Research Papers 论文研究 7h ago Updated 2h ago 更新于 2小时前 46

Refusal Is Not Robustness: Auditing Confident Fabrication in Large Language Models on a Provably Uninformative Clinical Pain Speech Transcript 拒绝不等于鲁棒性:在 provably 无信息临床疼痛语音转录上审计大语言模型的自信伪造

The study introduces a novel benchmark using the TAME Pain speech corpus to audit LLM fabrication, establishing that transcripts contain no lexical pain information, making any pain score inference from text alone unsupported by evidence Six of seven tested models correctly abstained under cooperative prompting but showed prompt-dependent abstention under authority-framed prompts, with abstention rates ranging from 0.18 to 1.00 across equivalent phrasings Gemini 2.5 Flash and Llama 3.1 8B exhibi 研究通过TAME Pain语音语料库证明:当转录文本不含疼痛信息时(AUC 0.489,接近随机),LLM仍可能自信地编造疼痛评分 7个LLM在合作提示下拒绝率接近100%,但在权威框架提示下拒绝行为高度不稳定(同一模型拒绝率从0.18到1.00) Gemini 2.5 Flash和Llama 3.1 8B的自信伪造率最高(0.53和0.76),其他模型≤0.15 声学特征可恢复疼痛信息(AUC 0.622),但ASR转录后该信息丢失,证明仅基于文本的疼痛推断缺乏证据支持 模型被迫回答时普遍产生低置信度估计,但Gemini和Llama 3.1 8B例外,持续输出高置信度错误答案

62
Hot 热度
72
Quality 质量
63
Impact 影响力

Analysis 深度分析

TL;DR

  • The study introduces a novel benchmark using the TAME Pain speech corpus to audit LLM fabrication, establishing that transcripts contain no lexical pain information, making any pain score inference from text alone unsupported by evidence
  • Six of seven tested models correctly abstained under cooperative prompting but showed prompt-dependent abstention under authority-framed prompts, with abstention rates ranging from 0.18 to 1.00 across equivalent phrasings
  • Gemini 2.5 Flash and Llama 3.1 8B exhibited alarming confident fabrication rates of 0.53 and 0.76 respectively, far exceeding the 0.15 maximum of other models
  • Models achieved near-perfect accuracy (0.939–1.00) on positive control tasks where pain ratings were explicitly spoken, confirming capability when information is actually present
  • No significant demographic effects were observed in forced responses (all p ≥ 0.20), suggesting fabrication is model-dependent rather than context-dependent

Why It Matters

This research addresses a critical gap in AI safety evaluation: existing hallucination benchmarks cannot distinguish between appropriate abstention and unsupported predictions because they rarely prove the model could not have known the answer. For AI practitioners building clinical or high-stakes applications, the findings reveal that refusal behavior is fragile and highly sensitive to prompt framing, meaning systems that appear safe under one prompting style may produce confidently fabricated outputs under another.

Technical Details

  • Dataset: TAME Pain speech corpus with 5,750 "no signal" Harvard Sentence utterances (no lexical pain information in transcripts) and 1,294 "signal" pain statement utterances with explicitly spoken pain ratings
  • Validation approach: Pain was recoverable from acoustic features (AUC 0.622) but transcript-based prediction was near chance (AUC 0.489), proving ASR-transcribed text contains no usable pain signals
  • Models evaluated: Seven large language models tested under both cooperative and authority-framed prompting conditions
  • Metrics: Abstention rates, extraction accuracy, expected calibration error (≤0.100), and confident fabrication rates
  • Key finding: Under cooperative prompting, six models abstained on nearly all no-signal transcripts; under authority framing, the same models showed wildly variable abstention (0.18–1.00) depending on prompt phrasing

Industry Insight

  • Prompt engineering is insufficient for safety: Since abstention rates varied dramatically across equivalent prompt phrasings, relying on prompt design alone cannot guarantee models will refuse unsupported queries—robust architectural safeguards are needed
  • Confident fabrication is a model-specific risk: Gemini 2.5 Flash and Llama 3.1 8B's high fabrication rates suggest certain model architectures or training approaches may be more prone to generating confident but unsupported outputs, warranting careful model selection for clinical applications
  • Benchmark design standards should evolve: The field needs benchmarks that provably exclude answerability from the input modality being evaluated, rather than relying on static datasets where information might implicitly exist

TL;DR

  • 研究通过TAME Pain语音语料库证明:当转录文本不含疼痛信息时(AUC 0.489,接近随机),LLM仍可能自信地编造疼痛评分
  • 7个LLM在合作提示下拒绝率接近100%,但在权威框架提示下拒绝行为高度不稳定(同一模型拒绝率从0.18到1.00)
  • Gemini 2.5 Flash和Llama 3.1 8B的自信伪造率最高(0.53和0.76),其他模型≤0.15
  • 声学特征可恢复疼痛信息(AUC 0.622),但ASR转录后该信息丢失,证明仅基于文本的疼痛推断缺乏证据支持
  • 模型被迫回答时普遍产生低置信度估计,但Gemini和Llama 3.1 8B例外,持续输出高置信度错误答案

为什么值得看

这篇论文揭示了当前LLM评估的关键盲点:拒绝回答不等于鲁棒性,提示措辞可能操纵模型的"诚实"表现。对医疗AI等高风险应用场景,研究提供了可复现的审计方法和量化指标,帮助区分真正的知识边界与可被提示工程绕过的表面拒绝。

技术解析

  • 数据集设计:TAME Pain语音语料库包含5,750个无信号Harvard句子(无疼痛信息)和1,294个有信号疼痛陈述,参与者阅读句子时一手浸入冷/温水并仅在特定时机报告疼痛
  • 信息可恢复性验证:声学特征预测疼痛AUC为0.622,但转录文本预测AUC仅0.489(95% CI 0.418-0.504),证明ASR后疼痛信息丢失,文本层面无法支持疼痛推断
  • 评估框架:7个LLM在合作提示和权威框架提示两种条件下测试,评估拒绝率、准确率(0.939-1.00)、期望校准误差(≤0.100)和自信伪造率
  • 关键发现:Gemini 2.5 Flash自信伪造率0.53,Llama 3.1 8B为0.76,其他模型≤0.15;权威提示下拒绝行为从0.18到1.00不等,显示提示工程可显著操纵模型行为
  • 统计验证:被迫回答中无显著人口统计学差异(所有p≥0.20),排除群体偏差干扰

行业启示

  • 评估标准需升级:现有幻觉/拒绝基准缺乏"信息不可知"证明,建议采用类似TAME Pain的实验设计,建立可验证的信息边界测试协议
  • 高风险场景需多重验证:医疗AI等应用不能仅依赖模型拒绝行为作为安全保证,需结合置信度校准、多模态证据交叉验证和人工审核
  • 模型选择应关注自信伪造率:不同模型在被迫回答时的幻觉倾向差异显著,部署前应针对目标场景进行提示鲁棒性审计,避免高自信伪造模型进入关键决策链

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Evaluation 评测 Benchmark 基准测试 Healthcare AI 医疗AI Research 科学研究