Agents are rapidly advancing in automating digital work, but enterprises face a harder challenge: moving beyond prototypes to deployed systems that deliver measurable business value. This path is complicated by fragmented frameworks, slow development, and the absence of standardized evaluation practices. Generalist agents have emerged as a promising direction, excelling on academic benchmarks and offering flexibility across tasks, applications, and modalities. Yet, evidence of their use in enterprise settings remains limited. This paper reports IBM’s experience developing and piloting the Computer Using Generalist Agent (CUGA). CUGA adopts a hierarchical planner--executor architecture with strong analytical foundations, achieving state-of-the-art performance on AppWorld and WebArena. Beyond benchmarks, it was evaluated in a Business-Process-Outsourcing talent acquisition pilot, addressing enterprise requirements for scalability, auditability, safety, and governance. In preliminary evaluations, CUGA approached the accuracy of specialized agents while suggesting reductions in development time and cost. We provide early evidence that generalist agents can operate at enterprise scale, distill key technical and organizational lessons, and outline requirements for transitioning research-grade architectures like CUGA into enterprise-ready systems.

This paper proposes a new non-native solver for Answer Set Programming (ASP) named maxmodels. It is based on translating logic normal programs into weighted MaxSAT clauses. We checked our proposal’s computational efficiency in the series of experiments on selected NP-hard optimization problems against the state-of-the-art ASP solvers dlv, clingo, and smodels. Our findings indicate that maxmodels outperforms other solvers in certain problems while also being competitive in others.

This paper introduces Bandit Recombination (BR), a novel genetic algorithm (GA) operator that leverages the multi-armed bandit (MAB) paradigm to enhance the exploitation of high-performance solutions in combinatorial optimization. BR dynamically balances exploration and exploitation via the ϵ-greedy strategy, using reward and pull matrices to guide recombination from an elite subpopulation. In a preliminary study on the NP-hard generalized assignment problem (GAP), BR-augmented GA achieves mean fitness improvements over baseline GA across diverse datasets.

This paper investigates the use of logic programming—specifically, Answer Set Programming (ASP)—as a declarative approach to orthogonal graph drawing on restricted grids. The drawing problem is encoded as a constraint satisfaction problem that enforces planarity, orthogonality, and spatial constraints within bounded integer grids. Experimental results on benchmark graphs indicate that the approach generates valid and aesthetically pleasing layouts, particularly for small and medium-sized instances. These findings demonstrate the promise of declarative programming in supporting constraint-driven graph layout methodologies.

In this paper, we explore the problem of finding all substrings, including those that overlap with others, that match given regular expressions in a long text. The standard methods (like finditer in Python or NextMatch in C#) are inefficient for the task due to the challenging requirement of finding all possible overlaps. We propose two new methods to tackle the problem: an efficient one based on automata and a suffix tree and an enhanced naive solution based solely on automata. Our approaches were evaluated in a series of experiments against each other.

This paper proposes a novel algorithm built on a reinforcement learning approach as an alternative to standard hyperparameter tuning methods for improving the feature selection process to obtain better classification performance. Various techniques have been proposed to handle the problem, and among them are metaheuristics, such as Particle Swarm Optimization (PSO). However, a new problem arises as learning algorithms must often be tuned in terms of their hyperparameters. We need to solve this problem separately, and vary of methods can be used for this matter. The hyperparameter optimization for the learning process of PSO can be seen as a sequence decision-making process. Hence, we can solve it with the multi-armed bandit problem algorithms.

The string site-removal problem (SSRP) originates from the field of computational biology and is known to be NP-hard. The problem is defined on the set of strings of equal length and an integer k, and its goal is to find the largest subset of the strings that become identical after removing at most k sites (columns). This paper proposes two new approaches to SSRP. The first is an exact algorithm that uses the quadratically constrained program (QCP). The second approach combines QCP and genetic algorithm (GA). In a series of experiments, we showed that our exact program performs faster on average than the state-of-the-art method (Integer Linear Programming, ILP). Regarding heuristic approaches, our new proposal gives the identical results as the state-of-the-art reference method (Large Neighborhood Search, LNS) with the same CPU time budget. All source codes and benchmark instances we made publicly available via GitHub.

Accounts Payable (AP) is a resource-intensive business process in large enterprises for paying vendors within contractual payment deadlines for goods and services procured from them. There are multiple verifications before payment to the supplier/vendor. After the validations, the invoice flows through several steps such as vendor identification, line-item matching for Purchase order (PO) based invoices, Accounting Code identification for Non- Purchase order (Non-PO) based invoices, tax code identification, etc. Currently, each of these steps is mostly manual and cumbersome making it labor-intensive, error-prone, and requiring constant training of agents. Automatically processing these invoices for payment without any manual intervention is quite difficult. To tackle this challenge, we have developed an automated end-to-end invoice processing system using AI-based modules for multiple steps of the invoice processing pipeline. It can be configured to an individual client’s requirements with minimal effort. Currently, the system is deployed in production for two clients. It has successfully processed around ~80k invoices out of which 76% invoices were processed with low or no manual intervention.

This paper presents a detailed study of the discrete particle swarm optimization algorithm (DPSO) applied to solve the dynamic traveling salesman problem which has many practical applications in planning, logistics and chip manufacturing. The dynamic version is especially important in practical applications in which new circumstances, e.g., a traffic jam or a machine failure, could force changes to the problem specification. The DPSO algorithm was enriched with a pheromone memory which is used to guide the search process similarly to the ant colony optimization algorithm. The paper extends our previous work on the DPSO algorithm in various ways. Firstly, the performance of the algorithm is thoroughly tested on a set of newly generated DTSP instances which differ in the number and the size of the changes. Secondly, the impact of the pheromone memory on the convergence of the DPSO is investigated and compared with the version without a pheromone memory. Moreover, the results are compared with two ant colony optimization algorithms, namely the $$\mathcal {MAX}$$ – $$\mathcal {MIN}$$ ant system (MMAS) and the population-based ant colony optimization (PACO). The results show that the DPSO is able to find high-quality solutions to the DTSP and its performance is competitive with the performance of the MMAS and the PACO algorithms. Moreover, the pheromone memory has a positive impact on the convergence of the algorithm, especially in the face of dynamic changes to the problem’s definition.