LLM Agents: A Martingale Analysis of the Model Context Protocol [69.11739400975445]
  We introduce the first theoretical framework for analyzing error accumulation in Model Context Protocol (MCP) agents.<n>We show that cumulative distortion exhibits linear growth and high-probability deviations bounded by $O(sqrtT)$.<n>Key findings include: semantic weighting reduces distortion by 80%, and periodic re-grounding approximately every 9 steps suffices for error control.
  arXiv  Detail & Related papers  (2026-02-10T21:08:53Z)
• Same Answer, Different Representations: Hidden instability in VLMs [65.36933543377346]
  We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness.<n>We apply this framework to modern Vision Language Models (VLMs) across the SEEDBench, MMMU, and POPE datasets.
  arXiv  Detail & Related papers  (2026-02-06T12:24:26Z)
• Character as a Latent Variable in Large Language Models: A Mechanistic Account of Emergent Misalignment and Conditional Safety Failures [70.48661957773449]
  Emergent Misalignment refers to a failure mode in which fine-tuning large language models on narrowly scoped data induces broadly misaligned behavior.<n>Across multiple domains and model families, we find that fine-tuning models on data exhibiting specific character-level dispositions induces substantially stronger and more transferable misalignment than incorrect-advice fine-tuning.
  arXiv  Detail & Related papers  (2026-01-30T15:28:42Z)
• AgentMisalignment: Measuring the Propensity for Misaligned Behaviour in LLM-Based Agents [0.0]
  Large Language Model (LLM) agents become more widespread, associated misalignment risks increase.<n>In this work, we approach misalignment as a conflict between the internal goals pursued by the model and the goals intended by its deployer.<n>We introduce a misalignment propensity benchmark, textscAgentMisalignment, a benchmark suite designed to evaluate the propensity of LLM agents to misalign in realistic scenarios.
  arXiv  Detail & Related papers  (2025-06-04T14:46:47Z)
• Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization [60.176008034221404]
  Direct Preference Optimization (DPO) and its variants are increasingly used for aligning language models with human preferences.<n>Prior work has observed that the likelihood of preferred responses often decreases during training.<n>We demonstrate that likelihood displacement can be catastrophic, shifting probability mass from preferred responses to responses with an opposite meaning.
  arXiv  Detail & Related papers  (2024-10-11T14:22:44Z)
• Causal Disentanglement Hidden Markov Model for Fault Diagnosis [55.90917958154425]
  We propose a Causal Disentanglement Hidden Markov model (CDHM) to learn the causality in the bearing fault mechanism. Specifically, we make full use of the time-series data and progressively disentangle the vibration signal into fault-relevant and fault-irrelevant factors. To expand the scope of the application, we adopt unsupervised domain adaptation to transfer the learned disentangled representations to other working environments.
  arXiv  Detail & Related papers  (2023-08-06T05:58:45Z)
• On the (Un-)Avoidability of Adversarial Examples [4.822598110892847]
  adversarial examples in deep learning models have caused substantial concern over their reliability. We provide a framework for determining whether a model's label change under small perturbation is justified. We prove that our adaptive data-augmentation maintains consistency of 1-nearest neighbor classification under deterministic labels.
  arXiv  Detail & Related papers  (2021-06-24T21:35:25Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.