雅虎

When audio and text conflict, speech-enabled language models follow the text 10 times more often than when arbitrating between two text sources, even when explicitly instructed to trust the audio. Using ALME, a benchmark of 57,602 controlled audio-text conflict stimuli across 8 languages, we find that Gemini 2.0 Flash exhibits 16.6% text dominance under audio-text conflict versus 1.6% under text-text conflict with identical reliability cues. This gap is not explained by audio quality: audio-only accuracy (97.2%) exceeds cascade accuracy (93.9%), indicating audio embeddings preserve more information than text transcripts. We propose that text dominance reflects an asymmetry not in information content but in arbitration accessibility: how easily the model can reason over competing representations. This framework explains otherwise puzzling findings. Forcing transcription before answering increases text dominance (19% to 33%), sacrificing audio's information advantage without improving accessibility. Framing text as “deliberately corrupted” reduces text dominance by 80%. A fine-tuning ablation provides interventional evidence: training only the audio projection layer increases text dominance (+26.5%), while LoRA on the language model halves it (-23.9%), localizing text dominance to the LLM's reasoning rather than the audio encoder. Experiments across four state-of-the-art audio-LLMs and 8 languages show consistent trends with substantial cross-linguistic and cross-model variation, establishing modality arbitration as a distinct reliability dimension not captured by standard speech benchmarks.

Software supply chain frameworks, such as the US NIST Secure Software Development Framework (SSDF), detail what tasks software development organizations are recommended or mandated to adopt to reduce security risk. However, to further reduce the risk of similar attacks occurring, software organizations benefit from knowing what tasks mitigate attack techniques the attackers are currently using to address specific threats, prioritize tasks, and close mitigation gaps. The goal of this study is to aid software organizations in reducing the risk of software supply chain attacks by systematically synthesizing how framework tasks mitigate the attack techniques used in the SolarWinds, Log4j, and XZ Utils attacks. We qualitatively analyzed 106 Cyber Threat Intelligence (CTI) reports of the 3 attacks to gather the attack techniques. We then systematically constructed a mapping between attack techniques and the 73 tasks enumerated in 10 software supply chain frameworks. Afterward, we established and ranked priority tasks that mitigate attack techniques. The three mitigation tasks with the highest scores are role-based access control, system monitoring, and boundary protection. Additionally, three mitigation tasks were missing from all ten frameworks, including sustainable open-source software and environmental scanning tools. Thus, software products would still be vulnerable to software supply chain attacks even if organizations adopted all recommended tasks.

Bayesian quadrature is a probabilistic, model-based approach to numerical integration, the estimation of intractable integrals, or expectations. Although Bayesian quadrature was popularised already in the 1980s, no systematic and comprehensive treatment has been published. The purpose of this survey is to fill this gap. We review the mathematical foundations of Bayesian quadrature from different points of view; present a systematic taxonomy for classifying different Bayesian quadrature methods along the three axes of modelling, inference, and sampling; collect general theoretical guarantees; and provide a controlled numerical study that explores and illustrates the effect of different choices along the axes of the taxonomy. We also provide a realistic assessment of practical challenges and limitations to application of Bayesian quadrature methods and include an up-to-date and nearly exhaustive bibliography that covers not only machine learning and statistics literature but all areas of mathematics and engineering in which Bayesian quadrature or equivalent methods have seen use.

Neural networks gain capabilities during training, but the internal changes that precede capability acquisition are not well understood. In particular, the relationship between geometric change and behavioral change, and the effect of task difficulty and model scale on that relationship, is unclear. We track geometric measures and linear probes across six transformer sizes (405K–151M parameters), eight algorithmic tasks (144 task×level×model combinations), and three Pythia language models (160M–2.8B). Across all settings, representations first collapse to a low-dimensional state, then recover, and only then does behavioral performance improve. Linear probes show that the model's hidden states already contain task-relevant information before the model can act on it. The collapse floor is task-specific, the collapse propagates top-down through the network, and of the geometric measures tested, only reliably precedes capability acquisition for hard tasks. Whether this precursor is detectable depends on task difficulty relative to model capacity. For hard tasks, there is a clear gap: geometry changes first, behavior follows. For easy tasks, the model learns so quickly that both happen simultaneously and no precursor is detectable. On Pythia-2.8B, a logical deduction task that is genuinely hard for the model shows a precursor gap of ∼49K training steps, while easy benchmarks show none. This suggests that geometric patterns observed in small proxy models can persist at larger scale when the task remains difficult relative to model capacity.

The proposed scalable LLM-as-judge paradigm provides the possibility of developing an automatic and reliable evaluation scheme for generative AI systems using the concept of domain-adaptive calibration. More specifically, the proposed scheme includes three distinct modules, such as Counterfactual Evaluation Transformer (CET) model, Domain Knowledge Graph Calibration Network (DKGCN), and Entropy-Guided Meta-Judge Optimizer (EGMJO). More specifically, CET evaluates robustness and reasoning stability of responses based on comparison between original answers and their counterfactual counterparts; DKGCN uses graph neural networks to embed relationships among various items within finance, healthcare, cybersecurity, and engineering domains, and EGMJO relies on entropy-based meta-learning to measure uncertainties, resolve contradictions among judges, and update threshold values. Consequently, these modules generate an accurate evaluation score enhanced with information on reliability and confidence, reaching 98.2% accuracy.