Large language models (LLMs) pose significant risks due to the potential for generating harmful content or users attempting to evade guardrails. Existing studies have developed LLM-based guard models designed to moderate the input and output of threat LLMs, ensuring adherence to safety policies by blocking content that violates these protocols upon deployment. However, limited attention has been given to the reliability and calibration of such guard models. In this work, we empirically conduct comprehensive investigations of confidence calibration for 9 existing LLM-based guard models on 12 benchmarks in both user input and model output classification. Our findings reveal that current LLM-based guard models tend to 1) produce overconfident predictions, 2) exhibit significant miscalibration when subjected to jailbreak attacks, and 3) demonstrate limited robustness to the outputs generated by different types of response models. Additionally, we assess the effectiveness of post-hoc calibration methods to mitigate miscalibration. We demonstrate the efficacy of temperature scaling and, for the first time, highlight the benefits of contextual calibration for confidence calibration of guard models, particularly in the absence of validation sets. Our analysis and experiments underscore the limitations of current LLM-based guard models and provide valuable insights for the future development of well-calibrated guard models toward more reliable content moderation. We also advocate for incorporating reliability evaluation of confidence calibration when releasing future LLM-based guard models.

Auto-regressive language Models (LMs) assign significant attention to the first token, even if it is not semantically important, which is known as **attention sink**. This phenomenon has been widely adopted in applications such as streaming/long context generation, KV cache optimization, inference acceleration, model quantization, and others. Despite its widespread use, a deep understanding of attention sink in LMs is still lacking. In this work, we first demonstrate that attention sinks exist universally in auto-regressive LMs with various inputs, even in small models. Furthermore, attention sink is observed to emerge during the LM pre-training, motivating us to investigate how *optimization*, *data distribution*, *loss function*, and *model architecture* in LM pre-training influence its emergence. We highlight that attention sink emerges after effective optimization on sufficient training data. The sink position is highly correlated with the loss function and data distribution. Most importantly, we find that attention sink acts more like key biases, *storing extra attention scores*, which could be non-informative and not contribute to the value computation. We also observe that this phenomenon (at least partially) stems from tokens' inner dependence on attention scores as a result of softmax normalization. After relaxing such dependence by replacing softmax attention with other attention operations, such as sigmoid attention without normalization, attention sinks do not emerge in LMs up to 1B parameters. The code is available at https://github.com/sail-sg/Attention-Sink.

Recent advancements in LLM pretraining have featured ever-expanding context windows to process longer sequences. However, our controlled study reveals that models pretrained with shorter context windows consistently outperform their long-context counterparts under a fixed token budget. This finding motivates us to explore an optimal context window scheduling strategy to better balance long-context capability with pretraining efficiency. To this end, we propose SkyLadder, a simple yet effective approach that implements a short-to-long context window transition. SkyLadder preserves strong standard benchmark performance, while matching or exceeding baseline results on long-context tasks. Through extensive experiments, we pretrain 1B-parameter models (up to 32K context) and 3B-parameter models (8K context) on 100B tokens, demonstrating that SkyLadder yields consistent gains of up to 3.7% on common benchmarks, while achieving up to 22% faster training speeds compared to baselines.

A multimodal large language model (MLLM) agent can receive instructions, capture images, retrieve histories from memory, and decide which tools to use. Nonetheless, red-teaming efforts have revealed that adversarial images/prompts can jailbreak an MLLM and cause unaligned behaviors. In this work, we report an even more severe safety issue in multi-agent environments, referred to as infectious jailbreak. It entails the adversary simply jailbreaking a single agent, and without any further intervention from the adversary, (almost) all agents will become infected exponentially fast and exhibit harmful behaviors. To validate the feasibility of infectious jailbreak, we simulate multi-agent environments containing up to one million LLaVA-1.5 agents, and employ randomized pair-wise chat as a proof-of-concept instantiation for multi-agent interaction. Our results show that feeding an (infectious) adversarial image into the memory of any randomly chosen agent is sufficient to achieve infectious jailbreak. Finally, we derive a simple principle for determining whether a defense mechanism can provably restrain the spread of infectious jailbreak, but how to design a practical defense that meets this principle remains an open question to investigate.

Automatic lyric transcription (ALT) refers to transcribing singing voices into lyrics, while automatic music transcription (AMT) refers to transcribing singing voices into note events, i.e., musical MIDI notes. Despite these two tasks having significant potential for practical application, they are still nascent. This is because the transcription of lyrics and note events solely from singing audio is notoriously difficult due to the presence of noise contamination, e.g., musical accompaniment, resulting in a degradation of both the intelligibility of sung lyrics and the recognizability of sung notes. To address this challenge, we propose a general framework for implementing multimodal ALT and AMT systems. Additionally, we curate the first multimodal singing dataset, comprising N20EMv1 and N20EMv2, which encompasses audio recordings and videos of lip movements, together with ground truth for lyrics and note events. For model construction, we propose adapting self-supervised learning models from the speech domain as acoustic encoders and visual encoders to alleviate the scarcity of labeled data. We also introduce a residual cross-attention mechanism to effectively integrate features from the audio and video modalities. Through extensive experiments, we demonstrate that our single-modal systems exhibit state-of-the-art performance on both ALT and AMT tasks. Subsequently, through single-modal experiments, we also explore the individual contributions of each modality to the multimodal system. Finally, we combine these and demonstrate the effectiveness of our proposed multimodal systems, particularly in terms of their noise robustness.

It is widely known that males and females typically possess different sound characteristics when singing, such as timbre and pitch, but it has never been explored whether these gender-based characteristics lead to a performance disparity in singing voice transcription (SVT), whose target includes pitch. Such a disparity could cause fairness issues and severely affect the user experience of downstream SVT applications. Motivated by this, we first demonstrate the female superiority of SVT systems, which is observed across different models and datasets. We find that different pitch distributions, rather than gender data imbalance, contribute to this disparity. To address this issue, we propose using an attribute predictor to predict gender labels and adversarially training the SVT system to enforce the gender-invariance of acoustic representations. Leveraging the prior knowledge that pitch distributions may contribute to the gender bias, we propose conditionally aligning acoustic representations between demographic groups by feeding note events to the attribute predictor. Empirical experiments on multiple benchmark SVT datasets show that our method significantly reduces gender bias (up to more than 50%) with negligible degradation of overall SVT performance, on both in-domain and out-of-domain singing data, thus offering a better fairness-utility trade-off.

Phonation mode detection predicts phonation modes and their temporal boundaries in singing and speech, holding promise for characterizing voice quality and vocal health. However, it is very challenging due to the domain disparities between training data and unannotated real-world recordings. To tackle this problem, we develop a disentangled adversarial domain adaptation network, which adapts the phonation mode detection model with the structure of the convolutional recurrent neural network pre-trained on the source domain to the target domain without phonation mode labels. Based on our curated sung and spoken dataset for phonation mode detection, we demonstrate that the subject and the singing-speech mismatches cause performance decline. By disentangling domain-invariant phonation mode and domain-specific embeddings, our method greatly enhances the effectiveness and explainability of unsupervised adversarial domain adaptation. Experiments show that the performance drop caused by the subject mismatch is alleviated via adaptation, resulting in 44.7% and 6.8% improvement of the F-score for singing and speech, respectively. The singing and speech domain adaptation experiment indicates that a model trained on singing data can be adapted to speech, yielding an F-score of 0.56, commensurate with the F-score of 0.59 achieved using a model exclusively trained on speech data. By further investigating the disentangled embeddings, we find that the phonation mode feature shared by singing and speech is invariant to pitch. These results inspire reliable and versatile applications in voice quality evaluation and paralinguistic information retrieval.

Automatic lyric transcription (ALT) is a nascent field of study attracting increasing interest from both the speech and music information retrieval communities, given its significant application potential. However, ALT with audio data alone is a notoriously difficult task due to instrumental accompaniment and musical constraints resulting in degradation of both the phonetic cues and the intelligibility of sung lyrics. To tackle this challenge, we propose the MultiModal Automatic Lyric Transcription system (MM-ALT), together with a new dataset, N20EM, which consists of audio recordings, videos of lip movements, and inertial measurement unit (IMU) data of an earbud worn by the performing singer. We first adapt the wav2vec 2.0 framework from automatic speech recognition (ASR) to the ALT task. We then propose a video-based ALT method and an IMU-based voice activity detection (VAD) method. In addition, we put forward the Residual Cross Attention (RCA) mechanism to fuse data from the three modalities (i.e., audio, video, and IMU). Experiments show the effectiveness of our proposed MM-ALT system, especially in terms of noise robustness. Project page is at https://n20em.github.io.

Human intelligence has shown remarkably lower latency and higher precision than most AI systems when processing non-stationary streaming data in real-time. Numerous neuroscience studies suggest that such abilities may be driven by internal predictive modeling. In this paper, we explore the possibility of introducing such a mechanism in unsupervised domain adaptation (UDA) for handling non-stationary streaming data for real-time streaming applications. We propose to formulate internal predictive modeling as a continuous-time Bayesian filtering problem within a stochastic dynamical system context. Such a dynamical system describes the dynamics of model parameters of a UDA model evolving with non-stationary streaming data. Building on such a dynamical system, we then develop extrapolative continuous-time Bayesian neural networks (ECBNN), which generalize existing Bayesian neural networks to represent temporal dynamics and allow us to extrapolate the distribution of model parameters before observing the incoming data, therefore effectively reducing the latency. Remarkably, our empirical results show that ECBNN is capable of continuously generating better distributions of model parameters along the time axis given historical data only, thereby achieving (1) training-free test-time adaptation with low latency, (2) gradually improved alignment between the source and target features and (3) gradually improved model performance over time during the real-time testing stage.

Monocular 3D human pose estimation is challenging due to depth ambiguity. Convolution-based and Graph-Convolution-based methods have been developed to extract 3D information from temporal cues in motion videos. Typically, in the lifting-based methods, most recent works adopt the transformer to model the temporal relationship of 2D keypoint sequences. These previous works usually consider all the joints of a skeleton as a whole and then calculate the temporal attention based on the overall characteristics of the skeleton. Nevertheless, the human skeleton exhibits obvious part-wise inconsistency of motion patterns. It is therefore more appropriate to consider each part's temporal behaviors separately. To deal with such part-wise motion inconsistency, we propose the Part Aware Temporal Attention module to extract the temporal dependency of each part separately. Moreover, the conventional attention mechanism in 3D pose estimation usually calculates attention within a short time interval. This indicates that only the correlation within the temporal context is considered. Whereas, we find that the part-wise structure of the human skeleton is repeating across different periods, actions, and even subjects. Therefore, the part-wise correlation at a distance can be utilized to further boost 3D pose estimation. We thus propose the Part Aware Dictionary Attention module to calculate the attention for the part-wise features of input in a dictionary, which contains multiple 3D skeletons sampled from the training set. Extensive experimental results show that our proposed part aware attention mechanism helps a transformer-based model to achieve state-of-the-art 3D pose estimation performance on two widely used public datasets. The codes and the trained models are released at https://github.com/thuxyz19/3D-HPE-PAA.