We present AgentSeer, an interactive observability framework for agentic AI systems. Unlike conventional tracing tools that expose raw spans or model-centric metrics, AgentSeer introduces a dual graph decomposition constructed through a deterministic rule-based parser: a temporal action graph, where each prompt or tool invocation is represented as a distinct action, and a component graph capturing architectural relations among agents, tools, and memory modules. Beyond visualization, AgentSeer enables action-level red teaming, where jailbreak payloads are systematically attached to every action node (including agent messages, tool calls, and memory retrievals) to uncover vulnerabilities invisible to model-level testing. Our demonstration features a six-agent hierarchical testbed with interactive visualization and deployment-oriented safety evaluation applied directly on the same prompts and contexts, systematically revealing high-risk interactions, context-dependent vulnerabilities, and emergent behaviors. By combining structured decomposition, automated red teaming, and rule-based reliability, AgentSeer establishes a safety-first methodology for observability in multi-agent AI.

Modern Agentic AI systems plan, reason, and act across multiple steps, creating execution patterns that are difficult to interpret. Existing observability platforms track prompt I/O and operational metrics but require manual inspection of traces to reconstruct structure and reasoning. We present AgentGraph, which converts execution logs into interactive knowledge graphs and actionable insights. Nodes represent agents, tasks, tools, data inputs/outputs, and humans, while typed edges capture relations such as inputs consumed, tasks delegated or sequenced, tools required or used, outputs produced and delivered, and interventions from agents or humans. Each graph element links to its exact trace span, ensuring verifiability. Building on this representation, AgentGraph enables two analyses: qualitative trace-grounded failure detection and optimisation recommendations, and quantitative robustness evaluation via perturbation testing and causal attribution.

The development of unbiased large language models is widely recognized as crucial, yet existing benchmarks fall short in detecting biases due to limited scope, contamination, and lack of a fairness baseline. SAGED(bias) is the first holistic benchmarking pipeline to address these problems. The pipeline encompasses five core stages: scraping materials, assembling benchmarks, generating responses, extracting numeric features, and diagnosing with disparity metrics. SAGED includes metrics for max disparity, such as impact ratio, and bias concentration, such as Max Z-scores. Noticing that metric tool bias and contextual bias in prompts can distort evaluation, SAGED implements counterfactual branching and baseline calibration for mitigation. For demonstration, we use SAGED on G20 Countries with popular 8b-level models including Gemma2, Llama3.1, Mistral, and Qwen2. With sentiment analysis, we find that while Mistral and Qwen2 show lower max disparity and higher bias concentration than Gemma2 and Llama3.1, all models are notably biased against countries like Russia and (except for Qwen2) China. With further experiments to have models role-playing U.S. presidents, we see bias amplifies and shifts in heterogeneous directions. Moreover, we see Qwen2 and Mistral not engage in role-playing, while Llama3.1 and Gemma2 role-play Trump notably more intensively than Biden and Harris, indicating role-playing performance bias in these models.

Abstract This work presents a new neuro-evolutionary model, called NEVE (Neuroevolutionary Ensemble), based on an ensemble of Multi-Layer Perceptron (MLP) neural networks for learning in nonstationary environments. NEVE makes use of quantum-inspired evolutionary models to automatically configure the ensemble members and combine their output. The quantum-inspired evolutionary models identify the most appropriate topology for each MLP network, select the most relevant input variables, determine the neural network weights and calculate the voting weight of each ensemble member. Four different approaches of NEVE are developed, varying the mechanism for detecting and treating concepts drifts, including proactive drift detection approaches. The proposed models were evaluated in real and artificial datasets, comparing the results obtained with other consolidated models in the literature. The results show that the accuracy of NEVE is higher in most cases and the best configurations are obtained using some mechanism for drift detection. These results reinforce that the neuroevolutionary ensemble approach is a robust choice for situations in which the datasets are subject to sudden changes in behaviour.

Genetic programming techniques allow flexibility in the optimization process, making it possible to use them in different areas of knowledge and providing new ways for specialists to advance in their areas more quickly and more accurately. This work presents a preliminary analysis of a new method using multi-gene genetic programming to multimodal numerical optimization. The new algorithm uses some niching techniques based on the clear procedure to maintain the population diversity, in order to perform a multivariate mapping among initial guesses to optimal parameters for a system. We used a multi-modal benchmark set with different characteristics and difficulty levels to evaluate this new algorithm. Statistical analysis suggested that this new multi-modal method using multi-gene genetic programming can be used for problems that requires more than a single solution.

Derivative traders are usually required to scan through hundreds, even thousands of possible trades on a daily-basis; a concrete case is the so-called Mid-Curve Calendar Spread (MCCS). The actual procedure in place is full of pitfalls and a more systematic approach where more information at hand is crossed and aggregated to find good trading picks can be highly useful and undoubtedly increase the trader's productivity. Therefore, in this work we propose an MCCS Recommendation System based on a stacking approach through Neural Networks. In order to suggest that such approach is methodologically and computationally feasible, we used a list of 15 different types of US Dollar MCCSs regarding expiration, forward and swap tenure. For each MCCS, we used 10 years of historical data ranging weekly from Sep/06 to Sep/16. Then, we started the modelling stage by: (i) fitting the base learners using as the input sensitivity metrics linked with the MCCS at time t, and its subsequent annualized returns as the output; (ii) feeding the prediction from each base model to a particular stacker; and (iii) making predictions and comparing different modelling methodologies by a set of performance metrics and benchmarks. After establishing a backtesting engine and setting performance metrics, our results suggest that our proposed Neural Network stacker compared favourably to other combination procedures.

A time series is the most commonly used representation for the evolution of a given variable over time. In a time series forecasting problem, a model aims at predicting the series' future values, assuming that all information needed to do so is contained in the series' past behavior. Since the phenomena described by the time series does not always exist in isolation, it is possible to enhance the model with historical data from other related time series. The structure formed by several different time series occurring in parallel, each featuring the same interval and dimension, is called a multivariate time series. This paper presents a methodology for the generation of a Fuzzy Inference System (FIS) for multivariate time series forecasting from historical data, aiming at good performance in both forecasting accuracy and rule base interpretability - in order to extract knowledge about the relationship between the modeled time series. Several aspects related to the operation and construction of such a FIS are investigated regarding complexity and semantic clarity. The model is evaluated by applying it to multivariate time series obtained from the complete M3 competition database and by comparing it to other methods in terms of accuracy. In addition knowledge extraction possibilities from the resulting rule base are explored.

This work presents a novel genetic fuzzy system for forecasting, called Genetic Programming Fuzzy Inference System for Forecasting problems (GPFIS-Forecast), which generates an interpretable fuzzy rule base by using Multi-Gene Genetic Programming to define the premises terms of fuzzy rules. The main differences between GPFIS-Forecast and other genetic fuzzy systems lie in its fuzzy inference process, because it: (i) enables premises to be include negation, t-conorm and linguistic hedge operators; (ii) applies methods to define a consequent term more compatible with a given premise; and (iii) makes use of aggregation operators to weigh fuzzy rules in accordance with their influence on the problem. GPFIS-Forecast has been tested in the NN3 Competition, in order to evaluate its performance in a benchmark problem. In this case, it has produced competitive results when compared to other forecasting approaches.

Most drift detection mechanisms designed for classification problems works in a post-event manner: after receiving the data set completely (patterns and class labels of the train and test set), they apply a sequence of procedures to identify some change in the class-conditional distribution - a concept drift. However, detecting changes after its occurrence can be in some situations harmful for the process under supervision. This paper proposes a pre-event approach for abrupt drift detection, called by A2D2. Briefly, this method is composed of three steps: (i) label the patterns from the test set, using an unsupervised method; (ii) compute some statistics from the train and test set, conditioned on the given class labels; and (iii) compare the train and test statistics using a multivariate hypothesis test. Also, it has been proposed a procedure for creating datasets with abrupt drift. This procedure was used in the sensivity analysis of A2D2, in order to understand the influence degree of each parameter on its final performance.

This work presents a Genetic Fuzzy Controller (GFC), called Genetic Programming Fuzzy Inference System for Control tasks (GPFIS-Control). It is based on Multi-Gene Genetic Programming, a variant of canonical Genetic Programming. The main characteristics and concepts of this approach are described, as well as its distinctions from other GFCs. Two benchmarks application of GPFIS-Control are considered: the Cart-Centering Problem and the Inverted Pendulum. In both cases results demonstrate the superiority and potentialities of GPFIS-Control in relation to other GFCs found in the literature.

One of the most important issues in oil \& gas industry is the lithological identification. Lithology is the macroscopic description of the physical characteristics of a rock. This work proposes a new methodology for lithological discrimination, using GPF-CLASS model (Genetic Programming for Fuzzy Classification) a Genetic Fuzzy System based on Multi-Gene Genetic Programming. The main advantage of our approach is the possibility to identify, through seismic patterns, the rock types in new regions without requiring opening wells. Thus, we seek for a reliable model that provides two flexibilities for the experts: evaluate the membership degree of a seismic pattern to the several rock types and the chance to analyze at linguistic level the model output. Therefore, the final tool must afford knowledge discovery and support to the decision maker. Also, we evaluate other 7 classification models (from statistics and computational intelligence), using a database from a well located in Brazilian coast. The results demonstrate the potentialities of GPF-CLASS model when comparing to other classifiers.

In our previous works [1, 2], we proposed NEVE, a model that uses a weighted ensemble of neural network classifiers for adaptive learning, trained by means of a quantum-inspired evolutionary algorithm (QIEA). We showed that the neuro-evolutionary classifiers were able to learn the dataset and to quickly respond to any drifts on the underlying data. Now, we are particularly interested on analyzing the influence of an unlimited ensemble, instead of the limited ensemble from NEVE. For that, we modified NEVE to work with unlimited ensembles, and we call this new algorithm NEVE++. To verity how the unlimited ensemble influences the results, we used four different datasets with concept drift in order to compare the accuracy of NEVE and NEVE++, using two other existing algorithms as reference.

This paper introduces two new hybrid models for clustering problems in which the input features and parameters of a spiking neural network (SNN) are optimized using evolutionary algorithms. We used two novel evolutionary approaches, the quantum-inspired evolutionary algorithm (QIEA) and the optimization by genetic programming (OGP) methods, to develop the quantum binary-real evolving SNN (QbrSNN) and the SNN optimized by genetic programming (SNN-OGP) neuro-evolutionary models, respectively. The proposed models are applied to 8 benchmark datasets, and a significantly higher clustering accuracy compared to a standard SNN without feature and parameter optimization is achieved with fewer iterations. When comparing QbrSNN and SNN-OGP, the former performed slightly better but at the expense of increased computational effort.

This work presents a Genetic Fuzzy Classification System (GFCS) called Genetic Programming Fuzzy Classification System (GPF-CLASS). This model differs from the traditional approach of GFCS, which uses the metaheuristic as a way to learn “if-then” fuzzy rules. This classical approach needs several changes and constraints on the use of genetic operators, evaluation and selection, which depends primarily on the metaheuristic used. Genetic Programming makes this implementation costly and explores few of its characteristics and potentialities. The GPF-CLASS model seeks for a greater integration with the metaheuristic: Multi-Gene Genetic Programming (MGGP), exploring its potential of terminals selection (input features) and functional form and at the same time aims to provide the user with a comprehension of the classification solution. Tests with 22 benchmarks datasets for classification have been performed and, as well as statistical analysis and comparisons with others Genetic Fuzzy Systems proposed in the literature.

This work describes the use of a weighted ensemble of neural network classifiers for adaptive learning. We train the neural networks by means of a quantum-inspired evolutionary algorithm (QIEA). The QIEA is also used to determine the best weights for each classifier belonging to the ensemble when a new block of data arrives. We show that the neuroevolutionary classifiers are able to learn the data set and to quickly respond to any drifts on the underlying data. We also compare the results reached by our model with an existing algorithm, Learn++.NSE, in two different nonstationary scenarios.

This work describes the use of a weighted ensemble of neural network classifiers for adaptive learning. We train the neural networks by means of a quantum-inspired evolutionary algorithm (QIEA). The QIEA is also used to determine the best weights for each classifier belonging to the ensemble when a new block of data arrives. After running several simulations using two different datasets and performing two different analysis of the results, we show that the proposed algorithm, named neuro-evolutionary ensemble (NEVE), was able to learn the data set and to quickly respond to any drifts on the underlying data, indicating that our model can be a good alternative to address concept drift problems. We also compare the results obtained by our model with an existing algorithm, Learn++.NSE, in two different nonstationary scenarios.