News

Our paper "Multi-Agent LLMs for Adaptive Acquisition in Bayesian Optimization" has been accepted at the IISE Annual Conference & Expo 2026, and has been selected as a finalist for the DAIS Best Student Paper Award!

• In this paper, we study how Large Language Models (LLMs) manage the exploration–exploitation trade-off in black-box optimization. Unlike classical Bayesian Optimization, where this trade-off is explicitly encoded through acquisition functions, LLM-based optimizers rely on implicit prompt-based reasoning, making their search behavior difficult to analyze or control. We show that single-agent LLMs, which jointly perform strategy selection and candidate generation in a single prompt, suffer from cognitive overload, leading to unstable dynamics and premature convergence.

• To address this, we propose a multi-agent framework that decomposes search policy learning into two roles: a strategy agent that assigns interpretable weights to exploration criteria (exploitation, informativeness, diversity, and representativeness), and a generation agent that produces candidates conditioned on the resulting policy. Empirical results across three benchmarks, Rosenbrock, hyperparameter tuning, and robot pushing, show that this decomposition stabilizes search behavior and achieves performance competitive with standard GP-based BO, while providing interpretable strategy trajectories that reveal how LLMs dynamically adapt across optimization landscapes. 

📃 Full Paper 

We are excited to announce that two of our papers have been accepted at AAAI 2026!

• The first paper, "Bi-Level Contextual Bandits for Individualized Resource Allocation under Delayed Feedback" (Mohammadsina Almasi, Hadis Anahideh), addresses the challenge of equitably allocating limited resources in high-stakes domains such as education and workforce development. We propose a bi-level contextual bandit framework that operates at two levels: a meta level that optimizes budget allocations across subgroups subject to fairness and capacity constraints, and a base level that identifies the most responsive individuals using a neural network trained on observational data. By explicitly modeling feedback delays through resource-specific delay kernels and cooldown windows, the algorithm continually refines its policy as new data arrive. Experiments on real-world datasets show higher cumulative outcomes and more equitable distribution compared to existing approaches. 

📃 Full Paper 

• The second paper, "Resolving Predictive Multiplicity for the Rashomon Set" (Parian Haghighat, Hadis Anahideh, Cynthia Rudin), tackles the problem of predictive multiplicity, where many equally accurate models disagree on individual predictions, undermining trust in high-stakes applications. We propose three complementary approaches: outlier correction, local patching, and pairwise reconciliation, each targeting a distinct source of inconsistency within the Rashomon set. The reconciled predictions can be distilled into a single interpretable model for deployment. Across multiple datasets, our methods reduce disagreement metrics while maintaining competitive accuracy. 

📃 Full Paper 

Our research group attended the INFORMS 2025 Annual Meeting in Atlanta, GA.

• We held four talk sessions to present our recent research projects:

  • • Almasi, Mohammadsina, and Hadis Anahideh. "Bi-Level Contextual Bandits for Individualized Resource Allocation under Delayed Feedback." 

    • Haghighat, Parian, Hadis Anahideh, and Cynthia Rudin. "Resolving Predictive Multiplicity for the Rashomon Set."

    • Carbonati, Andrea, Mohammadsina Almasi, and Hadis Anahideh. "Multi-Agent LLMs for Adaptive Acquisition in Bayesian Optimization."

    • Almasi, Mohammadsina, et al. "Adaptive Pareto Exploration (APEX) for Fairness-Aware Hyperparameter Optimization in FairPilot." Information and Software Technology 189.C (2026).

📃 Full Paper