Neural audio codecs have been widely adopted in audio-generative tasks because their compact and discrete representations are suitable for both large-language-model-style and regression-based generative models. However, most neural codecs struggle to model out-of-domain audio, resulting in error propagations to downstream generative tasks. In this paper, we first argue that information loss from codec compression degrades out-of-domain robustness. Then, we propose full-band 48~kHz ComplexDec with complex spectral input and output to ease the information loss while adopting the same 24~kbps bitrate as the baseline AuidoDec and ScoreDec. Objective and subjective evaluations demonstrate the out-of-domain robustness of ComplexDec trained using only the 30-hour VCTK corpus.
Large language models have achieved remarkable success in various tasks but suffer from high computational costs during inference, limiting their deployment in resource-constrained applications. To address this issue, we propose a novel Collaborative Inference with Token-lEvel Routing (CITER) framework that enables efficient collaboration between small and large language models (SLMs \& LLMs) through a token-level routing strategy. Specifically, CITER routes non-critical tokens to an SLM for efficiency and routes critical tokens to an LLM for generalization quality. We formulate router training as a policy optimization, where the router receives rewards based on both the quality of predictions and the inference costs of generation. This allows the router to learn to predict token-level routing scores and make routing decisions based on both the current token and the future impact of its decisions. To further accelerate the reward evaluation process, we introduce a shortcut which significantly reduces the costs of the reward estimation and improving the practicality of our approach. Extensive experiments on five benchmark datasets demonstrate that CITER reduces the inference costs while preserving high-quality generation, offering a promising solution for real-time and resource-constrained applications. Our data and code are available at https://github.com/aiming-lab/CITER.
The rise of Agent AI and Large Language Model-powered Multi-Agent Systems (LLM-MAS) has underscored the need for responsible and dependable system operation. Tools like LangChain and Retrieval-Augmented Generation have expanded LLM capabilities, enabling deeper integration into MAS through enhanced knowledge retrieval and reasoning. However, these advancements introduce critical challenges: LLM agents exhibit inherent unpredictability, and uncertainties in their outputs can compound across interactions, threatening system stability. To address these risks, a human-centered design approach with active dynamic moderation is essential. Such an approach enhances traditional passive oversight by facilitating coherent inter-agent communication and effective system governance, allowing MAS to achieve desired outcomes more efficiently.
As large language models (LLMs) continue to scale, their inference demands present significant challenges, particularly due to the high power consumption of AI accelerators in datacenters. These facilities require specialized cooling and power management systems, substantially increasing the total cost of ownership (TCO) for cloud service providers (CSPs). In this work, we analyze the computational characteristics and constraints of LLM inference from a TCO perspective, focusing on two representative accelerators: the Gaudi 2 and NVIDIA H100. We present a generalizable framework that enables CSPs to compare and select AI accelerators according to diverse operational requirements. Using this model, we analyze the impact of FP8 precision and LLM inference workload characteristics as key factors influencing TCO. We investigate FP8 quantization, which is gaining adoption in LLM training, as a technique to improve inference throughput while maintaining cost efficiency. Furthermore, our analysis of LLM inference workloads reveals that performance on thin GEMMs, which dominate the decode phase, can have a greater impact than theoretical hardware peak performance. By studying the interaction between power consumption, quantization strategies, and hardware architecture, we offer insights that support informed deployment decisions and guide future accelerator designs to improve the TCO of LLM inference.
Large Language Models (LLMs) have significantly advanced natural language processing applications, yet their widespread use raises concerns regarding inherent biases that may reduce utility or harm for particular social groups. Despite the advancement in addressing LLM bias, existing research has two major limitations. First, existing LLM bias evaluation focuses on the U.S. cultural context, making it challenging to reveal stereotypical biases of LLMs toward other cultures, leading to unfair development and use of LLMs. Second, current bias evaluation often assumes models are familiar with the target social groups. When LLMs encounter words beyond their knowledge boundaries that are unfamiliar in their training data, they produce irrelevant results in the local context due to hallucinations and overconfidence, which are not necessarily indicative of inherent bias. This research addresses these limitations with a Local Integrated Bias Recognition and Assessment Framework (LIBRA) for measuring bias using datasets sourced from local corpora without crowdsourcing. Implementing this framework, we develop a dataset comprising over 360,000 test cases in the New Zealand context. Furthermore, we propose the Enhanced Idealized CAT Score (EiCAT), integrating the iCAT score with a beyond knowledge boundary score (bbs) and a distribution divergence-based bias measurement to tackle the challenge of LLMs encountering words beyond knowledge boundaries. Our results show that the BERT family, GPT-2, and Llama-3 models seldom understand local words in different contexts. While Llama-3 exhibits larger bias, it responds better to different cultural contexts. The code and dataset are available at: https://github.com/ipangbo/LIBRA.
We present a theoretical framework showing that popular LLM alignment methods, including RLHF and its variants, can be understood as divergence estimators between aligned (safe or preferred) and unaligned (harmful or less preferred) distributions. This perspective explains the emergence of separation in the latent space between safe and harmful prompts after alignment. As an application of our general divergence framework, we propose KLDO, a novel KL divergence-based alignment method, and empirically validate its effectiveness. We further show that using compliance-refusal datasets, rather than standard preference-based datasets, leads to stronger separation and improved safety alignment. Finally, to quantify the separation effect, we propose a distance-based metric in the prompt representation space, which also acts as a statistically significant indicator for model safety.
Token filtering has been proposed to enhance utility of large language models (LLMs) by eliminating inconsequential tokens during training. While using fewer tokens should reduce computational workloads, existing studies have not succeeded in achieving higher efficiency. This is primarily due to the insufficient sparsity caused by filtering tokens only in the output layers, as well as inefficient sparse GEMM (General Matrix Multiplication), even when having sufficient sparsity. This paper presents Collider, a system unleashing the full efficiency of token filtering in LLM training. At its core, Collider filters activations of inconsequential tokens across all layers to maintain sparsity. Additionally, it features an automatic workflow that transforms sparse GEMM into dimension-reduced dense GEMM for optimized efficiency. Evaluations on three LLMs-TinyLlama-1.1B, Qwen2.5-1.5B, and Phi1.5-1.4B-demonstrate that Collider reduces backpropagation time by up to 35.1% and end-to-end training time by up to 22.0% when filtering 40% of tokens. Utility assessments of training TinyLlama on 15B tokens indicate that Collider sustains the utility advancements of token filtering by relatively improving model utility by 16.3% comparing to regular training, and reduces training time from 4.7 days to 3.5 days using 8 GPUs. Collider is designed for easy integration into existing LLM training frameworks, allowing systems already using token filtering to accelerate training with just one line of code.
The pervasiveness of the dissemination of fake news through social media platforms poses critical risks to the trust of the general public, societal stability, and democratic institutions. This challenge calls for novel methodologies in detection, which can keep pace with the dynamic and multi-modal nature of misinformation. Recent works include powering the detection using large language model advances in multimodal frameworks, methodologies using graphs, and adversarial training in the literature of fake news. Based on the different approaches which can bring success, some key highlights will be underlined: enhanced LLM-improves accuracy through more advanced semantics and cross-modality fusion for robust detections. The review further identifies critical gaps in adaptability to dynamic social media trends, real-time, and cross-platform detection capabilities, as well as the ethical challenges thrown up by the misuse of LLMs. Future directions underline the development of style-agnostic models, cross-lingual detection frameworks, and robust policies with a view to mitigating LLM-driven misinformation. This synthesis thus lays a concrete foundation for those researchers and practitioners committed to reinforcing fake news detection systems with complications that keep on growing in the digital landscape.
This study examines the potential of large language models (LLMs) to augment the academic peer review process by reliably evaluating the quality of economics research without introducing systematic bias. We conduct one of the first large-scale experimental assessments of four LLMs (GPT-4o, Claude 3.5, Gemma 3, and LLaMA 3.3) across two complementary experiments. In the first, we use nonparametric binscatter and linear regression techniques to analyze over 29,000 evaluations of 1,220 anonymized papers drawn from 110 economics journals excluded from the training data of current LLMs, along with a set of AI-generated submissions. The results show that LLMs consistently distinguish between higher- and lower-quality research based solely on textual content, producing quality gradients that closely align with established journal prestige measures. Claude and Gemma perform exceptionally well in capturing these gradients, while GPT excels in detecting AI-generated content. The second experiment comprises 8,910 evaluations designed to assess whether LLMs replicate human like biases in single blind reviews. By systematically varying author gender, institutional affiliation, and academic prominence across 330 papers, we find that GPT, Gemma, and LLaMA assign significantly higher ratings to submissions from top male authors and elite institutions relative to the same papers presented anonymously. These results emphasize the importance of excluding author-identifying information when deploying LLMs in editorial screening. Overall, our findings provide compelling evidence and practical guidance for integrating LLMs into peer review to enhance efficiency, improve accuracy, and promote equity in the publication process of economics research.
We present EditIQ, a completely automated framework for cinematically editing scenes captured via a stationary, large field-of-view and high-resolution camera. From the static camera feed, EditIQ initially generates multiple virtual feeds, emulating a team of cameramen. These virtual camera shots termed rushes are subsequently assembled using an automated editing algorithm, whose objective is to present the viewer with the most vivid scene content. To understand key scene elements and guide the editing process, we employ a two-pronged approach: (1) a large language model (LLM)-based dialogue understanding module to analyze conversational flow, coupled with (2) visual saliency prediction to identify meaningful scene elements and camera shots therefrom. We then formulate cinematic video editing as an energy minimization problem over shot selection, where cinematic constraints determine shot choices, transitions, and continuity. EditIQ synthesizes an aesthetically and visually compelling representation of the original narrative while maintaining cinematic coherence and a smooth viewing experience. Efficacy of EditIQ against competing baselines is demonstrated via a psychophysical study involving twenty participants on the BBC Old School dataset plus eleven theatre performance videos. Video samples from EditIQ can be found at https://editiq-ave.github.io/.
Reinforcement Learning from Human Feedback (RLHF) has emerged as a pivotal technique for large language model (LLM) alignment. This paper studies the setting of online RLHF and focus on improving sample efficiency. All existing algorithms in online RLHF, whether doing passive exploration or active exploration, suffer from a sample complexity that scales exponentially with the scale of the reward function. This fundamental limitation hinders their effectiveness in scenarios with heavily skewed preferences, e.g. questions with a unique correct solution. To address this, we introduce Self-Exploring Preference-Incentive Online Preference Optimization (SE-POPO), an online RLHF algorithm that for the first time achieves a sample complexity that scales polynomially with the reward scale, answering an open problem raised by Xie et al. (2024).. Theoretically, we demonstrate that the sample complexity of SE-POPO dominates that of existing exploration algorithms. Empirically, our systematic evaluation confirms that SE-POPO is more sample-efficient than both exploratory and non-exploratory baselines, in two primary application scenarios of RLHF as well as on public benchmarks, marking a significant step forward in RLHF algorithm design.
Recently, violence detection systems developed using unified multimodal models have achieved significant success and attracted widespread attention. However, most of these systems face two critical challenges: the lack of interpretability as black-box models and limited functionality, offering only classification or retrieval capabilities. To address these challenges, this paper proposes a novel interpretable violence detection system, termed the Three-in-One (TIO) System. The TIO system integrates knowledge graphs (KG) and graph attention networks (GAT) to provide three core functionalities: detection, retrieval, and explanation. Specifically, the system processes each video frame along with text descriptions generated by a large language model (LLM) for videos containing potential violent behavior. It employs ImageBind to generate high-dimensional embeddings for constructing a knowledge graph, uses GAT for reasoning, and applies lightweight time series modules to extract video embedding features. The final step connects a classifier and retriever for multi-functional outputs. The interpretability of KG enables the system to verify the reasoning process behind each output. Additionally, the paper introduces several lightweight methods to reduce the resource consumption of the TIO system and enhance its efficiency. Extensive experiments conducted on the XD-Violence and UCF-Crime datasets validate the effectiveness of the proposed system. A case study further reveals an intriguing phenomenon: as the number of bystanders increases, the occurrence of violent behavior tends to decrease.
Large pretrained models can be used as annotators, helping replace or augment crowdworkers and enabling distilling generalist models into smaller specialist models. Unfortunately, this comes at a cost: employing top-of-the-line models often requires paying thousands of dollars for API calls, while the resulting datasets are static and challenging to audit. To address these challenges, we propose a simple alternative: rather than directly querying labels from pretrained models, we task models to generate programs that can produce labels. These programs can be stored and applied locally, re-used and extended, and cost orders of magnitude less. Our system, Alchemist, obtains comparable to or better performance than large language model-based annotation in a range of tasks for a fraction of the cost: on average, improvements amount to a 12.9% enhancement while the total labeling costs across all datasets are reduced by a factor of approximately 500x.
This paper examines the output of cultural items generated by Chat Generative PreTrained Transformer Pro in response to three structured prompts to translate three anthologies of African poetry. The first prompt was broad, the second focused on poetic structure, and the third prompt emphasized cultural specificity. To support this analysis, four comparative tables were created. The first table presents the results of the cultural items produced after the three prompts, the second categorizes these outputs based on Aixela framework of Proper nouns and Common expressions, the third table summarizes the cultural items generated by human translators, a custom translation engine, and a Large Language Model. The final table outlines the strategies employed by Chat Generative PreTrained Transformer Pro following the culture specific prompt. Compared to the outputs of cultural items from reference human translation and the custom translation engine in prior studies the findings indicate that the culture oriented prompts used with Chat Generative PreTrained Transformer Pro did not yield significant enhancements of cultural items during the translation of African poetry from English to French. Among the fifty four cultural items, the human translation produced thirty three cultural items in repetition, the custom translation engine generated Thirty eight cultural items in repetition while Chat Generative PreTrained Transformer Pro produced forty one cultural items in repetition. The untranslated cultural items revealed inconsistencies in Large language models approach to translating cultural items in African poetry from English to French.
Large language models (LLMs) are increasingly utilized in healthcare applications. However, their deployment in clinical practice raises significant safety concerns, including the potential spread of harmful information. This study systematically assesses the vulnerabilities of seven LLMs to three advanced black-box jailbreaking techniques within medical contexts. To quantify the effectiveness of these techniques, we propose an automated and domain-adapted agentic evaluation pipeline. Experiment results indicate that leading commercial and open-source LLMs are highly vulnerable to medical jailbreaking attacks. To bolster model safety and reliability, we further investigate the effectiveness of Continual Fine-Tuning (CFT) in defending against medical adversarial attacks. Our findings underscore the necessity for evolving attack methods evaluation, domain-specific safety alignment, and LLM safety-utility balancing. This research offers actionable insights for advancing the safety and reliability of AI clinicians, contributing to ethical and effective AI deployment in healthcare.
In this paper, we show that current state-of-the-art LLM unlearning methods inherently reduce models' robustness, causing them to misbehave even when a single non-adversarial forget-token is in the retain-query. Toward understanding underlying causes, we reframe the unlearning process as backdoor attacks and defenses: forget-tokens act as backdoor triggers that, when activated in retain-queries, cause disruptions in unlearned models' behaviors, similar to successful backdoor attacks. To mitigate this vulnerability, we propose Random Noise Augmentation (RNA) -- a plug-and-play, model and method agnostic approach with theoretical guarantees for improving the robustness of unlearned models. Extensive experiments demonstrate that RNA significantly improves the robustness of unlearned models, maintains unlearning performances while introducing no additional computational overhead.
Scaling laws have emerged as important components of large language model (LLM) training as they can predict performance gains through scale, and provide guidance on important hyper-parameter choices that would otherwise be expensive. LLMs also rely on large, high-quality training datasets, like those sourced from (sometimes sensitive) user data. Training models on this sensitive user data requires careful privacy protections like differential privacy (DP). However, the dynamics of DP training are significantly different, and consequently their scaling laws are not yet fully understood. In this work, we establish scaling laws that accurately model the intricacies of DP LLM training, providing a complete picture of the compute-privacy-utility tradeoffs and the optimal training configurations in many settings.
The greatest ambition of mechanistic interpretability is to completely rewrite deep neural networks in a format that is more amenable to human understanding, while preserving their behavior and performance. In this paper, we attempt to partially rewrite a large language model using simple natural language explanations. We first approximate one of the feedforward networks in the LLM with a wider MLP with sparsely activating neurons - a transcoder - and use an automated interpretability pipeline to generate explanations for these neurons. We then replace the first layer of this sparse MLP with an LLM-based simulator, which predicts the activation of each neuron given its explanation and the surrounding context. Finally, we measure the degree to which these modifications distort the model's final output. With our pipeline, the model's increase in loss is statistically similar to entirely replacing the sparse MLP output with the zero vector. We employ the same protocol, this time using a sparse autoencoder, on the residual stream of the same layer and obtain similar results. These results suggest that more detailed explanations are needed to improve performance substantially above the zero ablation baseline.
Structured embedding transformations offer a promising approach for enhancing the efficiency and coherence of language model inference. The introduction of Structural Embedding Projection (SEP) provides a mechanism for refining token representations through projection matrices that integrate hierarchical and relational dependencies. The mathematical formulation of SEP enables embedding spaces to capture structured contextual relationships, thereby improving semantic fidelity without significantly increasing computational overhead. Experimental evaluations conducted on a range of linguistic datasets revealed that SEP contributed to reductions in perplexity and enhanced contextual coherence, demonstrating its potential to refine language model outputs. Computational efficiency assessments highlighted variations across different datasets, suggesting that the integration of structured embeddings introduced dataset-dependent trade-offs between inference speed and representational richness. The qualitative analysis of generated responses indicated that SEP enhanced narrative consistency and topic alignment, leading to improved fluency in multi-sentence text generation. The modifications to embedding layers required precise optimization to ensure stable training dynamics, as the introduction of structured transformations altered the traditional representation-learning process. The architectural adjustments necessary for SEP implementation influenced inference latency and memory consumption, requiring a balance between efficiency gains and additional processing demands. The impact of SEP on lexical diversity suggested that embedding modifications influenced the model's vocabulary usage, reflecting a more context-aware selection of generated tokens.
In this work, we present and evaluate SELMA, a Speech-Enabled Language Model for virtual Assistant interactions that integrates audio and text as inputs to a Large Language Model (LLM). SELMA is designed to handle three primary and two auxiliary tasks related to interactions with virtual assistants simultaneously within a single end-to-end model. We employ low-rank adaptation modules for parameter-efficient training of both the audio encoder and the LLM. Additionally, we implement a feature pooling strategy enabling the system to recognize global patterns and improve accuracy on tasks less reliant on individual sequence elements. Experimental results on Voice Trigger (VT) detection, Device-Directed Speech Detection (DDSD), and Automatic Speech Recognition (ASR), demonstrate that our approach both simplifies the typical input processing pipeline of virtual assistants significantly and also improves performance compared to dedicated models for each individual task. SELMA yields relative Equal-Error Rate improvements of 64% on the VT detection task, and 22% on DDSD, while also achieving word error rates close to the baseline.