Large Language Models (LLMs) are widely used in high-stakes applications, making their interpretability increasingly important.Existing interpretability methods are typically categorized into internal and external perspectives, which are often studied in isolation and tend to overlook two key aspects: causality and temporal dynamics.Explanations are often limited to surface correlations or static dependencies, failing to capture how influences evolve during autoregressive generation.To address these limitations, we propose a causal and dynamic interpretability framework for LLM generation.We first characterize the backdoor-adjusted causal effects of both the generated prefix and the prompt on the current token using the Structural Causal Model.Next, we introduce two metrics to quantify contextual causal influence and question-answer causal influence.Overall, our work provides a unified causal view of internal consistency and external alignment in LLM generation dynamics.1

The crux of data disclosure lies in the meticulous quantification and judicious trade-off between privacy leakage and utility. Firstly, the measurement of privacy leakage is the premise of sensitive data compliance disclosure. Existing solutions are mainly based on the qualitative perspective and the group perspective, which are unable to quantitatively measure the risk of individual privacy leakage. Secondly, in terms of balancing privacy leakage and utility, existing solutions overlook uninformed disclosure scenarios. In such scenarios, the two-dimensional privacy-utility game will be reduced to the optimization of a single parameter of mutual information. To address the aforementioned drawbacks, this paper proposes a data disclosure mechanism tailored for heterogeneous privacy profiles. Specifically, we decouple the multilevel privacy leakage through mutual information. By respectively addressing the informed and uninformed disclosure, we achieve the optimization of the joint potential energy surface of privacy and utility, breaking through the dilemma of excessive protection and utility collapse in data disclosure. On this basis, we conduct a comprehensive discussion on the two data disclosure strategies, namely introducing disturbance and linear obfuscation. The results of extensive simulations verify the effectiveness of the proposed mechanism. Finally, our work shows that balancing privacy and utility in data disclosure is unfeasible with disturbance-introducing strategies. In contrast, using linear obfuscation strategies can achieve such a balance, and the optimal approach is the disclosure scheme that minimizes the data disclosure loss.

The feverish personal data gold rush has made sensitive information leakage a non-negligible issue, turning it into a popular target for malicious attacks. Although some data may not seem to reveal private information directly, they could still be exploited through malicious intervention and inference; such data are referred to as risk data. Studies have claimed that such risk data exhibit the trait ofmacro-level collaborative leakage, meaning that individually harmless risk data can reveal sensitive information when combined. However, why the macro-level collaborative leakage will occur remains relatively uncharted. Hence, in this paper, we conduct rigorous quantitative analyses for the first time, to trace the root of the macro-level collaborative leakage. We conclude that this phenomenon arises from thecollaborative effectsamong pieces of risk data concerning sensitive information. In light of this, we formulate the sufficient condition for the occurrence of the macro-level collaborative leakage and investigate its presence in the Gaussian-distributed data. We highlight that, on the one hand, the Gaussian distribution can align the correlation between risk data and sensitive data at both the micro- and macro- levels, thereby preventing the macro-level collaborative leakage. This reveals the potential of the Gaussian distribution in enhancing data privacy protection from a macro perspective. On the other hand, the ability of the Gaussian distribution to counteract the macro-level collaborative leakage is inherently limited, which further corroborates the ubiquity of this phenomenon. Our insights underscore the need for more comprehensive security and privacy protection mechanisms to ensure data security and confidentiality.

The way of task posting serves as the main pillar in achieving an efficient crowdsourcing market. Pioneer solutions on task posting can be categorized as retail task posting and batch task posting. Unlike retail task posting, which simply matches the most suitable worker to tasks, batch task posting considers the collaborations not only between workers and tasks but also among tasks, which brings high efficiency, low costs, and satisfactory task completion rates. However, the state of the arts on batch task posting leverage specific attributes to combine tasks as bundles for posting, leading to limited scalability. Hence, we propose a causal analysis framework for batch crowdsourcing to achieve an attribute-independent batch crowdsourcing solution that disentangles multi-factors to uncover the posting merits of tasks bundled at optimal prices, based on which an approximately optimal algorithm is further introduced to form reasonable bundles for posting. Since batch crowdsourcing may incur losses due to short-term profit fluctuation, a risk assessment method is proposed to encourage the requestor to act properly for loss mitigations. Our work explores the causal analysis and risk assessment in batch crowdsourcing for the first time, with the following highlights: 1)generality. It proposes a composite metric for gauging task bundles which avoids the issue of attribute dependence in the state of the arts, resulting in better universality; 2)synergy. By collaboratively considering the “value” and “relative position” of variables, our work derives results reflecting causal relationships rather than naive correlations; and 3)precision. We not only elucidate the probability of risk in batch crowdsourcing but also delineate the rate function governing its probability decay. This allows a requestor to know when and how fast to halt batch task posting.