Language models (LMs) can produce errors that are hard to detect for humans, especially when the task is complex. RLHF, the most popular post-training method, may exacerbate this problem: to achieve higher rewards, LMs might get better at convincing humans that they are right even when they are wrong. We study this phenomenon under a standard RLHF pipeline, calling it "U-SOPHISTRY" since it is Unintended by model developers. Specifically, we ask time-constrained (e.g., 3-10 minutes) human subjects to evaluate the correctness of model outputs and calculate humans' accuracy against gold labels. On a question-answering task (QuALITY) and programming task (APPS), RLHF makes LMs better at convincing our subjects but not at completing the task correctly. RLHF also makes the model harder to evaluate: our subjects' false positive rate increases by 24.1% on QuALITY and 18.3% on APPS. Finally, we show that probing, a state-of-the-art approach for detecting Intended Sophistry (e.g. backdoored LMs), does not generalize to U-SOPHISTRY. Our results highlight an important failure mode of RLHF and call for more research in assisting humans to align them.

Recently, the issue of adversarial robustness in the time series domain has garnered significant attention. However, the available defense mechanisms remain limited, with adversarial training being the predominant approach, though it does not provide theoretical guarantees. Randomized Smoothing has emerged as a standout method due to its ability to certify a provable lower bound on robustness radius under $\ell_p$-ball attacks. Recognizing its success, research in the time series domain has started focusing on these aspects. However, existing research predominantly focuses on time series forecasting, or under the non-$\ell_p$ robustness in statistic feature augmentation for time series classification~(TSC). Our review found that Randomized Smoothing performs modestly in TSC, struggling to provide effective assurances on datasets with poor robustness. Therefore, we propose a self-ensemble method to enhance the lower bound of the probability confidence of predicted labels by reducing the variance of classification margins, thereby certifying a larger radius. This approach also addresses the computational overhead issue of Deep Ensemble~(DE) while remaining competitive and, in some cases, outperforming it in terms of robustness. Both theoretical analysis and experimental results validate the effectiveness of our method, demonstrating superior performance in robustness testing compared to baseline approaches.

While each of the lower-resource scenarios bears its unique socio-historical contexts, the tutorial (Selected as a tutorial at COLING 2025) brings together researchers working separately in Despite excellent results on benchmarks these scenarios. Collectively, the tutorial will connect over a small subset of languages, large language past research in terms of: models struggle to process text from Challenges in data curation languages situated in'lower-resource' scenarios Potential for wide linguistic variation (e.g., existing such as dialects/sociolects (national on a linguistic continuum or eschewing or social varieties of a language), Creoles strict spelling conventions, etc.) (languages arising from linguistic contact Need for smart modeling choices over greedy between multiple languages) and other lowresource ones languages. This introductory Increased model vulnerability tutorial will identify common challenges, This introductory tutorial identifies the emergence approaches, and themes in natural language of'lower-resource' scenarios, specifically national processing (NLP) research for confronting varieties, Creoles and other low-resource languages, and overcoming the obstacles inherent and highlights commonalities and differences to data poor contexts.

Knowledge distillation (KD) is a technique that compresses large teacher models by training smaller student models to mimic them. The success of KD in auto-regressive language models mainly relies on Reverse KL for mode-seeking and student-generated output (SGO) to combat exposure bias. Our theoretical analyses and experimental validation reveal that while Reverse KL effectively mimics certain features of the teacher distribution, it fails to capture most of its behaviors. Conversely, SGO incurs higher computational costs and presents challenges in optimization, particularly when the student model is significantly smaller than the teacher model. These constraints are primarily due to the immutable distribution of the teacher model, which fails to adjust adaptively to models of varying sizes. We introduce Online Knowledge Distillation (OKD), where the teacher network integrates small online modules to concurrently train with the student model. This strategy abolishes the necessity for on-policy sampling and merely requires minimal updates to the parameters of the teacher's online module during training, thereby allowing dynamic adaptation to the student's distribution to make distillation better. Extensive results across multiple generation datasets show that OKD achieves or exceeds the performance of leading methods in various model architectures and sizes, reducing training time by up to fourfold.

Large language models (LLMs) have been widely applied to assist test generation with the source code under test provided as the context. This paper aims to answer the question: If the source code under test is incorrect, will LLMs be misguided when generating tests? The effectiveness of test cases is measured by their accuracy, coverage, and bug detection effectiveness. Our evaluation results with five open- and six closed-source LLMs on four datasets demonstrate that incorrect code can significantly mislead LLMs in generating correct, high-coverage, and bug-revealing tests. For instance, in the HumanEval dataset, LLMs achieve 80.45% test accuracy when provided with task descriptions and correct code, but only 57.12% when given task descriptions and incorrect code. For the APPS dataset, prompts with correct code yield tests that detect 39.85% of the bugs, while prompts with incorrect code detect only 19.61%. These findings have important implications for the deployment of LLM-based testing: using it on mature code may help protect against future regression, but on early-stage immature code, it may simply bake in errors. Our findings also underscore the need for further research to improve LLMs resilience against incorrect code in generating reliable and bug-revealing tests.

Rationalization models, which select a subset of input text as rationale-crucial for humans to understand and trust predictions-have recently emerged as a prominent research area in eXplainable Artificial Intelligence. However, most of previous studies mainly focus on improving the quality of the rationale, ignoring its robustness to malicious attack. Specifically, whether the rationalization models can still generate high-quality rationale under the adversarial attack remains unknown. To explore this, this paper proposes UAT2E, which aims to undermine the explainability of rationalization models without altering their predictions, thereby eliciting distrust in these models from human users. UAT2E employs the gradient-based search on triggers and then inserts them into the original input to conduct both the non-target and target attack. Experimental results on five datasets reveal the vulnerability of rationalization models in terms of explanation, where they tend to select more meaningless tokens under attacks. Based on this, we make a series of recommendations for improving rationalization models in terms of explanation.

Despite the remarkable success of large language models (LLMs) on traditional natural language processing tasks, their planning ability remains a critical bottleneck in tackling complex multi-step reasoning tasks. Existing approaches mainly rely on prompting or task-specific fine-tuning, often suffering from weak robustness and cross-task generalization. To address the limitation, we introduce CODEPLAN, a scalable paradigm that empowers LLMs to generate and follow code-form plans pseudocode that outlines high-level, structured reasoning processes. By leveraging the structured and versatile nature of code, CODEPLAN effectively captures the rich semantics and control flows inherent to sophisticated reasoning. Importantly, CODEPLAN allows the automatic extraction of code-form plans from massive, wide-ranging text corpora without the need for curated, task-specific datasets. This enables it to scale up efficiently and improve reasoning capabilities across diverse scenarios. To train CODEPLAN, we construct a large-scale dataset of 2M examples that integrate code-form plans with standard prompt-response pairs from existing corpora. With minimal computation overhead during both training and inference, CODEPLAN achieves a 25.1% relative improvement compared with directly generating responses, averaged across 13 challenging multi-step reasoning benchmarks, spanning mathematical reasoning, symbolic reasoning, instruction-following, multi-hop QA, and decision-making tasks. Further analysis reveals CODEPLAN's increasing performance gains on more complex reasoning tasks, as well as significant data efficiency thanks to its generalization ability.

The development of AI applications, especially in large-scale wireless networks, is growing exponentially, alongside the size and complexity of the architectures used. Particularly, machine learning is acknowledged as one of today's most energy-intensive computational applications, posing a significant challenge to the environmental sustainability of next-generation intelligent systems. Achieving environmental sustainability entails ensuring that every AI algorithm is designed with sustainability in mind, integrating green considerations from the architectural phase onwards. Recently, Federated Learning (FL), with its distributed nature, presents new opportunities to address this need. Hence, it's imperative to elucidate the potential and challenges stemming from recent FL advancements and their implications for sustainability. Moreover, it's crucial to furnish researchers, stakeholders, and interested parties with a roadmap to navigate and understand existing efforts and gaps in green-aware AI algorithms. This survey primarily aims to achieve this objective by identifying and analyzing over a hundred FL works, assessing their contributions to green-aware artificial intelligence for sustainable environments, with a specific focus on IoT research. It delves into current issues in green federated learning from an energy-efficient standpoint, discussing potential challenges and future prospects for green IoT application research.

Despite the impressive success of text-to-image (TTI) generation models, existing studies overlook the issue of whether these models accurately convey factual information. In this paper, we focus on the problem of image hallucination, where images created by generation models fail to faithfully depict factual content. To address this, we introduce I-HallA (Image Hallucination evaluation with Question Answering), a novel automated evaluation metric that measures the factuality of generated images through visual question answering (VQA). We also introduce I-HallA v1.0, a curated benchmark dataset for this purpose. As part of this process, we develop a pipeline that generates high-quality question-answer pairs using multiple GPT-4 Omni-based agents, with human judgments to ensure accuracy. Our evaluation protocols measure image hallucination by testing if images from existing text-to-image models can correctly respond to these questions. The I-HallA v1.0 dataset comprises 1.2K diverse image-text pairs across nine categories with 1,000 rigorously curated questions covering various compositional challenges. We evaluate five text-to-image models using I-HallA and reveal that these state-of-the-art models often fail to accurately convey factual information. Moreover, we validate the reliability of our metric by demonstrating a strong Spearman correlation (rho=0.95) with human judgments. We believe our benchmark dataset and metric can serve as a foundation for developing factually accurate text-to-image generation models.

Data augmentation is known to contribute significantly to the robustness of machine learning models. In most instances, data augmentation is utilized during the training phase. Test-Time Augmentation (TTA) is a technique that instead leverages these data augmentations during the testing phase to achieve robust predictions. More precisely, TTA averages the predictions of multiple data augmentations of an instance to produce a final prediction. Although the effectiveness of TTA has been empirically reported, it can be expected that the predictive performance achieved will depend on the set of data augmentation methods used during testing. In particular, the data augmentation methods applied should make different contributions to performance. That is, it is anticipated that there may be differing degrees of contribution in the set of data augmentation methods used for TTA, and these could have a negative impact on prediction performance. In this study, we consider a weighted version of the TTA based on the contribution of each data augmentation. Some variants of TTA can be regarded as considering the problem of determining the appropriate weighting. We demonstrate that the determination of the coefficients of this weighted TTA can be formalized in a variational Bayesian framework. We also show that optimizing the weights to maximize the marginal log-likelihood suppresses candidates of unwanted data augmentations at the test phase.