Great Models Think Alike and this Undermines AI Oversight

As Language Model (LM) capabilities advance, evaluating and supervising them at scale is getting harder for humans. There is hope that other language models can automate both these tasks, which we refer to as "AI Oversight". We study how model similarity affects both aspects of AI oversight by proposing a probabilistic metric for LM similarity based on overlap in model mistakes. Using this metric, we first show that LLM-as-a-judge scores favor models similar to the judge, generalizing recent self-preference results. Then, we study training on LM annotations, and find complementary knowledge between the weak supervisor and strong student model plays a crucial role in gains from "weak-to-strong generalization". As model capabilities increase, it becomes harder to find their mistakes, and we might defer more to AI oversight. However, we observe a concerning trend -- model mistakes are becoming more similar with increasing capabilities, pointing to risks from correlated failures. Our work underscores the importance of reporting and correcting for model similarity, especially in the emerging paradigm of AI oversight.

A Cognac shot to forget bad memories: Corrective Unlearning in GNNs

Graph Neural Networks (GNNs) are increasingly being used for a variety of ML applications on graph data. As graph data does not follow the independently and identically distributed (i.i.d) assumption, adversarial manipulations or incorrect data can propagate to other data points through message passing, deteriorating the model's performance. To allow model developers to remove the adverse effects of manipulated entities from a trained GNN, we study the recently formulated problem of Corrective Unlearning. We find that current graph unlearning methods fail to unlearn the effect of manipulations even when the whole manipulated set is known. We introduce a new graph unlearning method, Cognac, which can unlearn the effect of the manipulation set even when only 5% of it is identified. It recovers most of the performance of a strong oracle with fully corrected training data, even beating retraining from scratch without the deletion set while being 8x more efficient. We hope our work guides GNN developers in fixing harmful effects due to issues in real-world data post-training.

Exploring the Abilities of Large Language Models to Solve Proportional Analogies via Knowledge-Enhanced Prompting

Making analogies is fundamental to cognition. Proportional analogies, which consist of four terms, are often used to assess linguistic and cognitive abilities. For instance, completing analogies like "Oxygen is to Gas as <blank> is to <blank>" requires identifying the semantic relationship (e.g., "type of") between the first pair of terms ("Oxygen" and "Gas") and finding a second pair that shares the same relationship (e.g., "Aluminum" and "Metal"). In this work, we introduce a 15K Multiple-Choice Question Answering (MCQA) dataset for proportional analogy completion and evaluate the performance of contemporary Large Language Models (LLMs) in various knowledge-enhanced prompt settings. Specifically, we augment prompts with three types of knowledge: exemplar, structured, and targeted. Our results show that despite extensive training data, solving proportional analogies remains challenging for current LLMs, with the best model achieving an accuracy of 55%. Notably, we find that providing targeted knowledge can better assist models in completing proportional analogies compared to providing exemplars or collections of structured knowledge.

Just KIDDIN: Knowledge Infusion and Distillation for Detection of INdecent Memes

Toxicity identification in online multimodal environments remains a challenging task due to the complexity of contextual connections across modalities (e.g., textual and visual). In this paper, we propose a novel framework that integrates Knowledge Distillation (KD) from Large Visual Language Models (LVLMs) and knowledge infusion to enhance the performance of toxicity detection in hateful memes. Our approach extracts sub-knowledge graphs from ConceptNet, a large-scale commonsense Knowledge Graph (KG) to be infused within a compact VLM framework. The relational context between toxic phrases in captions and memes, as well as visual concepts in memes enhance the model's reasoning capabilities. Experimental results from our study on two hate speech benchmark datasets demonstrate superior performance over the state-of-the-art baselines across AU-ROC, F1, and Recall with improvements of 1.1%, 7%, and 35%, respectively. Given the contextual complexity of the toxicity detection task, our approach showcases the significance of learning from both explicit (i.e. KG) as well as implicit (i.e. LVLMs) contextual cues incorporated through a hybrid neurosymbolic approach. This is crucial for real-world applications where accurate and scalable recognition of toxic content is critical for creating safer online environments.

Rethinking Thinking Tokens: Understanding Why They Underperform in Practice

Thinking Tokens (TT) have been proposed as an unsupervised method to facilitate reasoning in language models. However, despite their conceptual appeal, our findings show that TTs marginally improves performance and consistently underperforms compared to Chain-of-Thought (CoT) reasoning across multiple benchmarks. We hypothesize that this underperformance stems from the reliance on a single embedding for TTs, which results in inconsistent learning signals and introduces noisy gradients. This paper provides a comprehensive empirical analysis to validate this hypothesis and discusses the implications for future research on unsupervised reasoning in LLMs.

LLM Vocabulary Compression for Low-Compute Environments

We present a method to compress the final linear layer of language models, reducing memory usage by up to 3.4x without significant performance loss. By grouping tokens based on Byte Pair Encoding (BPE) merges, we prevent materialization of the memory-intensive logits tensor. Evaluations on the TinyStories dataset show that our method performs on par with GPT-Neo and GPT2 while significantly improving throughput by up to 3x, making it suitable for low-compute environments.

SceneGraMMi: Scene Graph-boosted Hybrid-fusion for Multi-Modal Misinformation Veracity Prediction

Misinformation undermines individual knowledge and affects broader societal narratives. Despite growing interest in the research community in multi-modal misinformation detection, existing methods exhibit limitations in capturing semantic cues, key regions, and cross-modal similarities within multi-modal datasets. We propose SceneGraMMi, a Scene Graph-boosted Hybrid-fusion approach for Multi-modal Misinformation veracity prediction, which integrates scene graphs across different modalities to improve detection performance. Experimental results across four benchmark datasets show that SceneGraMMi consistently outperforms state-of-the-art methods. In a comprehensive ablation study, we highlight the contribution of each component, while Shapley values are employed to examine the explainability of the model's decision-making process.

Towards identifying Source credibility on Information Leakage in Digital Gadget Market

The use of Social media to share content is on a constant rise. One of the capsize effect of information sharing on Social media includes the spread of sensitive information on the public domain. With the digital gadget market becoming highly competitive and ever-evolving, the trend of an increasing number of sensitive posts leaking information on devices in social media is observed. Many web-blogs on digital gadget market have mushroomed recently, making the problem of information leak all pervasive. Credible leaks on specifics of an upcoming device can cause a lot of financial damage to the respective organization. Hence, it is crucial to assess the credibility of the platforms that continuously post about a smartphone or digital gadget leaks. In this work, we analyze the headlines of leak web-blog posts and their corresponding official press-release. We first collect 54, 495 leak and press-release headlines for different smartphones. We train our custom NER model to capture the evolving smartphone names with an accuracy of 82.14% on manually annotated results. We further propose a credibility score metric for the web-blog, based on the number of falsified and authentic smartphone leak posts.

Towards Infusing Auxiliary Knowledge for Distracted Driver Detection

Distracted driving is a leading cause of road accidents globally. Identification of distracted driving involves reliably detecting and classifying various forms of driver distraction (e.g., texting, eating, or using in-car devices) from in-vehicle camera feeds to enhance road safety. This task is challenging due to the need for robust models that can generalize to a diverse set of driver behaviors without requiring extensive annotated datasets. In this paper, we propose KiD3, a novel method for distracted driver detection (DDD) by infusing auxiliary knowledge about semantic relations between entities in a scene and the structural configuration of the driver's pose. Specifically, we construct a unified framework that integrates the scene graphs, and driver pose information with the visual cues in video frames to create a holistic representation of the driver's actions.Our results indicate that KiD3 achieves a 13.64% accuracy improvement over the vision-only baseline by incorporating such auxiliary knowledge with visual information.

Multilingual Non-Factoid Question Answering with Silver Answers

Most existing Question Answering Datasets (QuADs) primarily focus on factoid-based short-context Question Answering (QA) in high-resource languages. However, the scope of such datasets for low-resource languages remains limited, with only a few works centered on factoid-based QuADs and none on non-factoid QuADs. Therefore, this work presents MuNfQuAD, a multilingual QuAD with non-factoid questions. It utilizes interrogative sub-headings from BBC news articles as questions and the corresponding paragraphs as silver answers. The dataset comprises over 370K QA pairs across 38 languages, encompassing several low-resource languages, and stands as the largest multilingual QA dataset to date. Based on the manual annotations of 790 QA-pairs from MuNfQuAD (golden set), we observe that 98\% of questions can be answered using their corresponding silver answer. Our fine-tuned Answer Paragraph Selection (APS) model outperforms the baselines. The APS model attained an accuracy of 80\% and 72\%, as well as a macro F1 of 72\% and 66\%, on the MuNfQuAD testset and the golden set, respectively. Furthermore, the APS model effectively generalizes certain a language within the golden set, even after being fine-tuned on silver labels.