This paper introduces M2M Gen, a multi modal framework for generating background music tailored to Japanese manga. The key challenges in this task are the lack of an available dataset or a baseline. To address these challenges, we propose an automated music generation pipeline that produces background music for an input manga book. Initially, we use the dialogues in a manga to detect scene boundaries and perform emotion classification using the characters faces within a scene. Then, we use GPT4o to translate this low level scene information into a high level music directive. Conditioned on the scene information and the music directive, another instance of GPT 4o generates page level music captions to guide a text to music model. This produces music that is aligned with the mangas evolving narrative. The effectiveness of M2M Gen is confirmed through extensive subjective evaluations, showcasing its capability to generate higher quality, more relevant and consistent music that complements specific scenes when compared to our baselines.

Fine-tuning Large Language Models (LLMs) has emerged as a common practice for tailoring models to individual needs and preferences. The choice of datasets for fine-tuning can be diverse, introducing safety concerns regarding the potential inclusion of harmful data samples. Manually filtering or avoiding such samples, however, can be labor-intensive and subjective. To address these difficulties, we propose a novel Safety-Aware Fine-Tuning (SAFT) framework designed to automatically detect and remove potentially harmful data, by leveraging a scoring function that exploits the subspace information of harmful and benign samples. Experimental results demonstrate the efficacy of SAFT across different LLMs and varying contamination rates, achieving reductions in harmfulness of up to 27.8%. Going beyond, we delve into the mechanism of our approach and validate its versatility in addressing practical challenges in real-world scenarios.

The integrity of the peer-review process is vital for maintaining scientific rigor and trust within the academic community. With the steady increase in the usage of large language models (LLMs) like ChatGPT in academic writing, there is a growing concern that AI-generated texts could compromise scientific publishing, including peer-reviews. Previous works have focused on generic AI-generated text detection or have presented an approach for estimating the fraction of peer-reviews that can be AI-generated. Our focus here is to solve a real-world problem by assisting the editor or chair in determining whether a review is written by ChatGPT or not. To address this, we introduce the Term Frequency (TF) model, which posits that AI often repeats tokens, and the Review Regeneration (RR) model, which is based on the idea that ChatGPT generates similar outputs upon re-prompting. We stress test these detectors against token attack and paraphrasing. Finally, we propose an effective defensive strategy to reduce the effect of paraphrasing on our models. Our findings suggest both our proposed methods perform better than the other AI text detectors. Our RR model is more robust, although our TF model performs better than the RR model without any attacks. We make our code, dataset, and model public.

Previous studies on question generation from videos have mostly focused on generating questions about common objects and attributes and hence are not entity-centric. In this work, we focus on the generation of entity-centric information-seeking questions from videos. Such a system could be useful for video-based learning, recommending ``People Also Ask'' questions, video-based chatbots, and fact-checking. Our work addresses three key challenges: identifying question-worthy information, linking it to entities, and effectively utilizing multimodal signals. Further, to the best of our knowledge, there does not exist a large-scale dataset for this task. Most video question generation datasets are on TV shows, movies, or human activities or lack entity-centric information-seeking questions. Hence, we contribute a diverse dataset of YouTube videos, VideoQuestions, consisting of 411 videos with 2265 manually annotated questions. We further propose a model architecture combining Transformers, rich context signals (titles, transcripts, captions, embeddings), and a combination of cross-entropy and contrastive loss function to encourage entity-centric question generation. Our best method yields BLEU, ROUGE, CIDEr, and METEOR scores of 71.3, 78.6, 7.31, and 81.9, respectively, demonstrating practical usability. We make the code and dataset publicly available. https://github.com/thePhukan/ECIS-VQG

Large Language Models (LLMs) are widely adopted across various tasks including Question-Answering (QA) tasks. More and more models including the fine-tuned models and the dataset used for fine-tuning are being released daily. This explosion in the number of models, and datasets emphasizes the importance of accurate automatic evaluation for out-of-the-model language model training as well as gauging their QA capabilities. However, varying question types (short-form, long-form, open-ended, etc.) and ambiguity in the grading rubric make it difficult to properly gauge each model's capability objectively for QA tasks. No single existing evaluation metric can capture the language model's QA answer capability for multiple quality types. For example, Exact Match (EM) is a widely adopted all-or-nothing evaluation metric that shows a high correlation with human-evaluated scores for short-form QA tasks but is too strict to give credit for any semantically identical answer. The lack of an objective grading rubric for varying QA types creates a bias in summary statistics. For example, half credit for an open-ended question is regarded equally as half credit for a simple factual question. In this paper, we (1) deploy statistical approaches to characterize various existing evaluation metrics, (2) the effectiveness of recent ChatGPT-o1-preview model [6] as QA grader, and (3) potential solution, a Mixture Of Grader (MOG), which first classifies each (question, gold answer) pair into one of the predefined QA type class and select the appropriate evaluation metric accordingly for an advanced automatic evaluation that better "correlates" to human evaluator.

In-Context Learning (ICL) emerges as a key feature for Large Language Models (LLMs), allowing them to adapt to new tasks by leveraging task-specific examples without updating model parameters. However, ICL faces challenges with increasing numbers of examples due to performance degradation and quadratic computational costs. In this paper, we propose Logit Arithmetic Reweighting Approach (LARA), a novel framework that enhances ICL by using logit-based ensembling of multiple demonstrations. Our approach divides long input demonstrations into parallelizable shorter inputs to significantly reduce memory requirements, and then effectively aggregate the information by reweighting logits of each group via a non-gradient optimization approach. We further introduce Binary LARA (B-LARA), a variant that constrains weights to binary values to simplify the search space and reduces memory usage by filtering out less informative demonstration groups. Experiments on BBH and MMLU demonstrate that LARA and B-LARA outperform all baseline methods in both accuracy and memory efficiency. We also conduct extensive analysis to show that LARA generalizes well to scenarios of varying numbers of examples from limited to many-shot demonstrations.

Aspect-based sentiment analysis (ABSA) is the challenging task of extracting sentiment along with its corresponding aspects and opinions from human language. Due to the inherent variability of natural language, aspect and opinion terms can be expressed in various surface forms, making their accurate identification complex. Current evaluation methods for this task often restrict answers to a single ground truth, penalizing semantically equivalent predictions that differ in surface form. To address this limitation, we propose a novel, fully automated pipeline that augments existing test sets with alternative valid responses for aspect and opinion terms. This approach enables a fairer assessment of language models by accommodating linguistic diversity, resulting in higher human agreement than single-answer test sets (up to 10%p improvement in Kendall's Tau score). Our experimental results demonstrate that Large Language Models (LLMs) show substantial performance improvements over T5 models when evaluated using our augmented test set, suggesting that LLMs' capabilities in ABSA tasks may have been underestimated. This work contributes to a more comprehensive evaluation framework for ABSA, potentially leading to more accurate assessments of model performance in information extraction tasks, particularly those involving span extraction.

Crafting effective features is a crucial yet labor-intensive and domain-specific task within machine learning pipelines. Fortunately, recent advancements in Large Language Models (LLMs) have shown promise in automating various data science tasks, including feature engineering. But despite this potential, evaluations thus far are primarily based on the end performance of a complete ML pipeline, providing limited insight into precisely how LLMs behave relative to human experts in feature engineering. To address this gap, we propose ELF-Gym, a framework for Evaluating LLM-generated Features. We curated a new dataset from historical Kaggle competitions, including 251 "golden" features used by top-performing teams. ELF-Gym then quantitatively evaluates LLM-generated features by measuring their impact on downstream model performance as well as their alignment with expert-crafted features through semantic and functional similarity assessments. This approach provides a more comprehensive evaluation of disparities between LLMs and human experts, while offering valuable insights into specific areas where LLMs may have room for improvement. For example, using ELF-Gym we empirically demonstrate that, in the best-case scenario, LLMs can semantically capture approximately 56% of the golden features, but at the more demanding implementation level this overlap drops to 13%. Moreover, in other cases LLMs may fail completely, particularly on datasets that require complex features, indicating broad potential pathways for improvement.

Large language models (LLMs) exhibit remarkable performance across diverse tasks, indicating their potential for expansion into large speech-text models (LSMs) by integrating speech capabilities. Although unified speech-text pre-training and multimodal data instruction-tuning offer considerable benefits, these methods generally entail significant resource demands and tend to overfit specific tasks. This study aims to refine the use of speech datasets for LSM training by addressing the limitations of vanilla instruction tuning. We explore the instruction-following dynamics within LSMs, identifying a critical issue termed speech anchor bias-a tendency for LSMs to over-rely on speech inputs, mistakenly interpreting the entire speech modality as directives, thereby neglecting textual instructions. To counteract this bias, we introduce a self-powered LSM that leverages augmented automatic speech recognition data generated by the model itself for more effective instruction tuning. Our experiments across a range of speech-based tasks demonstrate that self-powered LSM mitigates speech anchor bias and improves the fusion of speech and text modalities in LSMs. Data, code and scripts are freely available at https://github.com/ytf-philp/Self-powered-LSM.

Learning from preference feedback is a common practice for aligning large language models~(LLMs) with human value. Conventionally, preference data is learned and encoded into a scalar reward model that connects a value head with an LLM to produce a scalar score as preference or reward. However, scalar models lack interpretability and are known to be susceptible to biases in datasets. This paper investigates leveraging the generation capability of LLMs to address both limitations in one shot. Specifically, we prompt the pre-trained LLM to generate positive and negative judgments, both supported with rationales in natural language form. The self-generated contrastive judgment pairs are used to train the generative judge with Direct Preference Optimization (DPO). This proposal of training the generative Judge using self-generated Contrastive judgments (Con-J) ensures natural interpretability due to the generated rationales together with the judgments, as well as high robustness against bias without the need for an additional reward head. Experimental results show that the performance of Con-J is comparable to the scalar reward model trained on the same collection of preference data, and demonstrate its superior interpretability and robustness in encoding human preferences.