We revisit the likelihood ratio between a pretrained large language model (LLM) and its finetuned variant as a criterion for out-of-distribution (OOD) detection. The intuition behind such a criterion is that, the pretrained LLM has the prior knowledge about OOD data due to its large amount of training data, and once finetuned with the in-distribution data, the LLM has sufficient knowledge to distinguish their difference. Leveraging the power of LLMs, we show that, the likelihood ratio can serve as an effective OOD detection criterion. Moreover, we apply the proposed LLM-based likelihood ratio to detect OOD questions in question-answering (QA) systems, which can be used to improve the performance of specialized LLMs for general questions. Given that likelihood can be easily obtained by the loss functions within contemporary neural network frameworks, it is straightforward to implement this approach in practice with \textbf{three lines} of code. Since both the pretrained LLMs and its various finetuned models are widely available from online platforms such as Hugging Face, our proposed criterion can be effortlessly incorporated for OOD detection without the need for further training. We conduct comprehensive evaluation across on multiple settings, including far OOD, near OOD, spam detection, and QA scenarios, to demonstrate the effectiveness of the method. Code can be found at \url{https://github.com/andiac/LLMOODratio}

We propose a probabilistic perspective on adversarial examples, allowing us to embed subjective understanding of semantics as a distribution into the process of generating adversarial examples, in a principled manner. Despite significant pixel-level modifications compared to traditional adversarial attacks, our method preserves the overall semantics of the image, making the changes difficult for humans to detect. This extensive pixel-level modification enhances our method's ability to deceive classifiers designed to defend against adversarial attacks. Our empirical findings indicate that the proposed methods achieve higher success rates in circumventing adversarial defense mechanisms, while remaining difficult for human observers to detect.

This paper looks at the problem of designing medium access algorithm for wireless networks with the objective of providing high throughput and low delay performance to the users, while requiring only a modest computational effort at the transmitters and receivers. Additive inter-user interference at the receivers is an important physical layer characteristic of wireless networks. Today's Wi-Fi networks are based upon the abstraction of physical layer where inter-user interference is considered as noise leading to the 'collision' model in which users are required to co-ordinate their transmissions through Carrier Sensing Multiple Access (CSMA)-based schemes to avoid interference. This, in turn, leads to an inherent performance trade-off [1]: it is impossible to obtain high throughput and low delay by means of low complexity medium access algorithm (unless P=NP). As the main result, we establish that this trade-off is primarily due to treating interference as noise in the current wireless architecture. Concretely, we develop a simple medium access algorithm that allows for simultaneous transmissions of users to the same receiver by performing joint decoding at receivers, over time. For a receiver to be able to decode multiple transmissions quickly enough, we develop appropriate congestion control where each transmitter maintains a "window" of undecoded transmitted data that is adjusted based upon the "feedback" from the receiver. In summary, this provides an efficient, low complexity "online" code operating at varying rate, and the system as a whole experiences only small amount of delay (including decoding time) while operating at high throughput.

The latest large-scale data centers offer higher aggregate bandwidth and robustness by creating multiple paths in the core of the net- work. To utilize this bandwidth requires different flows take different paths, which poses a challenge. In short, a single-path transport seems ill-suited to such networks.

Recently new data center topologies have been proposed that offer higher aggregate bandwidth and location independence by creating multiple paths in the core of the network. To effectively use this bandwidth requires ensuring different flows take different paths, which poses a challenge.

The input-queued switch architecture is widely used in Internet routers, due to its ability to run at very high line speeds. A central problem in designing an input-queued switch is choosing the scheduling algorithm, i.e. deciding which packets to transfer from ingress ports to egress ports in a given timeslot. Important metrics for evaluating a scheduling algorithm are its throughput and average delay. The well-studied ‘Maximum-Weight’ algorithm has been proved to have maximal throughput [1]; later work [2]–[4] found a wider class of algorithms which also have maximal throughput. The delay performance of these algorithms is less well understood. In this paper, we present a new technique for analysing scheduling algorithms which can explain their delay performance. In particular, we are able to explain the empirical observations in [2] about the average delay in a parameterized class of algorithms akin to Maximum-Weight. We also propose an optimal scheduling algorithm. Our technique is based on critically-balanced fluid model equations.

As the Internet grows, it is becoming increasingly difficult to collect performance measurements, to monitor its state, and to perform simulations efficiently. This is because the size and the heterogeneity of the Internet makes it time-consuming and difficult to devise traffic models and analytic tools which would allow us to work with summary statistics. We explore a method to side step these problems by combining sampling, modeling, and simulation. Our hypothesis is this: if we take a sample of the input traffic and feed it into a suitably scaled version of the system, we can extrapolate from the performance of the scaled system to that of the original. Our main findings are as follows. When we scale an IP network which is shared by short- and long-lived TCP-like and UDP flows and which is controlled by a variety of active queue management schemes, then performance measures such as queueing delay and drop probability are left virtually unchanged. We show this in theory and in simulations. This makes it possible to capture the performance of large networks quite faithfully using smaller scale replicas.

In networks and in Web server farms, it is useful to collect performance measurements, to monitor the state of the system, and to perform simulations. However, the sheer volume of traffic in large high-speed network systems makes it hard to monitor their performance or to simulate them efficiently. And the heterogeneity of the Internet means it is time-consuming and difficult to devise the traffic models and analytic tools which would allow us to work with summary statistics. We explore a method to side-step these problems by combining sampling, modeling and simulation. Our hypothesis is this: if we take a sample of the input traffic, and feed it into a suitably scaled version of the system, we can extrapolate from the performance of the scaled system to that of the original. Our main findings are: When we scale an IP network which is shared by TCP-like, UDP and Web flows; and which is controlled by a variety of active queue management schemes, then performance measures such as queueing delay and drop probability are left virtually unchanged. We show this in theory and in simulations. This makes it possible to capture the performance of large networks quite faithfully using smaller scale replicas.

Calls that make large, persistent demands for network resources will be denied consistent service, unless the network employs adequate control mechanisms. Calls of this type include video conferences. Although overprovisioning network capacity would increase the likelihood of accepting these calls, it is a very expensive option to apply uniformly in a large network, especially as the calls require high bandwidth and low blocking probabilities. Such large calls typically require coordination of geographically distributed facilities and people at the end systems. So it is natural to book the network requirements ahead of their actual use. We present a new, effective admission control algorithm for booking ahead network services. The admission control is based on a novel application of effective bandwidth theory to the time domain. Systematic and comprehensive simulation experiments provide an understanding of how booking ahead affects call blocking and network utilization, considering call duration, number of links, bandwidth, routing, and the mix of book ahead versus immediate arrival traffic. Allowing some calls to book ahead radically reduces their chance of service denial, while allowing flexible and efficient sharing of network resources with normal calls that do not book ahead.