Open-source Large Language Models (LLMs) have recently gained popularity because of their comparable performance to proprietary LLMs. To efficiently fulfill domain-specialized tasks, open-source LLMs can be refined, without expensive accelerators, using low-rank adapters. However, it is still unknown whether low-rank adapters can be exploited to control LLMs. To address this gap, we demonstrate that an infected adapter can induce, on specific triggers, an LLM to output content defined by an adversary and to even maliciously use tools. To train a Trojan adapter, we propose two novel attacks, POLISHED and FUSION, that improve over prior approaches. POLISHED uses LLM-enhanced paraphrasing to polish benchmark poisoned datasets. In contrast, in the absence of a dataset, FUSION leverages an over-poisoning procedure to transform a benign adaptor. Our experiments validate that our attacks provide higher attack effectiveness than the baseline and, for the purpose of attracting downloads, preserves or improves the adapter's utility. Finally, we provide two case studies to demonstrate that the Trojan adapter can lead a LLM-powered autonomous agent to execute unintended scripts or send phishing emails. Our novel attacks represent the first study of supply chain threats for LLMs through the lens of Trojan plugins.
相關內容
Memory bandwidth is known to be a performance bottleneck for FPGA accelerators, especially when they deal with large multi-dimensional data-sets. A large body of work focuses on reducing of off-chip transfers, but few authors try to improve the efficiency of transfers. This paper addresses the later issue by proposing (i) a compiler-based approach to accelerator's data layout to maximize contiguous access to off-chip memory, and (ii) data packing and runtime compression techniques that take advantage of this layout to further improve memory performance. We show that our approach can decrease the I/O cycles up to $7\times$ compared to un-optimized memory accesses.
Optimal control (OC) algorithms such as Differential Dynamic Programming (DDP) take advantage of the derivatives of the dynamics to efficiently control physical systems. Yet, in the presence of nonsmooth dynamical systems, such class of algorithms are likely to fail due, for instance, to the presence of discontinuities in the dynamics derivatives or because of non-informative gradient. On the contrary, reinforcement learning (RL) algorithms have shown better empirical results in scenarios exhibiting non-smooth effects (contacts, frictions, etc). Our approach leverages recent works on randomized smoothing (RS) to tackle non-smoothness issues commonly encountered in optimal control, and provides key insights on the interplay between RL and OC through the prism of RS methods. This naturally leads us to introduce the randomized Differential Dynamic Programming (R-DDP) algorithm accounting for deterministic but non-smooth dynamics in a very sample-efficient way. The experiments demonstrate that our method is able to solve classic robotic problems with dry friction and frictional contacts, where classical OC algorithms are likely to fail and RL algorithms require in practice a prohibitive number of samples to find an optimal solution.
This work introduces a time domain personalized method (pGTFF0) to achieve intelligibility improvement of noisy speech for Autism Spectrum Disorder (ASD) situation. For this proposal, harmonic features estimated from speech frames are considered as center frequencies of Gammatone auditory filterbanks. A gain factor is further applied to the output of the filtered samples. The key goal is the emulation of an external noise filtering tailored for individuals with ASD. A perceptual listening test demonstrates that ASD volunteers attained lower intelligibility rates than Neurotypical (NT). The proposed solution is compared to three competing approaches considering four acoustic noises at different signal-to-noise ratios. Two objective measures (ESTOI and PESQ) are also adopted for evaluation. The experimental results show that the personalized solution outperformed the competing approaches in terms of intelligibility and quality improvement.
The prevalence of social media presents a growing opportunity to collect and analyse examples of English varieties. Whilst usage of these varieties was - and, in many cases, still is - used only in spoken contexts or hard-to-access private messages, social media sites like Twitter provide a platform for users to communicate informally in a scrapeable format. Notably, Indian English (Hinglish), Singaporean English (Singlish), and African-American English (AAE) can be commonly found online. These varieties pose a challenge to existing natural language processing (NLP) tools as they often differ orthographically and syntactically from standard English for which the majority of these tools are built. NLP models trained on standard English texts produced biased outcomes for users of underrepresented varieties. Some research has aimed to overcome the inherent biases caused by unrepresentative data through techniques like data augmentation or adjusting training models. We aim to address the issue of bias at its root - the data itself. We curate a dataset of tweets from countries with high proportions of underserved English variety speakers, and propose an annotation framework of six categorical classifications along a pseudo-spectrum that measures the degree of standard English and that thereby indirectly aims to surface the manifestations of English varieties in these tweets. Following best annotation practices, our growing corpus features 170,800 tweets taken from 7 countries, labeled by annotators who are from those countries and can communicate in regionally-dominant varieties of English. Our corpus highlights the accuracy discrepancies in pre-trained language identifiers between western English and non-western (i.e., less standard) English varieties. We hope to contribute to the growing literature identifying and reducing the implicit demographic discrepancies in NLP.
We systematically analyze the accuracy of Physics-Informed Neural Networks (PINNs) in approximating solutions to the critical Surface Quasi-Geostrophic (SQG) equation on two-dimensional periodic boxes. The critical SQG equation involves advection and diffusion described by nonlocal periodic operators, posing challenges for neural network-based methods that do not commonly exhibit periodic boundary conditions. In this paper, we present a novel approximation of these operators using their nonperiodic analogs based on singular integral representation formulas and use it to perform error estimates. This idea can be generalized to a larger class of nonlocal partial differential equations whose solutions satisfy prescribed boundary conditions, thereby initiating a new PINNs theory for equations with nonlocalities.
While Large Language Models (LLMs) have proven to be exceptional on a variety of tasks after alignment, they may still produce responses that contradict the context or world knowledge confidently, a phenomenon known as ``hallucination''. In this paper, we demonstrate that reducing the inconsistency between the external knowledge encapsulated in the training data and the intrinsic knowledge inherited in the pretraining corpus could mitigate hallucination in alignment. Specifically, we introduce a novel knowledge consistent alignment (KCA) approach, which involves automatically formulating examinations based on external knowledge for accessing the comprehension of LLMs. For data encompassing knowledge inconsistency, KCA implements several simple yet efficient strategies for processing. We illustrate the superior performance of the proposed KCA approach in mitigating hallucinations across six benchmarks using LLMs of different backbones and scales. Furthermore, we confirm the correlation between knowledge inconsistency and hallucination, signifying the effectiveness of reducing knowledge inconsistency in alleviating hallucinations. Our code, model weights, and data are public at \url{//github.com/fanqiwan/KCA}.
Generating proofs of unsatisfiability is a valuable capability of most SAT solvers, and is an active area of research for SMT solvers. This paper introduces the first method to efficiently generate proofs of unsatisfiability specifically for an important subset of SMT: SAT Modulo Monotonic Theories (SMMT), which includes many useful finite-domain theories (e.g., bit vectors and many graph-theoretic properties) and is used in production at Amazon Web Services. Our method uses propositional definitions of the theory predicates, from which it generates compact Horn approximations of the definitions, which lead to efficient DRAT proofs, leveraging the large investment the SAT community has made in DRAT. In experiments on practical SMMT problems, our proof generation overhead is minimal (7.41% geometric mean slowdown, 28.8% worst-case), and we can generate and check proofs for many problems that were previously intractable.
Large Language Models (LLMs) have shown excellent generalization capabilities that have led to the development of numerous models. These models propose various new architectures, tweaking existing architectures with refined training strategies, increasing context length, using high-quality training data, and increasing training time to outperform baselines. Analyzing new developments is crucial for identifying changes that enhance training stability and improve generalization in LLMs. This survey paper comprehensively analyses the LLMs architectures and their categorization, training strategies, training datasets, and performance evaluations and discusses future research directions. Moreover, the paper also discusses the basic building blocks and concepts behind LLMs, followed by a complete overview of LLMs, including their important features and functions. Finally, the paper summarizes significant findings from LLM research and consolidates essential architectural and training strategies for developing advanced LLMs. Given the continuous advancements in LLMs, we intend to regularly update this paper by incorporating new sections and featuring the latest LLM models.
Graph Neural Networks (GNNs) have recently become increasingly popular due to their ability to learn complex systems of relations or interactions arising in a broad spectrum of problems ranging from biology and particle physics to social networks and recommendation systems. Despite the plethora of different models for deep learning on graphs, few approaches have been proposed thus far for dealing with graphs that present some sort of dynamic nature (e.g. evolving features or connectivity over time). In this paper, we present Temporal Graph Networks (TGNs), a generic, efficient framework for deep learning on dynamic graphs represented as sequences of timed events. Thanks to a novel combination of memory modules and graph-based operators, TGNs are able to significantly outperform previous approaches being at the same time more computationally efficient. We furthermore show that several previous models for learning on dynamic graphs can be cast as specific instances of our framework. We perform a detailed ablation study of different components of our framework and devise the best configuration that achieves state-of-the-art performance on several transductive and inductive prediction tasks for dynamic graphs.
Visual Question Answering (VQA) models have struggled with counting objects in natural images so far. We identify a fundamental problem due to soft attention in these models as a cause. To circumvent this problem, we propose a neural network component that allows robust counting from object proposals. Experiments on a toy task show the effectiveness of this component and we obtain state-of-the-art accuracy on the number category of the VQA v2 dataset without negatively affecting other categories, even outperforming ensemble models with our single model. On a difficult balanced pair metric, the component gives a substantial improvement in counting over a strong baseline by 6.6%.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191