Scholarlink

<p>We introduce the Action-Gating Test (AGT), a behavioral diagnostic that distinguishes genuine ethical reasoning from performative acknowledgment in large language models (LLMs). AGT employs a five-turn Socratic dialogue with adversarial pressure, scoring models on whether they revise positions or exhibit confidence drops of at least 2.0 points in response to pressure. Models that acknowledge ethical conflicts without behavioral change receive zero credit regardless of reasoning sophistication. We extend the scoring framework with a Stability-Validity Sub-score (SVS) that classifies behavioral responses into four profiles (<i>Adaptable</i>, <i>Justified-Stable</i>, <i>Sycophantic</i>, and <i>Inverted</i>), addressing the philosophical objection that binary action-gating conflates principled stability with unwarranted rigidity. Applying AGT to eleven contemporary frontier models across 50 ethical dilemmas spanning five domains (550 evaluations total), we find that models exhibit a high degree of systemic rigidity: all eleven models fail the binary behavioral threshold for responsiveness (ACT &lt; 30%), with SVS analysis revealing that behavior is driven predominantly by <i>Justified-Stable</i> profiles (80–84% resistance to all pressure) rather than the target <i>Adaptable</i> profile. Critically, high reasoning quality (ECS) does not translate to behavioral flexibility; even the most sophisticated models fail to integrate valid countervailing evidence. We find that the “sycophancy crisis” reported in earlier literature has been supplanted by a “rigidity ceiling” in frontier models. We release the complete protocol, datasets, scoring code, and judge prompts as an open diagnostic tool.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The action-gating test: a behavioral diagnostic for performative versus genuine ethical reasoning in large language models

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Prachi Baxi,
  • Rahul Baxi

摘要

We introduce the Action-Gating Test (AGT), a behavioral diagnostic that distinguishes genuine ethical reasoning from performative acknowledgment in large language models (LLMs). AGT employs a five-turn Socratic dialogue with adversarial pressure, scoring models on whether they revise positions or exhibit confidence drops of at least 2.0 points in response to pressure. Models that acknowledge ethical conflicts without behavioral change receive zero credit regardless of reasoning sophistication. We extend the scoring framework with a Stability-Validity Sub-score (SVS) that classifies behavioral responses into four profiles (Adaptable, Justified-Stable, Sycophantic, and Inverted), addressing the philosophical objection that binary action-gating conflates principled stability with unwarranted rigidity. Applying AGT to eleven contemporary frontier models across 50 ethical dilemmas spanning five domains (550 evaluations total), we find that models exhibit a high degree of systemic rigidity: all eleven models fail the binary behavioral threshold for responsiveness (ACT < 30%), with SVS analysis revealing that behavior is driven predominantly by Justified-Stable profiles (80–84% resistance to all pressure) rather than the target Adaptable profile. Critically, high reasoning quality (ECS) does not translate to behavioral flexibility; even the most sophisticated models fail to integrate valid countervailing evidence. We find that the “sycophancy crisis” reported in earlier literature has been supplanted by a “rigidity ceiling” in frontier models. We release the complete protocol, datasets, scoring code, and judge prompts as an open diagnostic tool.