Search

Reward modeling is crucial for aligning large language models (LLMs) with human preferences, especially in reinforcement learning from human feedback (RLHF). However, current reward models mainly produce scalar scores and struggle to incorporate critiques in a natural language format. We hypothesize that predicting both critiques and the scalar reward would improve reward modeling ability. Motivated by this, we propose Critic-RM, a framework that improves reward models using self-generated critiques without extra supervision. Critic-RM employs a two-stage process: generating and filtering high-quality critiques, followed by joint fine-tuning on reward prediction and critique generation. Experiments across benchmarks show that Critic-RM improves reward modeling accuracy by 3.7%-7.3% compared to standard reward models and LLM judges, demonstrating strong performance and data efficiency. Additional studies further validate the effectiveness of generated critiques in rectifying flawed reasoning steps with 2.5%-3.2% gains in improving reasoning accuracy., Comment: 20 pages

In this paper, we present the first multi-modal FHE accelerator based on a unified architecture, which efficiently supports CKKS, TFHE, and their conversion scheme within a single accelerator. To achieve this goal, we first analyze the theoretical foundations of the aforementioned schemes and highlight their composition from a finite number of arithmetic kernels. Then, we investigate the challenges for efficiently supporting these kernels within a unified architecture, which include 1) concurrent support for NTT and FFT, 2) maintaining high hardware utilization across various polynomial lengths, and 3) ensuring consistent performance across diverse arithmetic kernels. To tackle these challenges, we propose a novel FHE accelerator named Trinity, which incorporates algorithm optimizations, hardware component reuse, and dynamic workload scheduling to enhance the acceleration of CKKS, TFHE, and their conversion scheme. By adaptive select the proper allocation of components for NTT and MAC, Trinity maintains high utilization across NTTs with various polynomial lengths and imbalanced arithmetic workloads. The experiment results show that, for the pure CKKS and TFHE workloads, the performance of our Trinity outperforms the state-of-the-art accelerator for CKKS (SHARP) and TFHE (Morphling) by 1.49x and 4.23x, respectively. Moreover, Trinity achieves 919.3x performance improvement for the FHE-conversion scheme over the CPU-based implementation. Notably, despite the performance improvement, the hardware overhead of Trinity is only 85% of the summed circuit areas of SHARP and Morphling., Comment: To be appeared in MICRO 2024. The first ASIC-based FHE accelerator which supports both CKKS, TFHE and their conversions. Provide new SOTA performance record for CKKS, TFHE and conversion

Fully homomorphic encryption (FHE) enables direct computation on encrypted data, making it a crucial technology for privacy protection. However, FHE suffers from significant performance bottlenecks. In this context, GPU acceleration offers a promising solution to bridge the performance gap. Existing efforts primarily focus on single-class FHE schemes, which fail to meet the diverse requirements of data types and functions, prompting the development of hybrid multi-class FHE schemes. However, studies have yet to thoroughly investigate specific GPU optimizations for hybrid FHE schemes. In this paper, we present an efficient GPU-based FHE scheme switching acceleration named Chameleon. First, we propose a scalable NTT acceleration design that adapts to larger CKKS polynomials and smaller TFHE polynomials. Specifically, Chameleon tackles synchronization issues by fusing stages to reduce synchronization, employing polynomial coefficient shuffling to minimize synchronization scale, and utilizing an SM-aware combination strategy to identify the optimal switching point. Second, Chameleon is the first to comprehensively analyze and optimize critical switching operations. It introduces CMux-level parallelization to accelerate LUT evaluation and a homomorphic rotation-free matrix-vector multiplication to improve repacking efficiency. Finally, Chameleon outperforms the state-of-the-art GPU implementations by 1.23x in CKKS HMUL and 1.15x in bootstrapping. It also achieves up to 4.87x and 1.51x speedups for TFHE gate bootstrapping compared to CPU and GPU versions, respectively, and delivers a 67.3x average speedup for scheme switching over CPU-based implementation., Comment: 15 pages, 14 figures

Model merging, such as model souping, is the practice of combining different models with the same architecture together without further training. In this work, we present a model merging methodology that addresses the difficulty of fine-tuning Large Language Models (LLMs) for target tasks in non-English languages, where task-specific data is often unavailable. We focus on mathematical reasoning and without in-language math data, facilitate cross-lingual transfer by composing language and math capabilities. Starting from the same pretrained model, we fine-tune separate "experts" on math instruction data in English and on generic instruction data in the target language. We then replace the top and bottom transformer layers of the math expert directly with layers from the language expert, which consequently enhances math performance in the target language. The resulting merged models outperform the individual experts and other merging methods on the math benchmark, MGSM, by 10% across four major languages where math instruction data is scarce. In addition, this layer swapping is simple, inexpensive, and intuitive, as it is based on an interpretative analysis of the most important parameter changes during the fine-tuning of each expert. The ability to successfully re-compose LLMs for cross-lingual transfer in this manner opens up future possibilities to combine model expertise, create modular solutions, and transfer reasoning capabilities across languages all post hoc., Comment: 11 main pages, 23 pages total, 9 figures, 5 tables

The wide deployment of Large Language Models (LLMs) has given rise to strong demands for optimizing their inference performance. Today's techniques serving this purpose primarily focus on reducing latency and improving throughput through algorithmic and hardware enhancements, while largely overlooking their privacy side effects, particularly in a multi-user environment. In our research, for the first time, we discovered a set of new timing side channels in LLM systems, arising from shared caches and GPU memory allocations, which can be exploited to infer both confidential system prompts and those issued by other users. These vulnerabilities echo security challenges observed in traditional computing systems, highlighting an urgent need to address potential information leakage in LLM serving infrastructures. In this paper, we report novel attack strategies designed to exploit such timing side channels inherent in LLM deployments, specifically targeting the Key-Value (KV) cache and semantic cache widely used to enhance LLM inference performance. Our approach leverages timing measurements and classification models to detect cache hits, allowing an adversary to infer private prompts with high accuracy. We also propose a token-by-token search algorithm to efficiently recover shared prompt prefixes in the caches, showing the feasibility of stealing system prompts and those produced by peer users. Our experimental studies on black-box testing of popular online LLM services demonstrate that such privacy risks are completely realistic, with significant consequences. Our findings underscore the need for robust mitigation to protect LLM systems against such emerging threats., Comment: This work was submitted for review on Sept. 5, 2024, and the initial version was uploaded to Arxiv on Sept. 30, 2024

The development and evaluation of Large Language Models (LLMs) have largely focused on individual capabilities. However, this overlooks the intersection of multiple abilities across different types of expertise that are often required for real-world tasks, which we term cross capabilities. To systematically explore this concept, we first define seven core individual capabilities and then pair them to form seven common cross capabilities, each supported by a manually constructed taxonomy. Building on these definitions, we introduce CrossEval, a benchmark comprising 1,400 human-annotated prompts, with 100 prompts for each individual and cross capability. To ensure reliable evaluation, we involve expert annotators to assess 4,200 model responses, gathering 8,400 human ratings with detailed explanations to serve as reference examples. Our findings reveal that, in both static evaluations and attempts to enhance specific abilities, current LLMs consistently exhibit the "Law of the Weakest Link," where cross-capability performance is significantly constrained by the weakest component. Specifically, across 58 cross-capability scores from 17 models, 38 scores are lower than all individual capabilities, while 20 fall between strong and weak, but closer to the weaker ability. These results highlight the under-performance of LLMs in cross-capability tasks, making the identification and improvement of the weakest capabilities a critical priority for future research to optimize performance in complex, multi-dimensional scenarios., Comment: Data, Code, & Benchmark: www.llm-cross-capabilities.org

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Fully Homomorphic Encryption (FHE), a novel cryptographic theory enabling computation directly on ciphertext data, offers significant security benefits but is hampered by substantial performance overhead. In recent years, a series of accelerator designs have significantly enhanced the performance of FHE applications, bringing them closer to real-world applicability. However, these accelerators face challenges related to large on-chip memory and area. Additionally, FHE algorithms undergo rapid development, rendering the previous accelerator designs less perfectly adapted to the evolving landscape of optimized FHE applications. In this paper, we conducted a detailed analysis of existing applications with the new FHE method, making two key observations: 1) the bottleneck of FHE applications shifts from NTT to the inner-product operation, and 2) the optimal {\alpha} of KeySwitch changes with the decrease in multiplicative level. Based on these observations, we designed an accelerator named Taiyi, which includes specific hardware for the inner-product operation and optimizes the NTT and BConv operations through algorithmic derivation. A comparative evaluation of Taiyi against previous state-of-the-art designs reveals an average performance improvement of 1.5x and reduces the area overhead by 15.7%., Comment: 14 pages, 15 figures

Integrity is critical for maintaining system security, as it ensures that only genuine software is loaded onto a machine. Although confidential virtual machines (CVMs) function within isolated environments separate from the host, it is important to recognize that users still encounter challenges in maintaining control over the integrity of the code running within the trusted execution environments (TEEs). The presence of a sophisticated operating system (OS) raises the possibility of dynamically creating and executing any code, making user applications within TEEs vulnerable to interference or tampering if the guest OS is compromised. To address this issue, this paper introduces NestedSGX, a framework which leverages virtual machine privilege level (VMPL), a recent hardware feature available on AMD SEV-SNP to enable the creation of hardware enclaves within the guest VM. Similar to Intel SGX, NestedSGX considers the guest OS untrusted for loading potentially malicious code. It ensures that only trusted and measured code executed within the enclave can be remotely attested. To seamlessly protect existing applications, NestedSGX aims for compatibility with Intel SGX by simulating SGX leaf functions. We have also ported the SGX SDK and the Occlum library OS to NestedSGX, enabling the use of existing SGX toolchains and applications in the system. Performance evaluations show that context switches in NestedSGX take about 32,000 -- 34,000 cycles, approximately $1.9\times$ -- $2.1\times$ higher than that of Intel SGX. NestedSGX incurs minimal overhead in most real-world applications, with an average overhead below 2% for computation and memory intensive workloads and below 15.68% for I/O intensive workloads., Comment: NDSS 2025 accepted version

The ability of large language models (LLMs) to process visual inputs has given rise to general-purpose vision systems, unifying various vision-language (VL) tasks by instruction tuning. However, due to the enormous diversity in input-output formats in the vision domain, existing general-purpose models fail to successfully integrate segmentation and multi-image inputs with coarse-level tasks into a single framework. In this work, we introduce VistaLLM, a powerful visual system that addresses coarse- and fine-grained VL tasks over single and multiple input images using a unified framework. VistaLLM utilizes an instruction-guided image tokenizer that filters global embeddings using task descriptions to extract compressed and refined features from numerous images. Moreover, VistaLLM employs a gradient-aware adaptive sampling technique to represent binary segmentation masks as sequences, significantly improving over previously used uniform sampling. To bolster the desired capability of VistaLLM, we curate CoinIt, a comprehensive coarse-to-fine instruction tuning dataset with 6.8M samples. We also address the lack of multi-image grounding datasets by introducing a novel task, AttCoSeg (Attribute-level Co-Segmentation), which boosts the model's reasoning and grounding capability over multiple input images. Extensive experiments on a wide range of V- and VL tasks demonstrate the effectiveness of VistaLLM by achieving consistent state-of-the-art performance over strong baselines across all downstream tasks. Our project page can be found at https://shramanpramanick.github.io/VistaLLM/., Comment: CVPR 2024 Highlight

Few-shot Relation Extraction (FSRE) aims to extract relational facts from a sparse set of labeled corpora. Recent studies have shown promising results in FSRE by employing Pre-trained Language Models (PLMs) within the framework of supervised contrastive learning, which considers both instances and label facts. However, how to effectively harness massive instance-label pairs to encompass the learned representation with semantic richness in this learning paradigm is not fully explored. To address this gap, we introduce a novel synergistic anchored contrastive pre-training framework. This framework is motivated by the insight that the diverse viewpoints conveyed through instance-label pairs capture incomplete yet complementary intrinsic textual semantics. Specifically, our framework involves a symmetrical contrastive objective that encompasses both sentence-anchored and label-anchored contrastive losses. By combining these two losses, the model establishes a robust and uniform representation space. This space effectively captures the reciprocal alignment of feature distributions among instances and relational facts, simultaneously enhancing the maximization of mutual information across diverse perspectives within the same relation. Experimental results demonstrate that our framework achieves significant performance enhancements compared to baseline models in downstream FSRE tasks. Furthermore, our approach exhibits superior adaptability to handle the challenges of domain shift and zero-shot relation extraction. Our code is available online at https://github.com/AONE-NLP/FSRE-SaCon.

Fine-tuning pre-trained language models (LMs) has become the de facto standard in many NLP tasks. Nevertheless, fine-tuned LMs are still prone to robustness issues, such as adversarial robustness and model calibration. Several perspectives of robustness for LMs have been studied independently, but lacking a unified consideration in multiple perspectives. In this paper, we propose Robustifying LMs via Adversarial perturbation with Selective Training (RoAST), a simple yet effective fine-tuning technique to enhance the multi-perspective robustness of LMs in a unified way. RoAST effectively incorporates two important sources for the model robustness, robustness on the perturbed inputs and generalizable knowledge in pre-trained LMs. To be specific, RoAST introduces adversarial perturbation during fine-tuning while the model parameters are selectively updated upon their relative importance to minimize unnecessary deviation. Under a unified evaluation of fine-tuned LMs by incorporating four representative perspectives of model robustness, we demonstrate the effectiveness of RoAST compared to state-of-the-art fine-tuning methods on six different types of LMs, which indicates its usefulness in practice., Comment: 33 pages, accepted at EMNLP 2023 Findings

Red-teaming is a common practice for mitigating unsafe behaviors in Large Language Models (LLMs), which involves thoroughly assessing LLMs to identify potential flaws and addressing them with responsible and accurate responses. While effective, manual red-teaming is costly, and existing automatic red-teaming typically discovers safety risks without addressing them. In this paper, we propose a Multi-round Automatic Red-Teaming (MART) method, which incorporates both automatic adversarial prompt writing and safe response generation, significantly increasing red-teaming scalability and the safety of the target LLM. Specifically, an adversarial LLM and a target LLM interplay with each other in an iterative manner, where the adversarial LLM aims to generate challenging prompts that elicit unsafe responses from the target LLM, while the target LLM is fine-tuned with safety aligned data on these adversarial prompts. In each round, the adversarial LLM crafts better attacks on the updated target LLM, while the target LLM also improves itself through safety fine-tuning. On adversarial prompt benchmarks, the violation rate of an LLM with limited safety alignment reduces up to 84.7% after 4 rounds of MART, achieving comparable performance to LLMs with extensive adversarial prompt writing. Notably, model helpfulness on non-adversarial prompts remains stable throughout iterations, indicating the target LLM maintains strong performance on instruction following.

Knowledge Distillation (KD) compresses computationally expensive pre-trained language models (PLMs) by transferring their knowledge to smaller models, allowing their use in resource-constrained or real-time settings. However, most smaller models fail to surpass the performance of the original larger model, resulting in sacrificing performance to improve inference speed. To address this issue, we propose Co-Training and Co-Distillation (CTCD), a novel framework that improves performance and inference speed together by co-training two models while mutually distilling knowledge. The CTCD framework successfully achieves this based on two significant findings: 1) Distilling knowledge from the smaller model to the larger model during co-training improves the performance of the larger model. 2) The enhanced performance of the larger model further boosts the performance of the smaller model. The CTCD framework shows promise as it can be combined with existing techniques like architecture design or data augmentation, replacing one-way KD methods, to achieve further performance improvement. Extensive ablation studies demonstrate the effectiveness of CTCD, and the small model distilled by CTCD outperforms the original larger model by a significant margin of 1.66 on the GLUE benchmark., Comment: Findings of EMNLP 2023

We present a series of long-context LLMs that support effective context windows of up to 32,768 tokens. Our model series are built through continual pretraining from Llama 2 with longer training sequences and on a dataset where long texts are upsampled. We perform extensive evaluation on language modeling, synthetic context probing tasks, and a wide range of research benchmarks. On research benchmarks, our models achieve consistent improvements on most regular tasks and significant improvements on long-context tasks over Llama 2. Notably, with a cost-effective instruction tuning procedure that does not require human-annotated long instruction data, the 70B variant can already surpass gpt-3.5-turbo-16k's overall performance on a suite of long-context tasks. Alongside these results, we provide an in-depth analysis on the individual components of our method. We delve into Llama's position encodings and discuss its limitation in modeling long dependencies. We also examine the impact of various design choices in the pretraining process, including the data mix and the training curriculum of sequence lengths -- our ablation experiments suggest that having abundant long texts in the pretrain dataset is not the key to achieving strong performance, and we empirically verify that long context continual pretraining is more efficient and similarly effective compared to pretraining from scratch with long sequences.

Aspect-based sentiment classification is a crucial problem in fine-grained sentiment analysis, which aims to predict the sentiment polarity of the given aspect according to its context. Previous works have made remarkable progress in leveraging attention mechanism to extract opinion words for different aspects. However, a persistent challenge is the effective management of semantic mismatches, which stem from attention mechanisms that fall short in adequately aligning opinions words with their corresponding aspect in multi-aspect sentences. To address this issue, we propose a novel Aspect-oriented Opinion Alignment Network (AOAN) to capture the contextual association between opinion words and the corresponding aspect. Specifically, we first introduce a neighboring span enhanced module which highlights various compositions of neighboring words and given aspects. In addition, we design a multi-perspective attention mechanism that align relevant opinion information with respect to the given aspect. Extensive experiments on three benchmark datasets demonstrate that our model achieves state-of-the-art results. The source code is available at https://github.com/AONE-NLP/ABSA-AOAN., Comment: 8 pages, 5 figure, ECAI 2023

The nanoconfinement of water can result in dramatic differences in its physical and chemical properties compared to bulk water. However, a detailed molecular-level understanding of these properties is still lacking. Vibrational spectroscopy, such as Raman and infrared, is a popular experimental tool for studying the structure and dynamics of water, and is often complemented by atomistic simulations to interpret experimental spectra, but there have been few theoretical spectroscopy studies of nanoconfined water using first-principles methods at ambient conditions, let alone under extreme pressure-temperature conditions. Here, we computed the Raman and IR spectra of water nanoconfined by graphene at ambient and extreme pressure-temperature conditions using ab intio simulations. Our results revealed alterations in the Raman stretching and low-frequency bands due to the graphene confinement. We also found spectroscopic evidence indicating that nanoconfinement considerably changes the tetrahedral hydrogen bond network, which is typically found in bulk water. Furthermore, we observed an unusual bending band in the Raman spectrum at ~10 GPa and 1000 K, which is attributed to the unique molecular structure of confined ionic water. Additionally, we found that at ~20 GPa and 1000 K, confined water transformed into a superionic fluid, making it challenging to identify the IR stretching band. Finally, we computed the ionic conductivity of confined water in the ionic and superionic phases. Our results highlight the efficacy of Raman and IR spectroscopy in studying the structure and dynamics of nanoconfined water in a large pressure-temperature range. Our predicted Raman and IR spectra can serve as a valuable guide for future experiments.

In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters. Our fine-tuned LLMs, called Llama 2-Chat, are optimized for dialogue use cases. Our models outperform open-source chat models on most benchmarks we tested, and based on our human evaluations for helpfulness and safety, may be a suitable substitute for closed-source models. We provide a detailed description of our approach to fine-tuning and safety improvements of Llama 2-Chat in order to enable the community to build on our work and contribute to the responsible development of LLMs.

ARM recently introduced the Confidential Compute Architecture (CCA) as part of the upcoming ARMv9-A architecture. CCA enables the support of confidential virtual machines (cVMs) within a separate world called the Realm world, providing protection from the untrusted normal world. While CCA offers a promising future for confidential computing, the widespread availability of CCA hardware is not expected in the near future, according to ARM's roadmap. To address this gap, we present virtCCA, an architecture that facilitates virtualized CCA using TrustZone, a mature hardware feature available on existing ARM platforms. Notably, virtCCA can be implemented on platforms equipped with the Secure EL2 (S-EL2) extension available from ARMv8.4 onwards, as well as on earlier platforms that lack S-EL2 support. virtCCA is fully compatible with the CCA specifications at the API level. We have developed the entire CCA software and firmware stack on top of virtCCA, including the enhancements to the normal world's KVM to support cVMs, and the TrustZone Management Monitor (TMM) that enforces isolation among cVMs and provides cVM life-cycle management. We have implemented virtCCA on real ARM servers, with and without S-EL2 support. Our evaluation, conducted on micro-benchmarks and macro-benchmarks, demonstrates that the overhead of running cVMs is acceptable compared to running normal-world VMs. Specifically, in a set of real-world workloads, the overhead of virtCCA-SEL2 is less than 29.5% for I/O intensive workloads, while virtCCA-EL3 outperforms the baseline in most cases.

Distillation from Weak Teacher (DWT) is a method of transferring knowledge from a smaller, weaker teacher model to a larger student model to improve its performance. Previous studies have shown that DWT can be effective in the vision domain and natural language processing (NLP) pre-training stage. Specifically, DWT shows promise in practical scenarios, such as enhancing new generation or larger models using pre-trained yet older or smaller models and lacking a resource budget. However, the optimal conditions for using DWT have yet to be fully investigated in NLP pre-training. Therefore, this study examines three key factors to optimize DWT, distinct from those used in the vision domain or traditional knowledge distillation. These factors are: (i) the impact of teacher model quality on DWT effectiveness, (ii) guidelines for adjusting the weighting value for DWT loss, and (iii) the impact of parameter remapping as a student model initialization technique for DWT., Comment: Findings of ACL 2023

Many real-world applications require making multiple predictions from the same text. Fine-tuning a large pre-trained language model for each downstream task causes computational burdens in the inference time due to several times of forward passes. To amortize the computational cost, freezing the language model and building lightweight models for downstream tasks based on fixed text representations are common solutions. Accordingly, how to learn fixed but general text representations that can generalize well to unseen downstream tasks becomes a challenge. Previous works have shown that the generalizability of representations can be improved by fine-tuning the pre-trained language model with some source tasks in a multi-tasking way. In this work, we propose a prefix-based method to learn the fixed text representations with source tasks. We learn a task-specific prefix for each source task independently and combine them to get the final representations. Our experimental results show that prefix-based training performs better than multi-tasking training and can update the text representations at a smaller computational cost than multi-tasking training., Comment: Paper accepted by EMNLP 2023 Findings

Prompt tuning is one of the successful approaches for parameter-efficient tuning of pre-trained language models. Despite being arguably the most parameter-efficient (tuned soft prompts constitute <0.1% of total parameters), it typically performs worse than other efficient tuning methods and is quite sensitive to hyper-parameters. In this work, we introduce Residual Prompt Tuning - a simple and efficient method that significantly improves the performance and stability of prompt tuning. We propose to reparameterize soft prompt embeddings using a shallow network with a residual connection. Our experiments show that Residual Prompt Tuning significantly outperforms prompt tuning on SuperGLUE benchmark. Notably, our method reaches +7 points improvement over prompt tuning with T5-Base and allows to reduce the prompt length by 10x without hurting performance. In addition, we show that our approach is robust to the choice of learning rate and prompt initialization, and is effective in few-shot settings., Comment: ACL Findings 2023

We present Multiscale Multiview Vision Transformers (MMViT), which introduces multiscale feature maps and multiview encodings to transformer models. Our model encodes different views of the input signal and builds several channel-resolution feature stages to process the multiple views of the input at different resolutions in parallel. At each scale stage, we use a cross-attention block to fuse information across different views. This enables the MMViT model to acquire complex high-dimensional representations of the input at different resolutions. The proposed model can serve as a backbone model in multiple domains. We demonstrate the effectiveness of MMViT on audio and image classification tasks, achieving state-of-the-art results.

Whether by processing videos with fixed resolution from start to end or incorporating pooling and down-scaling strategies, existing video transformers process the whole video content throughout the network without specially handling the large portions of redundant information. In this paper, we present a Supertoken Video Transformer (SVT) that incorporates a Semantic Pooling Module (SPM) to aggregate latent representations along the depth of visual transformer based on their semantics, and thus, reduces redundancy inherent in video inputs.~Qualitative results show that our method can effectively reduce redundancy by merging latent representations with similar semantics and thus increase the proportion of salient information for downstream tasks.~Quantitatively, our method improves the performance of both ViT and MViT while requiring significantly less computations on the Kinectics and Something-Something-V2 benchmarks.~More specifically, with our SPM, we improve the accuracy of MAE-pretrained ViT-B and ViT-L by 1.5% with 33% less GFLOPs and by 0.2% with 55% less FLOPs, respectively, on the Kinectics-400 benchmark, and improve the accuracy of MViTv2-B by 0.2% and 0.3% with 22% less GFLOPs on Kinectics-400 and Something-Something-V2, respectively.

We introduce Progressive Prompts - a simple and efficient approach for continual learning in language models. Our method allows forward transfer and resists catastrophic forgetting, without relying on data replay or a large number of task-specific parameters. Progressive Prompts learns a new soft prompt for each task and sequentially concatenates it with the previously learned prompts, while keeping the base model frozen. Experiments on standard continual learning benchmarks show that our approach outperforms state-of-the-art methods, with an improvement >20% in average test accuracy over the previous best-preforming method on T5 model. We also explore a more challenging continual learning setup with longer sequences of tasks and show that Progressive Prompts significantly outperforms prior methods.

Large multilingual language models typically rely on a single vocabulary shared across 100+ languages. As these models have increased in parameter count and depth, vocabulary size has remained largely unchanged. This \textit{vocabulary bottleneck} limits the representational capabilities of multilingual models like XLM-R. In this paper, we introduce a new approach for scaling to very large multilingual vocabularies by de-emphasizing token sharing between languages with little lexical overlap and assigning vocabulary capacity to achieve sufficient coverage for each individual language. Tokenizations using our vocabulary are typically more semantically meaningful and shorter compared to XLM-R. Leveraging this improved vocabulary, we train XLM-V, a multilingual language model with a one million token vocabulary. XLM-V outperforms XLM-R on every task we tested on ranging from natural language inference (XNLI), question answering (MLQA, XQuAD, TyDiQA), to named entity recognition (WikiAnn). XLM-V is particularly effective on low-resource language tasks and outperforms XLM-R by 11.2% and 5.8% absolute on MasakhaNER and Americas NLI, respectively., Comment: EMNLP 2023

In this paper, we propose TensorFHE, an FHE acceleration solution based on GPGPU for real applications on encrypted data. TensorFHE utilizes Tensor Core Units (TCUs) to boost the computation of Number Theoretic Transform (NTT), which is the part of FHE with highest time-cost. Moreover, TensorFHE focuses on performing as many FHE operations as possible in a certain time period rather than reducing the latency of one operation. Based on such an idea, TensorFHE introduces operation-level batching to fully utilize the data parallelism in GPGPU. We experimentally prove that it is possible to achieve comparable performance with GPGPU as with state-of-the-art ASIC accelerators. TensorFHE performs 913 KOPS and 88 KOPS for NTT and HMULT (key FHE kernels) within NVIDIA A100 GPGPU, which is 2.61x faster than state-of-the-art FHE implementation on GPGPU; Moreover, TensorFHE provides comparable performance to the ASIC FHE accelerators, which makes it even 2.9x faster than the F1+ with a specific workload. Such a pure software acceleration based on commercial hardware with high performance can open up usage of state-of-the-art FHE algorithms for a broad set of applications in real systems., Comment: To be appeared in the 29th IEEE International Symposium on High-Performance Computer Architecture (HPCA-29), 2023

Masked Language Modeling (MLM) has proven to be an essential component of Vision-Language (VL) pretraining. To implement MLM, the researcher must make two design choices: the masking strategy, which determines which tokens to mask, and the masking rate, which determines how many tokens to mask. Previous work has focused primarily on the masking strategy while setting the masking rate at a default of 15\%. In this paper, we show that increasing this masking rate improves downstream performance while simultaneously reducing performance gap among different masking strategies, rendering the uniform masking strategy competitive to other more complex ones. Surprisingly, we also discover that increasing the masking rate leads to gains in Image-Text Matching (ITM) tasks, suggesting that the role of MLM goes beyond language modeling in VL pretraining.

As cache-based side-channel attacks become serious security problems, various defenses have been proposed and deployed in both software and hardware. Consequently, cache-based side-channel attacks on processes co-residing on the same core are becoming extremely difficult. Most of recent attacks then shift their focus to the last-level cache (LLC). Although cache partitioning is currently the most promising defense against the attacks abusing LLC, it is ineffective in thwarting the side-channel attacks that automatically create eviction sets or bypass the user address space layout randomization. In fact, these attacks are largely undefended in current computer systems. We propose Remapped Cache Layout (\textsf{RCL}) -- a pure hardware defense against a broad range of conflict-based side-channel attacks. \textsf{RCL} obfuscates the mapping from address to cache sets; therefore, an attacker cannot accurately infer the location of her data in caches or using a cache set to infer her victim's data. To our best knowledge, it is the first defense to thwart the aforementioned largely undefended side-channel attacks . \textsf{RCL} has been implemented in a superscalar processor and detailed evaluation results show that \textsf{RCL} incurs only small costs in area, frequency and execution time.