The proliferation of Large Language Models (LLMs) with varying capabilities and costs has created a need for efficient model selection in AI systems. LLM routers address this need by dynamically choosing the most suitable model for a given query based on task requirements and budget constraints. However, existing routers face challenges in scalability and real-time adaptation, particularly in high-volume online environments. We present Eagle, a novel LLM routing approach that combines global and local ELO ranking modules to overcome these limitations. By evaluating both general and specialized LLM abilities, Eagle provides a scalable, training-free solution that enhances model selection quality while reducing computational overhead. Our experiments across multiple datasets show Eagle consistently outperforms baseline methods, with improvements of up to 23.52 percent in Area Under Curve (AUC) scores. Moreover, Eagle demonstrates remarkable efficiency, requiring only 1/20 of baseline methods' time for initialization and 100 to 200 times faster incremental updates in online scenarios, making it well-suited for dynamic, high-volume online serving environments.
Text ranking has witnessed significant advancements, attributed to the utilization of dual-encoder enhanced by Pre-trained Language Models (PLMs). Given the proliferation of available PLMs, selecting the most effective one for a given dataset has become a non-trivial challenge. As a promising alternative to human intuition and brute-force fine-tuning, Transferability Estimation (TE) has emerged as an effective approach to model selection. However, current TE methods are primarily designed for classification tasks, and their estimated transferability may not align well with the objectives of text ranking. To address this challenge, we propose to compute the expected rank as transferability, explicitly reflecting the model's ranking capability. Furthermore, to mitigate anisotropy and incorporate training dynamics, we adaptively scale isotropic sentence embeddings to yield an accurate expected rank score. Our resulting method, Adaptive Ranking Transferability (AiRTran), can effectively capture subtle differences between models. On challenging model selection scenarios across various text ranking datasets, it demonstrates significant improvements over previous classification-oriented TE methods, human intuition, and ChatGPT with minor time consumption.
Common workflows in machine learning and statistics rely on the ability to partition the information in a data set into independent portions. Recent work has shown that this may be possible even when conventional sample splitting is not (e.g., when the number of samples $n=1$, or when observations are not independent and identically distributed). However, the approaches that are currently available to decompose multivariate Gaussian data require knowledge of the covariance matrix. In many important problems (such as in spatial or longitudinal data analysis, and graphical modeling), the covariance matrix may be unknown and even of primary interest. Thus, in this work we develop new approaches to decompose Gaussians with unknown covariance. First, we present a general algorithm that encompasses all previous decomposition approaches for Gaussian data as special cases, and can further handle the case of an unknown covariance. It yields a new and more flexible alternative to sample splitting when $n>1$. When $n=1$, we prove that it is impossible to partition the information in a multivariate Gaussian into independent portions without knowing the covariance matrix. Thus, we use the general algorithm to decompose a single multivariate Gaussian with unknown covariance into dependent parts with tractable conditional distributions, and demonstrate their use for inference and validation. The proposed decomposition strategy extends naturally to Gaussian processes. In simulation and on electroencephalography data, we apply these decompositions to the tasks of model selection and post-selection inference in settings where alternative strategies are unavailable.
The task of determining whether two texts are paraphrases has long been a challenge in NLP. However, the prevailing notion of paraphrase is often quite simplistic, offering only a limited view of the vast spectrum of paraphrase phenomena. Indeed, we find that evaluating models in a paraphrase dataset can leave uncertainty about their true semantic understanding. To alleviate this, we create PARAPHRASUS, a benchmark designed for multi-dimensional assessment, benchmarking and selection of paraphrase detection models. We find that paraphrase detection models under our fine-grained evaluation lens exhibit trade-offs that cannot be captured through a single classification dataset. Furthermore, PARAPHRASUS allows prompt calibration for different use cases, tailoring LLM models to specific strictness levels. PARAPHRASUS includes 3 challenges spanning over 10 datasets, including 8 repurposed and 2 newly annotated; we release it along with a benchmarking library at https://github.com/impresso/paraphrasus
Purpose: To develop and evaluate an automated system for extracting structured clinical information from unstructured radiology and pathology reports using open-weights large language models (LMs) and retrieval augmented generation (RAG), and to assess the effects of model configuration variables on extraction performance. Methods and Materials: The study utilized two datasets: 7,294 radiology reports annotated for Brain Tumor Reporting and Data System (BT-RADS) scores and 2,154 pathology reports annotated for isocitrate dehydrogenase (IDH) mutation status. An automated pipeline was developed to benchmark the performance of various LMs and RAG configurations. The impact of model size, quantization, prompting strategies, output formatting, and inference parameters was systematically evaluated. Results: The best performing models achieved over 98% accuracy in extracting BT-RADS scores from radiology reports and over 90% for IDH mutation status extraction from pathology reports. The top model being medical fine-tuned llama3. Larger, newer, and domain fine-tuned models consistently outperformed older and smaller models. Model quantization had minimal impact on performance. Few-shot prompting significantly improved accuracy. RAG improved performance for complex pathology reports but not for shorter radiology reports. Conclusions: Open LMs demonstrate significant potential for automated extraction of structured clinical data from unstructured clinical reports with local privacy-preserving application. Careful model selection, prompt engineering, and semi-automated optimization using annotated data are critical for optimal performance. These approaches could be reliable enough for practical use in research workflows, highlighting the potential for human-machine collaboration in healthcare data extraction.
We propose a novel approach to select the best model of the data. Based on the exclusive properties of the nested models, we find the most parsimonious model containing the risk minimizer predictor. We prove the existence of probable approximately correct (PAC) bounds on the difference of the minimum empirical risk of two successive nested models, called successive empirical excess risk (SEER). Based on these bounds, we propose a model order selection method called nested empirical risk (NER). By the sorted NER (S-NER) method to sort the models intelligently, the minimum risk decreases. We construct a test that predicts whether expanding the model decreases the minimum risk or not. With a high probability, the NER and S-NER choose the true model order and the most parsimonious model containing the risk minimizer predictor, respectively. We use S-NER model selection in the linear regression and show that, the S-NER method without any prior information can outperform the accuracy of feature sorting algorithms like orthogonal matching pursuit (OMP) that aided with prior knowledge of the true model order. Also, in the UCR data set, the NER method reduces the complexity of the classification of UCR datasets dramatically, with a negligible loss of accuracy.
Conventional WENO3 methods are known to be highly dissipative at lower resolutions, introducing significant errors in the pre-asymptotic regime. In this paper, we employ a rational neural network to accurately estimate the local smoothness of the solution, dynamically adapting the stencil weights based on local solution features. As rational neural networks can represent fast transitions between smooth and sharp regimes, this approach achieves a granular reconstruction with significantly reduced dissipation, improving the accuracy of the simulation. The network is trained offline on a carefully chosen dataset of analytical functions, bypassing the need for differentiable solvers. We also propose a robust model selection criterion based on estimates of the interpolation's convergence order on a set of test functions, which correlates better with the model performance in downstream tasks. We demonstrate the effectiveness of our approach on several one-, two-, and three-dimensional fluid flow problems: our scheme generalizes across grid resolutions while handling smooth and discontinuous solutions. In most cases, our rational network-based scheme achieves higher accuracy than conventional WENO3 with the same stencil size, and in a few of them, it achieves accuracy comparable to WENO5, which uses a larger stencil.
Recent works for time-series forecasting more and more leverage the high predictive power of Deep Learning models. With this increase in model complexity, however, comes a lack in understanding of the underlying model decision process, which is problematic for high-stakes application scenarios. At the same time, simple, interpretable forecasting methods such as ARIMA still perform very well, sometimes on-par, with Deep Learning approaches. We argue that simple models are good enough most of the time, and that forecasting performance could be improved by choosing a Deep Learning method only for few, important predictions, increasing the overall interpretability of the forecasting process. In this context, we propose a novel online model selection framework which learns to identify these predictions. An extensive empirical study on various real-world datasets shows that our selection methodology performs comparable to state-of-the-art online model selections methods in most cases while being significantly more interpretable. We find that almost always choosing a simple autoregressive linear model for forecasting results in competitive performance, suggesting that the need for opaque black-box models in time-series forecasting might be smaller than recent works would suggest.
Visual Question-Answering (VQA) has become key to user experience, particularly after improved generalization capabilities of Vision-Language Models (VLMs). But evaluating VLMs for an application requirement using a standardized framework in practical settings is still challenging. This paper aims to solve that using an end-to-end framework. We present VQA360 - a novel dataset derived from established VQA benchmarks, annotated with task types, application domains, and knowledge types, for a comprehensive evaluation. We also introduce GoEval, a multimodal evaluation metric developed using GPT-4o, achieving a correlation factor of 56.71% with human judgments. Our experiments with state-of-the-art VLMs reveal that no single model excels universally, thus, making a right choice a key design decision. Proprietary models such as Gemini-1.5-Pro and GPT-4o-mini generally outperform others, but open-source models like InternVL-2-8B and CogVLM-2-Llama-3-19B also demonstrate competitive strengths, while providing additional advantages. Our framework can also be extended to other tasks.
The normalized maximum likelihood (NML) code length is widely used as a model selection criterion based on the minimum description length principle, where the model with the shortest NML code length is selected. A common method to calculate the NML code length is to use the sum (for a discrete model) or integral (for a continuous model) of a function defined by the distribution of the maximum likelihood estimator. While this method has been proven to correctly calculate the NML code length of discrete models, no proof has been provided for continuous cases. Consequently, it has remained unclear whether the method can accurately calculate the NML code length of continuous models. In this paper, we solve this problem affirmatively, proving that the method is also correct for continuous cases. Remarkably, completing the proof for continuous cases is non-trivial in that it cannot be achieved by merely replacing the sums in discrete cases with integrals, as the decomposition trick applied to sums in the discrete model case proof is not applicable to integrals in the continuous model case proof. To overcome this, we introduce a novel decomposition approach based on the coarea formula from geometric measure theory, which is essential to establishing our proof for continuous cases.
Ensemble learning is a meta-learning approach that combines the predictions of multiple learners, demonstrating improved accuracy and robustness. Nevertheless, ensembling models like Convolutional Neural Networks (CNNs) result in high memory and computing overhead, preventing their deployment in embedded systems. These devices are usually equipped with small batteries that provide power supply and might include energy-harvesting modules that extract energy from the environment. In this work, we propose E-QUARTIC, a novel Energy Efficient Edge Ensembling framework to build ensembles of CNNs targeting Artificial Intelligence (AI)-based embedded systems. Our design outperforms single-instance CNN baselines and state-of-the-art edge AI solutions, improving accuracy and adapting to varying energy conditions while maintaining similar memory requirements. Then, we leverage the multi-CNN structure of the designed ensemble to implement an energy-aware model selection policy in energy-harvesting AI systems. We show that our solution outperforms the state-of-the-art by reducing system failure rate by up to 40% while ensuring higher average output qualities. Ultimately, we show that the proposed design enables concurrent on-device training and high-quality inference execution at the edge, limiting the performance and energy overheads to less than 0.04%.
Characterising the intelligence of biological organisms is challenging. This work considers intelligent algorithms developed evolutionarily within neural systems. Mathematical analyses unveil a natural equivalence between canonical neural networks, variational Bayesian inference under a class of partially observable Markov decision processes, and differentiable Turing machines, by showing that they minimise the shared Helmholtz energy. Consequently, canonical neural networks can biologically plausibly equip Turing machines and conduct variational Bayesian inferences of external Turing machines in the environment. Applying Helmholtz energy minimisation at the species level facilitates deriving active Bayesian model selection inherent in natural selection, resulting in the emergence of adaptive algorithms. In particular, canonical neural networks with two mental actions can separately memorise transition mappings of multiple external Turing machines to form a universal machine. These propositions were corroborated by numerical simulations of algorithm implementation and neural network evolution. These notions offer a universal characterisation of biological intelligence emerging from evolution in terms of Bayesian model selection and belief updating.
Federated Learning (FL) enables training of a global model from distributed data, while preserving data privacy. However, the singular-model based operation of FL is open with uploading poisoned models compatible with the global model structure and can be exploited as a vulnerability to conduct model poisoning attacks. This paper proposes a multi-model based FL as a proactive mechanism to enhance the opportunity of model poisoning attack mitigation. A master model is trained by a set of slave models. To enhance the opportunity of attack mitigation, the structure of client models dynamically change within learning epochs, and the supporter FL protocol is provided. For a MEC system, the model selection problem is modeled as an optimization to minimize loss and recognition time, while meeting a robustness confidence. In adaption with dynamic network condition, a deep reinforcement learning based model selection is proposed. For a DDoS attack detection scenario, results illustrate a competitive accuracy gain under poisoning attack with the scenario that the system is without attack, and also a potential of recognition time improvement.
The rapid evolution of cyber threats necessitates innovative solutions for detecting and analyzing malicious activity. Honeypots, which are decoy systems designed to lure and interact with attackers, have emerged as a critical component in cybersecurity. In this paper, we present a novel approach to creating realistic and interactive honeypot systems using Large Language Models (LLMs). By fine-tuning a pre-trained open-source language model on a diverse dataset of attacker-generated commands and responses, we developed a honeypot capable of sophisticated engagement with attackers. Our methodology involved several key steps: data collection and processing, prompt engineering, model selection, and supervised fine-tuning to optimize the model's performance. Evaluation through similarity metrics and live deployment demonstrated that our approach effectively generates accurate and informative responses. The results highlight the potential of LLMs to revolutionize honeypot technology, providing cybersecurity professionals with a powerful tool to detect and analyze malicious activity, thereby enhancing overall security infrastructure.
The rapid development of Large Language Models (LLMs) for healthcare applications has spurred calls for holistic evaluation beyond frequently-cited benchmarks like USMLE, to better reflect real-world performance. While real-world assessments are valuable indicators of utility, they often lag behind the pace of LLM evolution, likely rendering findings obsolete upon deployment. This temporal disconnect necessitates a comprehensive upfront evaluation that can guide model selection for specific clinical applications. We introduce MEDIC, a framework assessing LLMs across five critical dimensions of clinical competence: medical reasoning, ethics and bias, data and language understanding, in-context learning, and clinical safety. MEDIC features a novel cross-examination framework quantifying LLM performance across areas like coverage and hallucination detection, without requiring reference outputs. We apply MEDIC to evaluate LLMs on medical question-answering, safety, summarization, note generation, and other tasks. Our results show performance disparities across model sizes, baseline vs medically finetuned models, and have implications on model selection for applications requiring specific model strengths, such as low hallucination or lower cost of inference. MEDIC's multifaceted evaluation reveals these performance trade-offs, bridging the gap between theoretical capabilities and practical implementation in healthcare settings, ensuring that the most promising models are identified and adapted for diverse healthcare applications.
Outlier detection (OD) has a vast literature as it finds numerous applications in environmental monitoring, cybersecurity, finance, and medicine to name a few. Being an inherently unsupervised task, model selection is a key bottleneck for OD (both algorithm and hyperparameter selection) without label supervision. There is a long list of techniques to choose from -- both classical algorithms and deep neural architectures -- and while several studies report their hyperparameter sensitivity, the literature is quite slim on unsupervised model selection -- limiting the effective use of OD in practice. In this paper we present FoMo-0D, for zero/0-shot OD exploring a transformative new direction that bypasses the hurdle of model selection altogether (!), thus breaking new ground. The fundamental idea behind FoMo-0D is the Prior-data Fitted Networks, recently introduced by Muller et al.(2022), which trains a Transformer model on a large body of synthetically generated data from a prior data distribution. In essence, FoMo-0D is a pretrained Foundation Model for zero/0-shot OD on tabular data, which can directly predict the (outlier/inlier) label of any test data at inference time, by merely a single forward pass -- making obsolete the need for choosing an algorithm/architecture, tuning its associated hyperparameters, and even training any model parameters when given a new OD dataset. Extensive experiments on 57 public benchmark datasets against 26 baseline methods show that FoMo-0D performs statistically no different from the top 2nd baseline, while significantly outperforming the majority of the baselines, with an average inference time of 7.7 ms per test sample.
Eating disorders are complex mental health conditions that affect millions of people around the world. Effective interventions on social media platforms are crucial, yet testing strategies in situ can be risky. We present a novel LLM-driven experimental testbed for simulating and assessing intervention strategies in ED-related discussions. Our framework generates synthetic conversations across multiple platforms, models, and ED-related topics, allowing for controlled experimentation with diverse intervention approaches. We analyze the impact of various intervention strategies on conversation dynamics across four dimensions: intervention type, generative model, social media platform, and ED-related community/topic. We employ cognitive domain analysis metrics, including sentiment, emotions, etc., to evaluate the effectiveness of interventions. Our findings reveal that civility-focused interventions consistently improve positive sentiment and emotional tone across all dimensions, while insight-resetting approaches tend to increase negative emotions. We also uncover significant biases in LLM-generated conversations, with cognitive metrics varying notably between models (Claude-3 Haiku $>$ Mistral $>$ GPT-3.5-turbo $>$ LLaMA3) and even between versions of the same model. These variations highlight the importance of model selection in simulating realistic discussions related to ED. Our work provides valuable information on the complex dynamics of ED-related discussions and the effectiveness of various intervention strategies.
The rapid evolution of mobile networks from 5G to 6G has necessitated the development of autonomous network management systems, such as Zero-Touch Networks (ZTNs). However, the increased complexity and automation of these networks have also escalated cybersecurity risks. Existing Intrusion Detection Systems (IDSs) leveraging traditional Machine Learning (ML) techniques have shown effectiveness in mitigating these risks, but they often require extensive manual effort and expert knowledge. To address these challenges, this paper proposes an Automated Machine Learning (AutoML)-based autonomous IDS framework towards achieving autonomous cybersecurity for next-generation networks. To achieve autonomous intrusion detection, the proposed AutoML framework automates all critical procedures of the data analytics pipeline, including data pre-processing, feature engineering, model selection, hyperparameter tuning, and model ensemble. Specifically, it utilizes a Tabular Variational Auto-Encoder (TVAE) method for automated data balancing, tree-based ML models for automated feature selection and base model learning, Bayesian Optimization (BO) for hyperparameter optimization, and a novel Optimized Confidence-based Stacking Ensemble (OCSE) method for automated model ensemble. The proposed AutoML-based IDS was evaluated on two public benchmark network security datasets, CICIDS2017 and 5G-NIDD, and demonstrated improved performance compared to state-of-the-art cybersecurity methods. This research marks a significant step towards fully autonomous cybersecurity in next-generation networks, potentially revolutionizing network security applications.
Digital Twin (DT) technology is expected to play a pivotal role in NextG wireless systems. However, a key challenge remains in the evaluation of data-driven algorithms within DTs, particularly the transfer of learning from simulations to real-world environments. In this work, we investigate the sim-to-real gap in developing a digital twin for the NSF PAWR Platform, POWDER. We first develop a 3D model of the University of Utah campus, incorporating geographical measurements and all rooftop POWDER nodes. We then assess the accuracy of various path loss models used in training modeling and control policies, examining the impact of each model on sim-to-real link performance predictions. Finally, we discuss the lessons learned from model selection and simulation design, offering guidance for the implementation of DT-enabled wireless networks.
This research introduces an innovative method for the early screening of cardiorespiratory diseases based on an acquisition protocol, which leverages commodity smartphone's Inertial Measurement Units (IMUs) and deep learning techniques. We collected, in a clinical setting, a dataset featuring recordings of breathing kinematics obtained by accelerometer and gyroscope readings from five distinct body regions. We propose an end-to-end deep learning pipeline for early cardiorespiratory disease screening, incorporating a preprocessing step segmenting the data into individual breathing cycles, and a recurrent bidirectional module capturing features from diverse body regions. We employed Leave-one-out-cross-validation with Bayesian optimization for hyperparameter tuning and model selection. The experimental results consistently demonstrated the superior performance of a bidirectional Long-Short Term Memory (Bi-LSTM) as a feature encoder architecture, yielding an average sensitivity of $0.81 \pm 0.02$, specificity of $0.82 \pm 0.05$, F1 score of $0.81 \pm 0.02$, and accuracy of $80.2\% \pm 3.9$ across diverse seed variations. We also assessed generalization capabilities on a skewed distribution, comprising exclusively healthy patients not used in training, revealing a true negative rate of $74.8 \% \pm 4.5$. The sustained accuracy of predictions over time during breathing cycles within a single patient underscores the efficacy of the preprocessing strategy, highlighting the model's ability to discern significant patterns throughout distinct phases of the respiratory cycle. This investigation underscores the potential usefulness of widely available smartphones as devices for timely cardiorespiratory disease screening in the general population, in at-home settings, offering crucial assistance to public health efforts (especially during a pandemic outbreaks, such as the recent COVID-19).