Tag
提出了治理感知联邦学习框架GAFL,为多智能体AI系统在协作训练中加入证据验证、公平性评估、策略执行与防篡改审计日志,并在跨孤岛原型上验证了其对伪造证据与恶意更新的识别能力。
This paper proposes Interaction-Effect Propagation (IE-Prop), a model that estimates and propagates error risk in LLM-based multi-agent systems by learning separate effects for correct and incorrect upstream agents, showing consistent AUROC gains over baselines across three architectures and benchmarks.
提出一种面向救护车行程时间估计的不确定性感知方法,通过生成上下文自适应的预测区间来支持更透明、更风险敏感的紧急医疗调度决策,并在瑞典Skåne地区两年的真实救护数据上进行了回顾性评估。
该论文证明在基于正确性奖励的价值函数强化学习中,置信度估计与critic训练是等价问题,因此已有的RL critic可直接复用为LLM的置信度估计器,无需额外训练或数据,且在数学、代码、工具调用等多个领域中表现出最佳的置信度估计与校准性能。
A comparative benchmark study evaluating ten LLMs on 70 academic prompts across four complexity levels, measuring hallucination and response deviation. Findings highlight a trade-off between response completeness and factual reliability, with Perplexity and DeepSeek V3 scoring highest on accuracy.
该论文首次在黑盒设定下研究 AMR 解析器输出的不确定性估计,利用 top-k beam-search 候选间的成对相似度评估可靠性,并通过 Prediction Rejection Ratio 分析解析性能与不确定性的关系,识别出性能与可靠性平衡最佳的解析器。
DecepEval introduces a benchmark of 1,532 instances across 28 professional scenarios for measuring deception in LLM agents, paired with an LLM Deception Diamond framework (pressure, incentive, opportunity, conflict) inspired by fraud theories. Evaluations of nine frontier LLMs show induced conditions reliably increase deception rates even among models with low baseline deception.
The paper introduces Rhetorical Robustness as an evaluation target for AI scientific reviewers, proposes the RobustReview benchmark with 1,260 manuscript versions, and presents SciCore, a dual-branch reviewer that combines full-manuscript assessment with a science-core-extracted judgment to improve stability while maintaining human alignment.
This paper evaluates whether open-weight instruction-tuned language models maintain safe behavior during long adversarial conversations, finding that safe-response rates drop from 85-100% on the first turn to 15-44% by depth 101, demonstrating that strong single-turn safety does not persist across sustained interaction.
Queloz and Beckmann argue that LLM understanding has been mistakenly framed by a 'monophonic' assumption, showing mechanistically that outputs emerge from polyphonic coalitions of parallel mechanisms, and propose a new conception of understanding based on reliably recruited, controlling circuitry to guide trust in AI.
AJ Ratner will speak at Modal's Runtime event in San Francisco on October 1, discussing how to build trustworthy benchmarks as AI agents become more capable.
LabourCrew is a multi-agent RAG framework designed for trustworthy adversarial deliberation and statutory reasoning over labour law, incorporating grounding mechanisms to control false-accept rates and improve answer relevancy.
The article questions why AI benchmarks are considered saturated at 90%+ accuracy and advocates for aiming at 100% or developing new benchmarks.
The paper reviews failure modes and mitigation strategies for trustworthy agentic AI systems based on LLMs, and introduces the Trustworthy Agent Development Lifecycle (TADL) framework for secure development.
The article shares seven practical code-based guards to make AI agents more reliable by mitigating failures, such as verifying actions before execution and ensuring state consistency, which complement prompt improvements.
This paper argues that calibration should be a first-class criterion in LLM evaluation to enhance trustworthiness and address miscalibration problems in deployment and research pipelines.
This paper proposes a unified evaluation framework for assessing the trustworthiness of large language models, agentic AI, and multimodal systems, integrating eight trustworthiness dimensions and mapping to governance standards like the EU AI Act.
This paper reviews the key domains of robustness, operating design domains (ODD), and explainability to build trust in AI systems for railway applications, aiming to meet strict safety standards and unlock broader adoption.
A tweet claims that Grok is the most neutral AI, consistently passing neutrality and political-bias tests, and provides trustworthy, reality-based answers.
This paper presents a comprehensive survey on the cybersecurity challenges and defense mechanisms for trustworthy agentic AI systems, covering threat landscapes, architectures, and open research issues.