Interactions with billion-scale large language models typically yield long-form responses due to their extensive parametric capacities, along with retrieval-augmented features. While detailed responses provide insightful viewpoint of a specific subject, they frequently generate redundant and less engaging content that does not meet user interests. In this work, we focus on the role of query outlining (i.e., selected sequence of queries) in scenarios that users request a specific range of information, namely coverage-conditioned ($C^2$) scenarios. For simulating $C^2$ scenarios, we construct QTree, 10K sets of information-seeking queries decomposed with various perspectives on certain topics. By utilizing QTree, we train QPlanner, a 7B language model generating customized query outlines that follow coverage-conditioned queries. We analyze the effectiveness of generated outlines through automatic and human evaluation, targeting on retrieval-augmented generation (RAG). Moreover, the experimental results demonstrate that QPlanner with alignment training can further provide outlines satisfying diverse user interests. Our resources are available at https://github.com/youngerous/qtree.

Recent years have witnessed significant progress in generative models for music, featuring diverse architectures that balance output quality, diversity, speed, and user control. This study explores a user-friendly graphical interface enabling the drawing of masked regions for inpainting by an Hourglass Diffusion Transformer (HDiT) model trained on MIDI piano roll images. To enhance note generation in specified areas, masked regions can be "repainted" with extra noise. The non-latent HDiTs linear scaling with pixel count allows efficient generation in pixel space, providing intuitive and interpretable controls such as masking throughout the network and removing the need to operate in compressed latent spaces such as those provided by pretrained autoencoders. We demonstrate that, in addition to inpainting of melodies, accompaniment, and continuations, the use of repainting can help increase note density yielding musical structures closely matching user specifications such as rising, falling, or diverging melody and/or accompaniment, even when these lie outside the typical training data distribution. We achieve performance on par with prior results while operating at longer context windows, with no autoencoder, and can enable complex geometries for inpainting masks, increasing the options for machine-assisted composers to control the generated music.

Large Language Models (LLMs) have significantly advanced various NLP tasks. However, these models often risk generating unsafe text that perpetuates biases. Current approaches to produce unbiased outputs from LLMs can reduce biases but at the expense of knowledge retention. In this research, we address the question of whether producing safe (unbiased) outputs through LLMs can retain knowledge and language understanding. In response, we developed the Safety and Responsible Large Language Model (\textbf{SR}$_{\text{LLM}}$), an LLM that has been instruction fine-tuned on top of already safe LLMs (e.g., Llama2 or related) to diminish biases in generated text. To achieve our goals, we compiled a specialized dataset designed to train our model in identifying and correcting biased text. We conduct experiments, both on this custom data and out-of-distribution test sets, to show the bias reduction and knowledge retention. The results confirm that \textbf{SR}$_{\text{LLM}}$ outperforms traditional fine-tuning and prompting methods in both reducing biases and preserving the integrity of language knowledge. The significance of our findings lies in demonstrating that instruction fine-tuning can provide a more robust solution for bias reduction in LLMs. We have made our code and data available at \href{https://github.com/shainarazavi/Safe-Responsible-LLM}{Safe-LLM}.

In traditional RAG framework, the basic retrieval units are normally short. The common retrievers like DPR normally work with 100-word Wikipedia paragraphs. Such a design forces the retriever to search over a large corpus to find the `needle' unit. In contrast, the readers only need to extract answers from the short retrieved units. Such an imbalanced `heavy' retriever and `light' reader design can lead to sub-optimal performance. In order to alleviate the imbalance, we propose a new framework LongRAG, consisting of a `long retriever' and a `long reader'. LongRAG processes the entire Wikipedia into 4K-token units, which is 30x longer than before. By increasing the unit size, we significantly reduce the total units from 22M to 700K. This significantly lowers the burden of retriever, which leads to a remarkable retrieval score: answer recall@1=71% on NQ (previously 52%) and answer recall@2=72% (previously 47%) on HotpotQA (full-wiki). Then we feed the top-k retrieved units ($\approx$ 30K tokens) to an existing long-context LLM to perform zero-shot answer extraction. Without requiring any training, LongRAG achieves an EM of 62.7% on NQ, which is the best known result. LongRAG also achieves 64.3% on HotpotQA (full-wiki), which is on par of the SoTA model. Our study offers insights into the future roadmap for combining RAG with long-context LLMs.

Generating fine-grained video descriptions is a fundamental challenge in video understanding. In this work, we introduce Tarsier, a family of large-scale video-language models designed to generate high-quality video descriptions. Tarsier employs CLIP-ViT to encode frames separately and then uses an LLM to model temporal relationships. Despite its simple architecture, we demonstrate that with a meticulously designed two-stage training procedure, the Tarsier models exhibit substantially stronger video description capabilities than any existing open-source model, showing a $+51.4\%$ advantage in human side-by-side evaluation over the strongest model. Additionally, they are comparable to state-of-the-art proprietary models, with a $+12.3\%$ advantage against GPT-4V and a $-6.7\%$ disadvantage against Gemini 1.5 Pro. Besides video description, Tarsier proves to be a versatile generalist model, achieving new state-of-the-art results across nine public benchmarks, including multi-choice VQA, open-ended VQA, and zero-shot video captioning. Our second contribution is the introduction of a new benchmark for evaluating video description models, consisting of a new challenging dataset featuring videos from diverse sources and varying complexity, along with an automatic method specifically designed to assess the quality of fine-grained video descriptions. We make our models and evaluation benchmark publicly available at \url{https://github.com/bytedance/tarsier}.

The goal of multi-object tracking is to detect and track all objects in a scene while maintaining unique identifiers for each, by associating their bounding boxes across video frames. This association relies on matching motion and appearance patterns of detected objects. This task is especially hard in case of scenarios involving dynamic and non-linear motion patterns. In this paper, we introduce DeepMoveSORT, a novel, carefully engineered multi-object tracker designed specifically for such scenarios. In addition to standard methods of appearance-based association, we improve motion-based association by employing deep learnable filters (instead of the most commonly used Kalman filter) and a rich set of newly proposed heuristics. Our improvements to motion-based association methods are severalfold. First, we propose a new transformer-based filter architecture, TransFilter, which uses an object's motion history for both motion prediction and noise filtering. We further enhance the filter's performance by careful handling of its motion history and accounting for camera motion. Second, we propose a set of heuristics that exploit cues from the position, shape, and confidence of detected bounding boxes to improve association performance. Our experimental evaluation demonstrates that DeepMoveSORT outperforms existing trackers in scenarios featuring non-linear motion, surpassing state-of-the-art results on three such datasets. We also perform a thorough ablation study to evaluate the contributions of different tracker components which we proposed. Based on our study, we conclude that using a learnable filter instead of the Kalman filter, along with appearance-based association is key to achieving strong general tracking performance.

Transformers exhibit impressive capabilities but are often regarded as black boxes due to challenges in understanding the complex nonlinear relationships between features. Interpreting machine learning models is of paramount importance to mitigate risks, and mechanistic interpretability is in particular of current interest as it opens up a window for guiding manual modifications and reverse-engineering solutions. In this work, we introduce contextual decomposition for transformers (CD-T), extending a prior work on CD for RNNs and CNNs, to address mechanistic interpretation computationally efficiently. CD-T is a flexible interpretation method for transformers. It can capture contributions of combinations of input features or source internal components (e.g. attention heads, feed-forward networks) to (1) final predictions or (2) the output of any target internal component. Using CD-T, we propose a novel algorithm for circuit discovery. On a real-world pathology report classification task: we show CD-T distills a more faithful circuit of attention heads with improved computational efficiency (speed up 2x) than a prior benchmark, path patching. As a versatile interpretation method, CD-T also exhibits exceptional capabilities for local interpretations. CD-T is shown to reliably find words and phrases of contrasting sentiment/topic on SST-2 and AGNews datasets. Through human experiments, we demonstrate CD-T enables users to identify the more accurate of two models and to better trust a model's outputs compared to alternative interpretation methods such as SHAP and LIME.

Heterogeneous graphs can well describe the complex entity relationships in the real world. For example, online shopping networks contain multiple physical types of consumers and products, as well as multiple relationship types such as purchasing and favoriting. More and more scholars pay attention to this research because heterogeneous graph representation learning shows strong application potential in real-world scenarios. However, the existing heterogeneous graph models use data augmentation techniques to enhance the use of graph structure information, which only captures the graph structure information from the spatial topology, ignoring the information displayed in the spectrum dimension of the graph structure. To address the issue that heterogeneous graph representation learning methods fail to model spectral information, this paper introduces a spectral-enhanced graph contrastive learning model (SHCL) and proposes a spectral augmentation algorithm for the first time in heterogeneous graph neural networks. The proposed model learns an adaptive topology augmentation scheme through the heterogeneous graph itself, disrupting the structural information of the heterogeneous graph in the spectrum dimension, and ultimately improving the learning effect of the model. Experimental results on multiple real-world datasets demonstrate substantial advantages of the proposed model.

Advancements in DeepFake (DF) audio models pose a significant threat to voice authentication systems, leading to unauthorized access and the spread of misinformation. We introduce a defense mechanism, SecureSpectra, addressing DF threats by embedding orthogonal, irreversible signatures within audio. SecureSpectra leverages the inability of DF models to replicate high-frequency content, which we empirically identify across diverse datasets and DF models. Integrating differential privacy into the pipeline protects signatures from reverse engineering and strikes a delicate balance between enhanced security and minimal performance compromises. Our evaluations on Mozilla Common Voice, LibriSpeech, and VoxCeleb datasets showcase SecureSpectra's superior performance, outperforming recent works by up to 71% in detection accuracy. We open-source SecureSpectra to benefit the research community.

Audio analysis is useful in many application scenarios. The state-of-the-art audio analysis approaches assume that the data distribution at training and deployment time will be the same. However, due to various real-life environmental factors, the data may encounter drift in its distribution or can encounter new classes in the late future. Thus, a one-time trained model might not perform adequately. In this paper, we characterize continual learning (CL) approaches in audio analysis. In this paper, we characterize continual learning (CL) approaches, intended to tackle catastrophic forgetting arising due to drifts. As there is no CL dataset for audio analysis, we use DCASE 2020 to 2023 datasets to create various CL scenarios for audio-based monitoring tasks. We have investigated the following CL and non-CL approaches: EWC, LwF, SI, GEM, A-GEM, GDumb, Replay, Naive, cumulative, and joint training. The study is very beneficial for researchers and practitioners working in the area of audio analysis for developing adaptive models. We observed that Replay achieved better results than other methods in the DCASE challenge data. It achieved an accuracy of 70.12% for the domain incremental scenario and an accuracy of 96.98% for the class incremental scenario.