Pre-trained language models can be surprisingly adept at tasks they were not explicitly trained on, but how they implement these capabilities is poorly understood.

Social media platforms have experienced a significant rise in toxic content, including abusive language and discriminatory remarks, presenting growing challenges for content moderation. Some users evade censorship by deliberately disguising toxic words through homophonic cloak, which necessitates the task of unveiling cloaked toxicity. Existing methods are mostly designed for English texts, while Chinese cloaked toxicity unveiling has not been solved yet. To tackle the issue, we propose C$^2$TU, a novel training-free and prompt-free method for Chinese cloaked toxic content unveiling. It first employs substring matching to identify candidate toxic words based on Chinese homo-graph and toxic lexicon. Then it filters those candidates that are non-toxic and corrects cloaks to be their corresponding toxicities. Specifically, we develop two model variants for filtering, which are based on BERT and LLMs, respectively. For LLMs, we address the auto-regressive limitation in computing word occurrence probability and utilize the full semantic contexts of a text sequence to reveal cloaked toxic words. Extensive experiments demonstrate that C$^2$TU can achieve superior performance on two Chinese toxic datasets. In particular, our method outperforms the best competitor by up to 71% on the F1 score and 35% on accuracy, respectively. Our code and data are available at https://github.com/XDxc-cuber/C2TU-Chinese-cloaked-toxicity-unveiling.

Test-time prompt tuning enhances zero-shot generalization of vision-language models but tends to ignore the relatedness among test samples during inference. Online test-time prompt tuning provides a simple way to leverage the information in previous test samples, albeit with the risk of prompt collapse due to error accumulation. To enhance test-time prompt tuning, we propose DynaPrompt, short for dynamic test-time prompt tuning, exploiting relevant data distribution information while reducing error accumulation. Built on an online prompt buffer, DynaPrompt adaptively selects and optimizes the relevant prompts for each test sample during tuning. Specifically, we introduce a dynamic prompt selection strategy based on two metrics: prediction entropy and probability difference. For unseen test data information, we develop dynamic prompt appending, which allows the buffer to append new prompts and delete the inactive ones. By doing so, the prompts are optimized to exploit beneficial information on specific test data, while alleviating error accumulation. Experiments on fourteen datasets demonstrate the effectiveness of dynamic test-time prompt tuning. Despite achieving remarkable successes, foundation models such as Contrastive Language-Image Pretraining (CLIP) (Radford et al., 2021) still suffer from distribution shifts when adapting to downstream tasks (Zhou et al., 2022a;b; Xiao et al., 2024). To improve test-time adaptation of the model in the presence of distribution shifts, recent works introduce learnable prompts at test time. The methods freeze the CLIP model parameters while only tuning the learnable prompts for test data. As shown in Figure 1a, test-time prompt tuning (TPT) (Shu et al., 2022) adapts the prompt to each test sample individually, which is widely followed by recent works (Ma et al., 2023; Samadh et al., 2023; Yoon et al., 2024).

Discriminative learning when training and test data belong to different distributions is a challenging and complex task. Often times we have very few or no labeled data from the test or target distribution but may have plenty of labeled data from multiple related sources with different distributions. The difference in distributions may be both in marginal and conditional probabilities. Most of the existing domain adaptation work focuses on the marginal probability distribution difference between the domains, assuming that the conditional probabilities are similar. However in many real world applications, conditional probability distribution differences are as commonplace as marginal probability differences. In this paper we propose a two-stage domain adaptation methodology which combines weighted data from multiple sources based on marginal probability differences (first stage) as well as conditional probability differences (second stage), with the target domain data. The weights for minimizing the marginal probability differences are estimated independently, while the weights for minimizing conditional probability differences are computed simultaneously by exploiting the potential interaction among multiple sources. We also provide a theoretical analysis on the generalization performance of the proposed multi-source domain adaptation formulation using the weighted Rademacher complexity measure. Empirical comparisons with existing state-of-the-art domain adaptation methods using three real-world datasets demonstrate the effectiveness of the proposed approach.

Pre-trained language models can be surprisingly adept at tasks they were not explicitly trained on, but how they implement these capabilities is poorly understood. In this paper, we investigate the basic mathematical abilities often acquired by pre-trained language models. Concretely, we use mechanistic interpretability techniques to explain the (limited) mathematical abilities of GPT-2 small. As a case study, we examine its ability to take in sentences such as "The war lasted from the year 1732 to the year 17", and predict valid two-digit end years (years > 32). We first identify a circuit, a small subset of GPT-2 small's computational graph that computes this task's output. Then, we explain the role of each circuit component, showing that GPT-2 small's final multi-layer perceptrons boost the probability of end years greater than the start year. Finally, we find related tasks that activate our circuit. Our results suggest that GPT-2 small computes greater-than using a complex but general mechanism that activates across diverse contexts.

Discriminative learning when training and test data belong to different distributions is a challenging and complex task. Often times we have very few or no labeled data from the test or target distribution but may have plenty of labeled data from multiple related sources with different distributions. The difference in distributions may be in both marginal and conditional probabilities. Most of the existing domain adaptation work focuses on the marginal probability distribution difference between the domains, assuming that the conditional probabilities are similar. However in many real world applications, conditional probability distribution differences are as commonplace as marginal probability differences. In this paper we propose a two-stage domain adaptation methodology which combines weighted data from multiple sources based on marginal probability differences (first stage) as well as conditional probability differences (second stage), with the target domain data.

Textual adversarial attacks expose the vulnerabilities of text classifiers and can be used to improve their robustness. Existing context-aware methods solely consider the gold label probability and use the greedy search when searching an attack path, often limiting the attack efficiency. To tackle these issues, we propose PDBS, a context-aware textual adversarial attack model using Probability Difference guided Beam Search. The probability difference is an overall consideration of all class label probabilities, and PDBS uses it to guide the selection of attack paths. In addition, PDBS uses the beam search to find a successful attack path, thus avoiding suffering from limited search space. Extensive experiments and human evaluation demonstrate that PDBS outperforms previous best models in a series of evaluation metrics, especially bringing up to a +19.5% attack success rate. Ablation studies and qualitative analyses further confirm the efficiency of PDBS.

Fairness is crucial for neural networks which are used in applications with important societal implication. Recently, there have been multiple attempts on improving fairness of neural networks, with a focus on fairness testing (e.g., generating individual discriminatory instances) and fairness training (e.g., enhancing fairness through augmented training). In this work, we propose an approach to formally verify neural networks against fairness, with a focus on independence-based fairness such as group fairness. Our method is built upon an approach for learning Markov Chains from a user-provided neural network (i.e., a feed-forward neural network or a recurrent neural network) which is guaranteed to facilitate sound analysis. The learned Markov Chain not only allows us to verify (with Probably Approximate Correctness guarantee) whether the neural network is fair or not, but also facilities sensitivity analysis which helps to understand why fairness is violated. We demonstrate that with our analysis results, the neural weights can be optimized to improve fairness. Our approach has been evaluated with multiple models trained on benchmark datasets and the experiment results show that our approach is effective and efficient.

Discriminative learning when training and test data belong to different distributions is a challenging and complex task. Often times we have very few or no labeled data from the test or target distribution but may have plenty of labeled data from multiple related sources with different distributions. The difference in distributions may be in both marginal and conditional probabilities. Most of the existing domain adaptation work focuses on the marginal probability distribution difference between the domains, assuming that the conditional probabilities are similar. However in many real world applications, conditional probability distribution differences are as commonplace as marginal probability differences.