CMoE: Fast Carving of Mixture-of-Experts for Efficient LLM Inference

Large language models (LLMs) achieve impressive performance by scaling model parameters, but this comes with significant inference overhead. Feed-forward networks (FFNs), which dominate LLM parameters, exhibit high activation sparsity in hidden neurons. To exploit this, researchers have proposed using a mixture-of-experts (MoE) architecture, where only a subset of parameters is activated. However, existing approaches often require extensive training data and resources, limiting their practicality. We propose CMoE (Carved MoE), a novel framework to efficiently carve MoE models from dense models. CMoE achieves remarkable performance through efficient expert grouping and lightweight adaptation. First, neurons are grouped into shared and routed experts based on activation rates. Next, we construct a routing mechanism without training from scratch, incorporating a differentiable routing process and load balancing. Using modest data, CMoE produces a well-designed, usable MoE from a 7B dense model within five minutes. With lightweight fine-tuning, it achieves high-performance recovery in under an hour. We make our code publicly available at https://github.com/JarvisPei/CMoE.

KVTuner: Sensitivity-Aware Layer-wise Mixed Precision KV Cache Quantization for Efficient and Nearly Lossless LLM Inference

KV cache quantization can improve Large Language Models (LLMs) inference throughput and latency in long contexts and large batch-size scenarios while preserving LLMs effectiveness. However, current methods have three unsolved issues: overlooking layer-wise sensitivity to KV cache quantization, high overhead of online fine-grained decision-making, and low flexibility to different LLMs and constraints. Therefore, we thoroughly analyze the inherent correlation of layer-wise transformer attention patterns to KV cache quantization errors and study why key cache is more important than value cache for quantization error reduction. We further propose a simple yet effective framework KVTuner to adaptively search for the optimal hardware-friendly layer-wise KV quantization precision pairs for coarse-grained KV cache with multi-objective optimization and directly utilize the offline searched configurations during online inference. To reduce the computational cost of offline calibration, we utilize the intra-layer KV precision pair pruning and inter-layer clustering to reduce the search space. Experimental results show that we can achieve nearly lossless 3.25-bit mixed precision KV cache quantization for LLMs like Llama-3.1-8B-Instruct and 4.0-bit for sensitive models like Qwen2.5-7B-Instruct on mathematical reasoning tasks. The maximum inference throughput can be improved by 38.3% compared with KV8 quantization over various context lengths.

FuseGPT: Learnable Layers Fusion of Generative Pre-trained Transformers

Generative Pre-trained Transformers (GPTs) have demonstrated remarkable performance across diverse domains through the extensive scaling of model parameters. Recent works observe the redundancy across the transformer blocks and develop compression methods by structured pruning of the unimportant blocks. However, such straightforward elimination will always provide irreversible performance degradation. In this paper, we propose FuseGPT, a novel methodology to recycle the pruned transformer blocks to further recover the model performance. Firstly we introduce a new importance detection metric, Macro Influence (MI), to detect the long-term influence of each transformer block by calculating their loss of information after removal. Then we propose group-level layers fusion, which adopts the parameters in layers of the unimportant blocks and injects them into the corresponding layers inside the neighboring blocks. The fusion is not one-off but through iterative parameter updates by lightweight group-level fine-tuning. Specifically, these injected parameters are frozen but weighted with learnable rank decomposition matrices to reduce the overhead during fine-tuning. Our approach not only works well on large language models but also on large multimodal models. The experiments have shown that, by using modest amounts of data, FuseGPT can outperform previous works in both perplexity and zero-shot task performance.

State Chrono Representation for Enhancing Generalization in Reinforcement Learning

In reinforcement learning with image-based inputs, it is crucial to establish a robust and generalizable state representation. Recent advancements in metric learning, such as deep bisimulation metric approaches, have shown promising results in learning structured low-dimensional representation space from pixel observations, where the distance between states is measured based on task-relevant features. However, these approaches face challenges in demanding generalization tasks and scenarios with non-informative rewards. This is because they fail to capture sufficient long-term information in the learned representations. To address these challenges, we propose a novel State Chrono Representation (SCR) approach. SCR augments state metric-based representations by incorporating extensive temporal information into the update step of bisimulation metric learning. It learns state distances within a temporal framework that considers both future dynamics and cumulative rewards over current and long-term future states. Our learning strategy effectively incorporates future behavioral information into the representation space without introducing a significant number of additional parameters for modeling dynamics. Extensive experiments conducted in DeepMind Control and Meta-World environments demonstrate that SCR achieves better performance comparing to other recent metric-based methods in demanding generalization tasks. The codes of SCR are available in https://github.com/jianda-chen/SCR.

Improving the Generalization of Unseen Crowd Behaviors for Reinforcement Learning based Local Motion Planners

Deploying a safe mobile robot policy in scenarios with human pedestrians is challenging due to their unpredictable movements. Current Reinforcement Learning-based motion planners rely on a single policy to simulate pedestrian movements and could suffer from the over-fitting issue. Alternatively, framing the collision avoidance problem as a multi-agent framework, where agents generate dynamic movements while learning to reach their goals, can lead to conflicts with human pedestrians due to their homogeneity. To tackle this problem, we introduce an efficient method that enhances agent diversity within a single policy by maximizing an information-theoretic objective. This diversity enriches each agent's experiences, improving its adaptability to unseen crowd behaviors. In assessing an agent's robustness against unseen crowds, we propose diverse scenarios inspired by pedestrian crowd behaviors. Our behavior-conditioned policies outperform existing works in these challenging scenes, reducing potential collisions without additional time or travel.

Overcoming Negative Transfer by Online Selection: Distant Domain Adaptation for Fault Diagnosis

Unsupervised domain adaptation (UDA) has achieved remarkable success in fault diagnosis, bringing significant benefits to diverse industrial applications. While most UDA methods focus on cross-working condition scenarios where the source and target domains are notably similar, real-world applications often grapple with severe domain shifts. We coin the term `distant domain adaptation problem' to describe the challenge of adapting from a labeled source domain to a significantly disparate unlabeled target domain. This problem exhibits the risk of negative transfer, where extraneous knowledge from the source domain adversely affects the target domain performance. Unfortunately, conventional UDA methods often falter in mitigating this negative transfer, leading to suboptimal performance. In response to this challenge, we propose a novel Online Selective Adversarial Alignment (OSAA) approach. Central to OSAA is its ability to dynamically identify and exclude distant source samples via an online gradient masking approach, focusing primarily on source samples that closely resemble the target samples. Furthermore, recognizing the inherent complexities in bridging the source and target domains, we construct an intermediate domain to act as a transitional domain and ease the adaptation process. Lastly, we develop a class-conditional adversarial adaptation to address the label distribution disparities while learning domain invariant representation to account for potential label distribution disparities between the domains. Through detailed experiments and ablation studies on two real-world datasets, we validate the superior performance of the OSAA method over state-of-the-art methods, underscoring its significant utility in practical scenarios with severe domain shifts.

Decomposing Label Space, Format and Discrimination: Rethinking How LLMs Respond and Solve Tasks via In-Context Learning

In-context Learning (ICL) has emerged as a powerful capability alongside the development of scaled-up large language models (LLMs). By instructing LLMs using few-shot demonstrative examples, ICL enables them to perform a wide range of tasks without updating millions of parameters. However, the precise contributions of demonstrations towards improving end-task performance have not been thoroughly investigated in recent analytical studies. In this paper, we empirically decompose the overall performance of ICL into three dimensions, label space, format, and discrimination, and we evaluate four general-purpose LLMs across a diverse range of tasks. Counter-intuitively, we find that the demonstrations have a marginal impact on provoking discriminative knowledge of language models. However, ICL exhibits significant efficacy in regulating the label space and format which helps LLMs to respond in desired label words. We then demonstrate this ability functions similar to detailed instructions for LLMs to follow. We additionally provide an in-depth analysis of the mechanism of retrieval helping with ICL and find that retrieving the most semantically similar examples notably boosts model's discriminative capability.

Backdoor Attacks on Dense Passage Retrievers for Disseminating Misinformation

Dense retrievers and retrieval-augmented language models have been widely used in various NLP applications. Despite being designed to deliver reliable and secure outcomes, the vulnerability of retrievers to potential attacks remains unclear, raising concerns about their security. In this paper, we introduce a novel scenario where the attackers aim to covertly disseminate targeted misinformation, such as hate speech or advertisement, through a retrieval system. To achieve this, we propose a perilous backdoor attack triggered by grammar errors in dense passage retrieval. Our approach ensures that attacked models can function normally for standard queries but are manipulated to return passages specified by the attacker when users unintentionally make grammatical mistakes in their queries. Extensive experiments demonstrate the effectiveness and stealthiness of our proposed attack method. When a user query is error-free, our model consistently retrieves accurate information while effectively filtering out misinformation from the top-k results. However, when a query contains grammar errors, our system shows a significantly higher success rate in fetching the targeted content.

Adapt in Contexts: Retrieval-Augmented Domain Adaptation via In-Context Learning

Large language models (LLMs) have showcased their capability with few-shot inference known as in-context learning. However, in-domain demonstrations are not always readily available in real scenarios, leading to cross-domain in-context learning. Besides, LLMs are still facing challenges in long-tail knowledge in unseen and unfamiliar domains. The above limitations demonstrate the necessity of Unsupervised Domain Adaptation (UDA). In this paper, we study the UDA problem under an in-context learning setting to adapt language models from the source domain to the target domain without any target labels. The core idea is to retrieve a subset of cross-domain elements that are the most similar to the query, and elicit language model to adapt in an in-context manner by learning both target domain distribution and the discriminative task signal simultaneously with the augmented cross-domain in-context examples. We devise different prompting and training strategies, accounting for different LM architectures to learn the target distribution via language modeling. With extensive experiments on Sentiment Analysis (SA) and Named Entity Recognition (NER) tasks, we thoroughly study the effectiveness of ICL for domain transfer and demonstrate significant improvements over baseline models.

SOUL: Towards Sentiment and Opinion Understanding of Language

Sentiment analysis is a well-established natural language processing task, with sentiment polarity classification being one of its most popular and representative tasks. However, despite the success of pre-trained language models in this area, they often fall short of capturing the broader complexities of sentiment analysis. To address this issue, we propose a new task called Sentiment and Opinion Understanding of Language (SOUL). SOUL aims to evaluate sentiment understanding through two subtasks: Review Comprehension (RC) and Justification Generation (JG). RC seeks to validate statements that focus on subjective information based on a review text, while JG requires models to provide explanations for their sentiment predictions. To enable comprehensive evaluation, we annotate a new dataset comprising 15,028 statements from 3,638 reviews. Experimental results indicate that SOUL is a challenging task for both small and large language models, with a performance gap of up to 27% when compared to human performance. Furthermore, evaluations conducted with both human experts and GPT-4 highlight the limitations of the small language model in generating reasoning-based justifications. These findings underscore the challenging nature of the SOUL task for existing models, emphasizing the need for further advancements in sentiment analysis to address its complexities. The new dataset and code are available at https://github.com/DAMO-NLP-SG/SOUL.