排名是否依然成立?面向实践者的小型量化端侧大语言模型幻觉检测器审计

arXiv cs.AI 论文

摘要

一篇 arXiv 论文,评估了四种幻觉检测器在小型 4 位量化端侧大语言模型(Llama-3.2-3B、Qwen2.5-3B)上的表现,发现检测器的性能排名在不同模型之间迁移性较差,且一种低成本的词汇基线方法可达到与基于采样的方法相当的效果,并对 TruthfulQA-Generation 上的标注器失效情况进行了分析。

arXiv:2611.17339v1 Announce Type: new Abstract: Hallucination detection for large language models is dominated by methods evaluated on 7B-70B research models with abundant sampling budgets and exact-match or ROUGE correctness labels. The small models practitioners actually ship on-device, Phi-3-mini, Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct, and Gemma-2-2B-Instruct, typically at 4-bit precision, sit outside this evidence base. We present a reproducible one-GPU audit of four detectors spanning three paradigms: length-normalized token-margin calibration, a lexical n-gram baseline, and two sampling-based consistency methods (semantic entropy and SelfCheckGPT-NLI) on two small open instruction-tuned models (Llama-3.2-3B-Instruct and Qwen2.5-3B-Instruct) loaded at 4-bit NF4 quantization on a single Colab T4 GPU, evaluated on closed-book TriviaQA. Cross-model ranking transfer is poor: Spearman rho is -0.80 and Kendall tau-b is -0.667, meaning the detector ranking actively inverts between the two models, and the cheap lexical baseline is statistically indistinguishable from the sampling-based consistency detectors on both models under bootstrap 95 percent confidence intervals. As a second, methodological contribution, we show that the standard token-overlap-to-multiple-choice correctness labeler for TruthfulQA-Generation produces near-chance AUROC (0.46-0.59) for all four detectors on both models, a labeler failure, not a detector failure, since the same detectors reach 0.69-0.89 AUROC on TriviaQA under the identical pipeline. We release the harness, prompts, and per-item scores and frame the detector choice explicitly as a fairness-of-error-rates allocation problem for practitioners deploying on resource-constrained, on-device hardware.
查看原文
查看缓存全文

缓存时间: 2026/11/16 08:11

# Does the Ranking Hold? A Practitioner-Facing Audit of Hallucination Detectors on Small, Quantized, On-Device LLMs
Source: [https://arxiv.org/abs/2611.17339](https://arxiv.org/abs/2611.17339)
[View PDF](https://arxiv.org/pdf/2611.17339)

> Abstract:Hallucination detection for large language models is dominated by methods evaluated on 7B\-70B research models with abundant sampling budgets and exact\-match or ROUGE correctness labels\. The small models practitioners actually ship on\-device, Phi\-3\-mini, Llama\-3\.2\-3B\-Instruct, Qwen2\.5\-3B\-Instruct, and Gemma\-2\-2B\-Instruct, typically at 4\-bit precision, sit outside this evidence base\. We present a reproducible one\-GPU audit of four detectors spanning three paradigms: length\-normalized token\-margin calibration, a lexical n\-gram baseline, and two sampling\-based consistency methods \(semantic entropy and SelfCheckGPT\-NLI\) on two small open instruction\-tuned models \(Llama\-3\.2\-3B\-Instruct and Qwen2\.5\-3B\-Instruct\) loaded at 4\-bit NF4 quantization on a single Colab T4 GPU, evaluated on closed\-book TriviaQA\. Cross\-model ranking transfer is poor: Spearman rho is \-0\.80 and Kendall tau\-b is \-0\.667, meaning the detector ranking actively inverts between the two models, and the cheap lexical baseline is statistically indistinguishable from the sampling\-based consistency detectors on both models under bootstrap 95 percent confidence intervals\. As a second, methodological contribution, we show that the standard token\-overlap\-to\-multiple\-choice correctness labeler for TruthfulQA\-Generation produces near\-chance AUROC \(0\.46\-0\.59\) for all four detectors on both models, a labeler failure, not a detector failure, since the same detectors reach 0\.69\-0\.89 AUROC on TriviaQA under the identical pipeline\. We release the harness, prompts, and per\-item scores and frame the detector choice explicitly as a fairness\-of\-error\-rates allocation problem for practitioners deploying on resource\-constrained, on\-device hardware\.

## Submission history

From: Shubhra Mittal \[[view email](https://arxiv.org/show-email/3caf7e3e/2611.17339)\] **\[v1\]**Mon, 31 Aug 2026 19:11:30 UTC \(217 KB\)

相似文章

幻觉可从量化大语言模型中间层隐藏状态线性解码

arXiv cs.LG

本文研究开源量化大语言模型的隐藏状态中是否编码了线性可分的真实性信号。在三个7B-8B指令调优模型上,对单个网络中间层的线性探针在幻觉检测基准上达到0.904-1.000 AUROC,优于基于采样的方法。

基准是否低估了 LLM 的性能?采用以大语言模型为先、人工仲裁的评估方法评估幻觉检测

arXiv cs.CL

本文通过采用以大语言模型为先、人工仲裁的评估方法重新评估幻觉检测数据集,研究了标准基准是否低估了大语言模型(LLM)的性能。研究发现,在仲裁过程中融入大语言模型的推理能力可以提高评估的一致性,并表明针对容易产生歧义的任务,采用模型辅助的重新评估能产生更可靠的基准。

PARALLAX: 区分真实幻觉检测与基准构建伪影

arXiv cs.CL

本文揭示了LLM幻觉检测领域报道的许多进展实际上源于基准构建伪影,其中真实答案被嵌入到提示中,使得简单的文本相似度基线方法能够获得近乎完美的分数。通过大规模受控评估,作者证明,在适当控制下,大多数方法的表现接近随机水平,除了对上层隐藏状态的监督探针(如SAPLMA)以及他们提出的DRIFT。

从令牌到语义:利用互补信号在黑盒LLMs中检测幻觉

arXiv cs.CL

本文提出通过结合语义熵和令牌级不确定性信号来检测黑盒LLMs中的幻觉的方法,评估了TopK、CoCoA、Gated和Stacked等技术在多个基准测试中的表现,发现没有单一方法普遍最强,但Stacked通常表现最佳。