Hongliu Cao

Hongliu Cao, Ilias Driouich, Eoin Thomas

Large Language Model (LLM)-based agents are increasingly adopted in high-stakes settings, but current benchmarks evaluate mainly whether a task was completed, not how. We introduce Procedure-Aware Evaluation (PAE), a framework that formalizes agent procedures as structured observations and exposes consistency relationships between what agents observe, communicate, and execute. PAE evaluates agents along complementary axes (Utility, Efficiency, Interaction Quality, Procedural Integrity) and applies multi-dimensional gating that categorically disqualifies corrupt outcomes. Evaluating state-of-the-art LLM agents on tau-bench yields findings at the axis, compliance, and benchmark levels. At the axis level, the dimensions capture non-redundant failure modes: utility masks reliability gaps, speed does not imply precision, and conciseness does not predict intent adherence. At the procedural compliance level, 27-78% of benchmark reported successes are corrupt successes concealing violations across interaction and integrity. Furthermore, gating substantially collapses Pass^4 rate and affects model rankings. The analysis of corrupt success cases reveals distinctive per-model failure signatures: GPT-5 spreads errors across policy, execution, and intent dimensions; Kimi-K2-Thinking concentrates 78% of violations in policy faithfulness and compliance; and Mistral-Large-3 is dominated by faithfulness failures. At the benchmark level, our analysis exposes structural flaws in the benchmark design, including task scope gaps, contradictory reward signals, and simulator artifacts that produce accidental successes.

Hongliu Cao

The rapid advancement of Language Model technologies has opened new opportunities, but also introduced new challenges related to bias and fairness. This paper explores the uncharted territory of potential biases in state-of-the-art universal text embedding models towards specific document and query writing styles within Information Retrieval (IR) systems. Our investigation reveals that different embedding models exhibit different preferences of document writing style, while more informal and emotive styles are less favored by most embedding models. In terms of query writing styles, many embedding models tend to match the style of the query with the style of the retrieved documents, but some show a consistent preference for specific styles. Text embedding models fine-tuned on synthetic data generated by LLMs display a consistent preference for certain style of generated data. These biases in text embedding based IR systems can inadvertently silence or marginalize certain communication styles, thereby posing a significant threat to fairness in information retrieval. Finally, we also compare the answer styles of Retrieval Augmented Generation (RAG) systems based on different LLMs and find out that most text embedding models are biased towards LLM's answer styles when used as evaluation metrics for answer correctness. This study sheds light on the critical issue of writing style based bias in IR systems, offering valuable insights for the development of more fair and robust models.

Hongliu Cao, Ilias Driouich, Robin Singh et al.

Large Language Models (LLMs) have demonstrated impressive performance across diverse domains, yet they still encounter challenges such as insufficient domain-specific knowledge, biases, and hallucinations. This underscores the need for robust evaluation methodologies to accurately assess LLM-based applications. Traditional evaluation methods, which rely on word overlap or text embeddings, are inadequate for capturing the nuanced semantic information necessary to evaluate dynamic, open-ended text generation. Recent research has explored leveraging LLMs to mimic human reasoning and decision-making processes for evaluation purposes known as LLM-as-a-judge framework. However, these existing frameworks have two significant limitations. First, they lack the flexibility to adapt to different text styles, including various answer and ground truth styles, thereby reducing their generalization performance. Second, the evaluation scores produced by these frameworks are often skewed and hard to interpret, showing a low correlation with human judgment. To address these challenges, we propose a novel dynamic multi-agent system that automatically designs personalized LLM judges for various natural language generation applications. This system iteratively refines evaluation prompts and balances the trade-off between the adaptive requirements of downstream tasks and the alignment with human perception. Our experimental results show that the proposed multi-agent LLM Judge framework not only enhances evaluation accuracy compared to existing methods but also produces evaluation scores that better align with human perception.

Ilias Driouich, Hongliu Cao, Eoin Thomas

Retrieval-augmented generation (RAG) systems improve large language model outputs by incorporating external knowledge, enabling more informed and context-aware responses. However, the effectiveness and trustworthiness of these systems critically depends on how they are evaluated, particularly on whether the evaluation process captures real-world constraints like protecting sensitive information. While current evaluation efforts for RAG systems have primarily focused on the development of performance metrics, far less attention has been given to the design and quality of the underlying evaluation datasets, despite their pivotal role in enabling meaningful, reliable assessments. In this work, we introduce a novel multi-agent framework for generating synthetic QA datasets for RAG evaluation that prioritize semantic diversity and privacy preservation. Our approach involves: (1) a Diversity agent leveraging clustering techniques to maximize topical coverage and semantic variability, (2) a Privacy Agent that detects and mask sensitive information across multiple domains and (3) a QA curation agent that synthesizes private and diverse QA pairs suitable as ground truth for RAG evaluation. Extensive experiments demonstrate that our evaluation sets outperform baseline methods in diversity and achieve robust privacy masking on domain-specific datasets. This work offers a practical and ethically aligned pathway toward safer, more comprehensive RAG system evaluation, laying the foundation for future enhancements aligned with evolving AI regulations and compliance standards.

Hongliu Cao, Minh Nhat Do, Alexis Ravanel et al.

Face recognition has been used more and more in real world applications in recent years. However, when the skin color bias is coupled with intra-personal variations like harsh illumination, the face recognition task is more likely to fail, even during human inspection. Face normalization methods try to deal with such challenges by removing intra-personal variations from an input image while keeping the identity the same. However, most face normalization methods can only remove one or two variations and ignore dataset biases such as skin color bias. The outputs of many face normalization methods are also not realistic to human observers. In this work, a style based face normalization model (StyleFNM) is proposed to remove most intra-personal variations including large changes in pose, bad or harsh illumination, low resolution, blur, facial expressions, and accessories like sunglasses among others. The dataset bias is also dealt with in this paper by controlling a pretrained GAN to generate a balanced dataset of passport-like images. The experimental results show that StyleFNM can generate more realistic outputs and can improve significantly the accuracy and fairness of face recognition systems.

Hongliu Cao, Ilias El Baamrani, Eoin Thomas

As the digitization of travel industry accelerates, analyzing and understanding travelers' behaviors becomes increasingly important. However, traveler data frequently exhibit high data sparsity due to the relatively low frequency of user interactions with travel providers. Compounding this effect the multiplication of devices, accounts and platforms while browsing travel products online also leads to data dispersion. To deal with these challenges, probabilistic traveler matching can be used. Most existing solutions for user matching are not suitable for traveler matching as a traveler's browsing history is typically short and URLs in the travel industry are very heterogeneous with many tokens. To deal with these challenges, we propose the similarity based multi-view information fusion to learn a better user representation from URLs by treating the URLs as multi-view data. The experimental results show that the proposed multi-view user representation learning can take advantage of the complementary information from different views, highlight the key information in URLs and perform significantly better than other representation learning solutions for the user matching task.

Hongliu Cao

Negation plays an important role in various natural language processing tasks such as Natural Language Inference and Sentiment Analysis tasks. Numerous prior studies have found that contextual text embedding models such as BERT, ELMO, RoBERTa or XLNet face challenges in accurately understanding negation. Recent advancements in universal text embeddings have demonstrated superior performance over contextual text embeddings in various tasks. However, due to the bias in popular evaluation benchmarks, the negation awareness capacity of these models remains unclear. To bridge the gap in existing literature, an in-depth analysis is initiated in this work to study the negation awareness of cutting-edge universal text embedding models. Our findings reveal a significant lack of negation awareness in these models, often interpreting negated text pairs as semantically similar. To efficiently deal with the conflict that different tasks need different trade-offs between topic and negation information among other semantic information, a data-efficient and computational-efficient embedding re-weighting method is proposed without modifying the parameters of text embedding models. The proposed solution is able to improve text embedding models' negation awareness significantly on both simple negation understanding task and complex negation understanding task. Furthermore, the proposed solution can also significantly improve the negation awareness of Large Language Model based task-specific high dimensional universal text embeddings.

Hongliu Cao

In light of emerging legal requirements and policies focused on privacy protection, there is a growing trend of companies across various industries adopting Federated Learning (FL). This decentralized approach involves multiple clients or silos, collaboratively training a global model under the coordination of a central server while utilizing their private local data. Unlike traditional methods that necessitate data sharing and transmission, Cross-Silo FL allows clients to share model updates rather than raw data, thereby enhancing privacy. Despite its growing adoption, the carbon impact associated with Cross-Silo FL remains poorly understood due to the limited research in this area. This study seeks to bridge this gap by evaluating the sustainability of Cross-Silo FL throughout the entire AI product lifecycle, extending the analysis beyond the model training phase alone. We systematically compare this decentralized method with traditional centralized approaches and present a robust quantitative framework for assessing the costs and CO2 emissions in real-world Cross-Silo FL environments. Our findings indicate that the energy consumption and costs of model training are comparable between Cross-Silo Federated Learning and Centralized Learning. However, the additional data transfer and storage requirements inherent in Centralized Learning can result in significant, often overlooked CO2 emissions. Moreover, we introduce an innovative data and application management system that integrates Cross-Silo FL and analytics, aiming at improving the sustainability and economic efficiency of IT enterprises.

Hongliu Cao, Eoin Thomas

With the digitization of travel industry, it is more and more important to understand users from their online behaviors. However, online travel industry data are more challenging to analyze due to extra sparseness, dispersed user history actions, fast change of user interest and lack of direct or indirect feedbacks. In this work, a new similarity method is proposed to measure the destination similarity in terms of implicit user interest. By comparing the proposed method to several other widely used similarity measures in recommender systems, the proposed method achieves a significant improvement on travel data. Key words: Destination similarity, Travel industry, Recommender System, Implicit user interest

Simon Bernard, Hongliu Cao, Robert Sabourin et al.

Many classification problems are naturally multi-view in the sense their data are described through multiple heterogeneous descriptions. For such tasks, dissimilarity strategies are effective ways to make the different descriptions comparable and to easily merge them, by (i) building intermediate dissimilarity representations for each view and (ii) fusing these representations by averaging the dissimilarities over the views. In this work, we show that the Random Forest proximity measure can be used to build the dissimilarity representations, since this measure reflects similarities between features but also class membership. We then propose a Dynamic View Selection method to better combine the view-specific dissimilarity representations. This allows to take a decision, on each instance to predict, with only the most relevant views for that instance. Experiments are conducted on several real-world multi-view datasets, and show that the Dynamic View Selection offers a significant improvement in performance compared to the simple average combination and two state-of-the-art static view combinations.

Hongliu Cao, Simon Bernard, Robert Sabourin et al.

Multi-view learning is a learning task in which data is described by several concurrent representations. Its main challenge is most often to exploit the complementarities between these representations to help solve a classification/regression task. This is a challenge that can be met nowadays if there is a large amount of data available for learning. However, this is not necessarily true for all real-world problems, where data are sometimes scarce (e.g. problems related to the medical environment). In these situations, an effective strategy is to use intermediate representations based on the dissimilarities between instances. This work presents new ways of constructing these dissimilarity representations, learning them from data with Random Forest classifiers. More precisely, two methods are proposed, which modify the Random Forest proximity measure, to adapt it to the context of High Dimension Low Sample Size (HDLSS) multi-view classification problems. The second method, based on an Instance Hardness measurement, is significantly more accurate than other state-of-the-art measurements including the original RF Proximity measurement and the Large Margin Nearest Neighbor (LMNN) metric learning measurement.

Hongliu Cao, Simon Bernard, Laurent Heutte et al.

Cancer diagnosis and treatment often require a personalized analysis for each patient nowadays, due to the heterogeneity among the different types of tumor and among patients. Radiomics is a recent medical imaging field that has shown during the past few years to be promising for achieving this personalization. However, a recent study shows that most of the state-of-the-art works in Radiomics fail to identify this problem as a multi-view learning task and that multi-view learning techniques are generally more efficient. In this work, we propose to further investigate the potential of one family of multi-view learning methods based on Multiple Classifiers Systems where one classifier is learnt on each view and all classifiers are combined afterwards. In particular, we propose a random forest based dynamic weighted voting scheme, which personalizes the combination of views for each new patient for classification tasks. The proposed method is validated on several real-world Radiomics problems.

Hongliu Cao, Simon Bernard, Laurent Heutte et al.

Breast cancer is one of the most common types of cancer and leading cancer-related death causes for women. In the context of ICIAR 2018 Grand Challenge on Breast Cancer Histology Images, we compare one handcrafted feature extractor and five transfer learning feature extractors based on deep learning. We find out that the deep learning networks pretrained on ImageNet have better performance than the popular handcrafted features used for breast cancer histology images. The best feature extractor achieves an average accuracy of 79.30%. To improve the classification performance, a random forest dissimilarity based integration method is used to combine different feature groups together. When the five deep learning feature groups are combined, the average accuracy is improved to 82.90% (best accuracy 85.00%). When handcrafted features are combined with the five deep learning feature groups, the average accuracy is improved to 87.10% (best accuracy 93.00%).

Hongliu Cao, Simon Bernard, Laurent Heutte et al.

Radiomics is a term which refers to the analysis of the large amount of quantitative tumor features extracted from medical images to find useful predictive, diagnostic or prognostic information. Many recent studies have proved that radiomics can offer a lot of useful information that physicians cannot extract from the medical images and can be associated with other information like gene or protein data. However, most of the classification studies in radiomics report the use of feature selection methods without identifying the machine learning challenges behind radiomics. In this paper, we first show that the radiomics problem should be viewed as an high dimensional, low sample size, multi view learning problem, then we compare different solutions proposed in multi view learning for classifying radiomics data. Our experiments, conducted on several real world multi view datasets, show that the intermediate integration methods work significantly better than filter and embedded feature selection methods commonly used in radiomics.