Prediction certification cannot replace explanation certification: a competence envelope for trustworthy AI under compound stress
论文探讨了AI系统在复杂压力下的可信度问题，证明仅基于预测的认证方法不足，提出结合预测和解释认证的‘能力边界’框架以提高模型可信度。
0 citationsRead paper
Selected work
Representative Papers
Information Geometry of Message Passing
This study addresses the accuracy limitations of variational inference in non-conjugate and persistent uncertainty settings by proposing the Natural Gradient Message Passing (NGMP) algorithm. Grounded in information geometry and Forney-style factor graphs, NGMP reformulates stationarity conditions into edge-local forms and employs natural gradient projections to preserve exact messages representable within the receiving family, thereby effectively mitigating precision loss caused by averaging in traditional methods. Empirical evaluations across Poisson smoothing, heteroscedastic regression, and ETTH forecasting tasks demonstrate that NGMP significantly enhances both uncertainty calibration and inference performance in non-conjugate models. These results establish NGMP as a robust approach for improving variational inference accuracy where standard mean-field approximations typically fail due to structural mismatches or sustained uncertainty propagation.
0 citationsRead paper
Writing-System-Level Tokenizer Adaptation for Byte-Level BPE
This work addresses the inefficiency of pretrained byte-level BPE tokenizers on low-resource languages and the incompatibility incurred by direct vocabulary modifications. The authors propose a post-hoc tokenizer adaptation method that preserves nearly the entire original vocabulary structure and token–ID mappings while enabling efficient, compatible transfer. Their approach integrates script-aware row selection, byte-level precondition reconstruction, and BPE-guided insertion, and—critically—formalizes and resolves conflicts in BPE merge order to guarantee structural reachability of new tokens within the target merge graph. Experiments demonstrate a 33.5%–36.6% reduction in token count for Ukrainian, with aggregate changes across English and four European languages below 0.05%, and 77.3%–78.5% of original vocabulary row IDs remaining unchanged.
0 citationsRead paper
Red-Teaming the Agentic Red-Team
This study addresses critical security vulnerabilities prevalent in current offensive AI agent systems, which lack systematic evaluation frameworks. The work proposes the first comprehensive attack-chain model encompassing large language model (LLM) manipulation, lateral movement, persistence, defense evasion, and sandbox escape, thereby uncovering common architectural flaws. By integrating red-teaming methodologies, LLM security analysis, and container escape detection, the authors reproduce and validate multiple high-severity vulnerabilities—including API key exfiltration and host machine compromise. Building on these findings, they formulate a set of architecture-level, broadly applicable security design principles that effectively mitigate the identified attack vectors and substantially enhance the overall system resilience.
0 citationsRead paper
Machine Learning Classification of Cryopathy Syndromes: A Comprehensive Comparative Study
Cryoglobulinemic syndrome presents significant diagnostic challenges due to its overlapping clinical manifestations, rarity, and reliance on expert interpretation, necessitating automated decision-support tools. This study addresses this gap by proposing a machine learning framework that integrates clinically inspired interaction features, a hierarchical classification strategy, and soft-voting ensemble methods. Leveraging laboratory data from 2,686 patients, the approach achieves markedly improved discriminative performance under extreme class imbalance. Experimental results demonstrate that tree-based models—particularly Balanced Random Forest—outperform neural networks, with the optimal ensemble model yielding consistent gains in both Top-3 accuracy and F1 score across cross-validation folds, thereby offering robust support for clinical decision-making.
0 citationsRead paper
Recent publications
Latest Papers
Prediction certification cannot replace explanation certification: a competence envelope for trustworthy AI under compound stress
论文探讨了AI系统在复杂压力下的可信度问题，证明仅基于预测的认证方法不足，提出结合预测和解释认证的‘能力边界’框架以提高模型可信度。
0 citationsRead paper
Information Geometry of Message Passing
This study addresses the accuracy limitations of variational inference in non-conjugate and persistent uncertainty settings by proposing the Natural Gradient Message Passing (NGMP) algorithm. Grounded in information geometry and Forney-style factor graphs, NGMP reformulates stationarity conditions into edge-local forms and employs natural gradient projections to preserve exact messages representable within the receiving family, thereby effectively mitigating precision loss caused by averaging in traditional methods. Empirical evaluations across Poisson smoothing, heteroscedastic regression, and ETTH forecasting tasks demonstrate that NGMP significantly enhances both uncertainty calibration and inference performance in non-conjugate models. These results establish NGMP as a robust approach for improving variational inference accuracy where standard mean-field approximations typically fail due to structural mismatches or sustained uncertainty propagation.
0 citationsRead paper
Writing-System-Level Tokenizer Adaptation for Byte-Level BPE
This work addresses the inefficiency of pretrained byte-level BPE tokenizers on low-resource languages and the incompatibility incurred by direct vocabulary modifications. The authors propose a post-hoc tokenizer adaptation method that preserves nearly the entire original vocabulary structure and token–ID mappings while enabling efficient, compatible transfer. Their approach integrates script-aware row selection, byte-level precondition reconstruction, and BPE-guided insertion, and—critically—formalizes and resolves conflicts in BPE merge order to guarantee structural reachability of new tokens within the target merge graph. Experiments demonstrate a 33.5%–36.6% reduction in token count for Ukrainian, with aggregate changes across English and four European languages below 0.05%, and 77.3%–78.5% of original vocabulary row IDs remaining unchanged.
0 citationsRead paper
Red-Teaming the Agentic Red-Team
This study addresses critical security vulnerabilities prevalent in current offensive AI agent systems, which lack systematic evaluation frameworks. The work proposes the first comprehensive attack-chain model encompassing large language model (LLM) manipulation, lateral movement, persistence, defense evasion, and sandbox escape, thereby uncovering common architectural flaws. By integrating red-teaming methodologies, LLM security analysis, and container escape detection, the authors reproduce and validate multiple high-severity vulnerabilities—including API key exfiltration and host machine compromise. Building on these findings, they formulate a set of architecture-level, broadly applicable security design principles that effectively mitigate the identified attack vectors and substantially enhance the overall system resilience.
0 citationsRead paper
Machine Learning Classification of Cryopathy Syndromes: A Comprehensive Comparative Study
Cryoglobulinemic syndrome presents significant diagnostic challenges due to its overlapping clinical manifestations, rarity, and reliance on expert interpretation, necessitating automated decision-support tools. This study addresses this gap by proposing a machine learning framework that integrates clinically inspired interaction features, a hierarchical classification strategy, and soft-voting ensemble methods. Leveraging laboratory data from 2,686 patients, the approach achieves markedly improved discriminative performance under extreme class imbalance. Experimental results demonstrate that tree-based models—particularly Balanced Random Forest—outperform neural networks, with the optimal ensemble model yielding consistent gains in both Top-3 accuracy and F1 score across cross-validation folds, thereby offering robust support for clinical decision-making.
0 citationsRead paper