SIREN-Bench: Behavior-Driven Generation and Evaluation of Emergency-Vehicle Interactions
为解决紧急车辆与普通车辆交互的安全评估问题,本文提出SIREN平台,通过SUMO和CARLA联合仿真生成行为驱动的交互场景,并基于此进行多项任务测试。
为解决紧急车辆与普通车辆交互的安全评估问题,本文提出SIREN平台,通过SUMO和CARLA联合仿真生成行为驱动的交互场景,并基于此进行多项任务测试。
研究通过监督集成框架,利用少量特定领域数据集上的异构UQ评分输出训练分类器,以提高LLM在无证据情况下的幻觉检测的鲁棒性。
This work addresses the prevalence of functional errors in code generated by large language models (LLMs) and the limited efficacy of existing uncertainty quantification (UQ) methods for code generation. The authors systematically evaluate the transferability of various UQ approaches and introduce a novel paradigm grounded in functional equivalence: leveraging LLMs to assess whether generated code is functionally equivalent to a reference implementation. Building on this, they define code-specific UQ metrics such as “functional entropy.” By integrating token-level probabilities, sampling strategies, and LLM-driven functional equivalence judgments, their method achieves state-of-the-art performance, attaining the highest AUROC in 11 out of 15 model–benchmark combinations and demonstrating substantially improved calibration over existing techniques.
Existing uncertainty quantification methods struggle to effectively detect hallucinations in long-form text generation. This work proposes the first fine-grained uncertainty quantification framework tailored for long texts, which systematically evaluates generated content through a three-stage pipeline: response decomposition, unit-level scoring, and response aggregation. The framework introduces a unified taxonomy that integrates and extends existing black-box consistency-based methods, enabling fair comparison and modular component selection. It employs black-box consistency scorers—including claim–response entailment, claim-level, and sentence-level scoring—combined with uncertainty-aware decoding. Experimental results demonstrate that claim–response entailment scoring achieves the best performance, claim-level scoring outperforms sentence-level scoring, and uncertainty-aware decoding significantly enhances factual consistency in long-form generation.
Hallucinations in large language models (LLMs) pose serious threats to the safety and reliability of downstream applications. To address this, we introduce the first open-source Python toolkit for LLM hallucination detection. Our core method is a response-level confidence scoring system grounded in uncertainty quantification: it integrates multiple state-of-the-art uncertainty estimation techniques—including logit entropy, sampling variance, and calibration-aware confidence—to produce interpretable, normalized confidence scores in the [0,1] range. The toolkit is designed for plug-and-play deployment, modular extensibility, and seamless integration with mainstream LLM frameworks (e.g., Hugging Face Transformers, vLLM). Extensive experiments across multiple benchmark datasets demonstrate that our approach significantly improves hallucination detection accuracy, achieving an average +12.3% F1-score gain over baseline methods. This advancement enhances both the trustworthiness of generated content and the operational safety of LLM deployments.
为解决紧急车辆与普通车辆交互的安全评估问题,本文提出SIREN平台,通过SUMO和CARLA联合仿真生成行为驱动的交互场景,并基于此进行多项任务测试。
研究通过监督集成框架,利用少量特定领域数据集上的异构UQ评分输出训练分类器,以提高LLM在无证据情况下的幻觉检测的鲁棒性。
This work addresses the prevalence of functional errors in code generated by large language models (LLMs) and the limited efficacy of existing uncertainty quantification (UQ) methods for code generation. The authors systematically evaluate the transferability of various UQ approaches and introduce a novel paradigm grounded in functional equivalence: leveraging LLMs to assess whether generated code is functionally equivalent to a reference implementation. Building on this, they define code-specific UQ metrics such as “functional entropy.” By integrating token-level probabilities, sampling strategies, and LLM-driven functional equivalence judgments, their method achieves state-of-the-art performance, attaining the highest AUROC in 11 out of 15 model–benchmark combinations and demonstrating substantially improved calibration over existing techniques.
Existing uncertainty quantification methods struggle to effectively detect hallucinations in long-form text generation. This work proposes the first fine-grained uncertainty quantification framework tailored for long texts, which systematically evaluates generated content through a three-stage pipeline: response decomposition, unit-level scoring, and response aggregation. The framework introduces a unified taxonomy that integrates and extends existing black-box consistency-based methods, enabling fair comparison and modular component selection. It employs black-box consistency scorers—including claim–response entailment, claim-level, and sentence-level scoring—combined with uncertainty-aware decoding. Experimental results demonstrate that claim–response entailment scoring achieves the best performance, claim-level scoring outperforms sentence-level scoring, and uncertainty-aware decoding significantly enhances factual consistency in long-form generation.
Hallucinations in large language models (LLMs) pose serious threats to the safety and reliability of downstream applications. To address this, we introduce the first open-source Python toolkit for LLM hallucination detection. Our core method is a response-level confidence scoring system grounded in uncertainty quantification: it integrates multiple state-of-the-art uncertainty estimation techniques—including logit entropy, sampling variance, and calibration-aware confidence—to produce interpretable, normalized confidence scores in the [0,1] range. The toolkit is designed for plug-and-play deployment, modular extensibility, and seamless integration with mainstream LLM frameworks (e.g., Hugging Face Transformers, vLLM). Extensive experiments across multiple benchmark datasets demonstrate that our approach significantly improves hallucination detection accuracy, achieving an average +12.3% F1-score gain over baseline methods. This advancement enhances both the trustworthiness of generated content and the operational safety of LLM deployments.