While LLMs excel at general tasks, they struggle in specialized domains like finance, requiring diverse skills in domain knowledge, mathematical reasoning, and multilingual processing. Merging domain-specific Continual Pre-training (CPT) "experts" offers a practical alternative to costly and unstable multi-skill training. However, unlike established Supervised Fine-Tuning (SFT) model-based merging, CPT model merging remains largely unexplored. We address this gap by creating financial LLMs from experts in finance, math, and Japanese. We propose a three-stage evaluation focusing on knowledge recovery, complementarity, and emergence, and assess three merging methods (Task Arithmetic, TIES, and DARE-TIES) on a comprehensive financial benchmark curated from 18 tasks across 8 established datasets. Results show that merging an expert with its base model recovers general knowledge lost during CPT, while merging experts improves performance and can yield emergent cross-domain skills. Among the methods, Task Arithmetic performs strongly but is hyperparameter-sensitive, whereas TIES is more robust. Our findings also suggest that while model similarity correlates with merging success, emergent skills depend on more complex factors. This work presents the first foundational analysis of CPT model merging, establishing a principled framework and providing clear guidance for building multi-skill LLMs from existing assets.

Recent work has demonstrated state-of-the-art results in large language model (LLM) hallucination detection and mitigation through consistency-based approaches which involve aggregating multiple responses sampled from a single LLM for a given prompt. These approaches help offset limitations stemming from the imperfect data on which LLMs are trained, which includes biases and under-representation of information required at deployment time among other limitations which can lead to hallucinations. We show that extending these single-model consistency methods to combine responses from multiple LLMs with different training data, training schemes and model architectures can result in substantial further improvements in hallucination detection and mitigation capabilities beyond their single-model consistency counterparts. We evaluate this "consortium consistency" approach across many model teams from a pool of 15 LLMs and explore under what conditions it is beneficial to team together different LLMs in this manner. Further, we show that these performance improvements often come with reduced inference costs, offsetting a significant drawback with single-model consistency methods.

Model merging aims to combine multiple expert models into a more capable single model, offering benefits such as reduced storage and serving costs, improved generalization, and support for decentralized model development. Despite its promise, previous studies have primarily focused on merging a few small models. This leaves many unanswered questions about the effect of scaling model size and how it interplays with other key factors -- like the base model quality and number of expert models -- , to affect the merged model's performance. This work systematically evaluates the utility of model merging at scale, examining the impact of these different factors. We experiment with merging fully fine-tuned models using 4 popular merging methods -- Averaging, Task~Arithmetic, Dare, and TIES -- across model sizes ranging from 1B-64B parameters and merging up to 8 different expert models. We evaluate the merged models on both held-in tasks, i.e., the expert's training tasks, and zero-shot generalization to unseen held-out tasks. Our experiments provide several new insights about model merging at scale and the interplay between different factors. First, we find that merging is more effective when experts are created from strong base models, i.e., models with good zero-shot performance. Second, larger models facilitate easier merging. Third merging consistently improves generalization capabilities. Notably, when merging 8 large expert models, the merged models often generalize better compared to the multitask trained models. Fourth, we can better merge more expert models when working with larger models. Fifth, different merging methods behave very similarly at larger scales. Overall, our findings shed light on some interesting properties of model merging while also highlighting some limitations. We hope that this study will serve as a reference point on large-scale merging for upcoming research.

Merging has become a widespread way to cheaply combine individual models into a single model that inherits their capabilities and attains better performance. This popularity has spurred rapid development of many new merging methods, which are typically validated in disparate experimental settings and frequently differ in the assumptions made about model architecture, data availability, and computational budget. In this work, we characterize the relative merits of different merging methods by evaluating them in a shared experimental setting and precisely identifying the practical requirements of each method. Specifically, our setting focuses on using merging for compositional generalization of capabilities in image classification, image generation, and natural language processing. Additionally, we measure the computational costs of different merging methods as well as how they perform when scaling the number of models being merged. Taken together, our results clarify the state of the field of model merging and provide a comprehensive and rigorous experimental setup to test new methods.

As a multitude of capable machine learning (ML) models become widely available in forms such as open-source software and public APIs, central questions remain regarding their use in real-world applications, especially in high-stakes decision-making. Is there always one best model that should be used? When are the models likely to be error-prone? Should a black-box or interpretable model be used? In this work, we develop a prescriptive methodology to address these key questions, introducing a tree-based approach, Optimal Predictive-Policy Trees (OP2T), that yields interpretable policies for adaptively selecting a predictive model or ensemble, along with a parameterized option to reject making a prediction. We base our methods on learning globally optimized prescriptive trees. Our approach enables interpretable and adaptive model selection and rejection while only assuming access to model outputs. By learning policies over different feature spaces, including the model outputs, our approach works with both structured and unstructured datasets. We evaluate our approach on real-world datasets, including regression and classification tasks with both structured and unstructured data. We demonstrate that our approach provides both strong performance against baseline methods while yielding insights that help answer critical questions about which models to use, and when.

There is a long history in machine learning of model ensembling, beginning with boosting and bagging and continuing to the present day. Much of this history has focused on combining models for classification and regression, but recently there is interest in more complex settings such as ensembling policies in reinforcement learning. Strong connections have also emerged between ensembling and multicalibration techniques. In this work, we further investigate these themes by considering a setting in which we wish to ensemble models for multidimensional output predictions that are in turn used for downstream optimization. More precisely, we imagine we are given a number of models mapping a state space to multidimensional real-valued predictions. These predictions form the coefficients of a linear objective that we would like to optimize under specified constraints. The fundamental question we address is how to improve and combine such models in a way that outperforms the best of them in the downstream optimization problem. We apply multicalibration techniques that lead to two provably efficient and convergent algorithms. The first of these (the white box approach) requires being given models that map states to output predictions, while the second (the \emph{black box} approach) requires only policies (mappings from states to solutions to the optimization problem). For both, we provide convergence and utility guarantees. We conclude by investigating the performance and behavior of the two algorithms in a controlled experimental setting.

Over the past years, Machine Learning-as-a-Service (MLaaS) has received a surging demand for supporting Machine Learning-driven services to offer revolutionized user experience across diverse application areas. MLaaS provides inference service with low inference latency based on an ML model trained using a dataset collected from numerous individual data owners. Recently, for the sake of data owners' privacy and to comply with the "right to be forgotten (RTBF)" as enacted by data protection legislation, many machine unlearning methods have been proposed to remove data owners' data from trained models upon their unlearning requests. However, despite their promising efficiency, almost all existing machine unlearning methods handle unlearning requests independently from inference requests, which unfortunately introduces a new security issue of inference service obsolescence and a privacy vulnerability of undesirable exposure for machine unlearning in MLaaS. In this paper, we propose the ERASER framework for machinE unleaRning in MLaAS via an inferencE seRving-aware approach. ERASER strategically choose appropriate unlearning execution timing to address the inference service obsolescence issue. A novel inference consistency certification mechanism is proposed to avoid the violation of RTBF principle caused by postponed unlearning executions, thereby mitigating the undesirable exposure vulnerability. ERASER offers three groups of design choices to allow for tailor-made variants that best suit the specific environments and preferences of various MLaaS systems. Extensive empirical evaluations across various settings confirm ERASER's effectiveness, e.g., it can effectively save up to 99% of inference latency and 31% of computation overhead over the inference-oblivion baseline.

Training large language models (LLMs) is a costly endeavour in terms of time and computational resources. The large amount of training data used during the unsupervised pre-training phase makes it difficult to verify all data and, unfortunately, undesirable data may be ingested during training. Re-training from scratch is impractical and has led to the creation of the 'unlearning' discipline where models are modified to "unlearn" undesirable information without retraining. However, any modification can alter the behaviour of LLMs, especially on key dimensions such as fairness. This is the first work that examines this interplay between unlearning and fairness for LLMs. In particular, we focus on a popular unlearning framework known as SISA [Bourtoule et al., 2021], which creates an ensemble of models trained on disjoint shards. We evaluate the performance-fairness trade-off for SISA, and empirically demsontrate that SISA can indeed reduce fairness in LLMs. To remedy this, we propose post-processing bias mitigation techniques for ensemble models produced by SISA. We adapt the post-processing fairness improvement technique from [Hardt et al., 2016] to design three methods that can handle model ensembles, and prove that one of the methods is an optimal fair predictor for ensemble of models. Through experimental results, we demonstrate the efficacy of our post-processing framework called 'FairSISA'.

The South Australia region of the Australian National Electricity Market (NEM) displays some of the highest levels of price volatility observed in modern electricity markets. This paper outlines an approach to probabilistic forecasting under these extreme conditions, including spike filtration and several post-processing steps. We propose using quantile regression as an ensemble tool for probabilistic forecasting, with our combined forecasts achieving superior results compared to all constituent models. Within our ensemble framework, we demonstrate that averaging models with varying training length periods leads to a more adaptive model and increased prediction accuracy. The applicability of the final model is evaluated by comparing our median forecasts with the point forecasts available from the Australian NEM operator, with our model outperforming these NEM forecasts by a significant margin.

Ensembling certifiably robust neural networks is a promising approach for improving the \emph{certified robust accuracy} of neural models. Black-box ensembles that assume only query-access to the constituent models (and their robustness certifiers) during prediction are particularly attractive due to their modular structure. Cascading ensembles are a popular instance of black-box ensembles that appear to improve certified robust accuracies in practice. However, we show that the robustness certifier used by a cascading ensemble is unsound. That is, when a cascading ensemble is certified as locally robust at an input $x$ (with respect to $\epsilon$), there can be inputs $x'$ in the $\epsilon$-ball centered at $x$, such that the cascade's prediction at $x'$ is different from $x$ and thus the ensemble is not locally robust. Our theoretical findings are accompanied by empirical results that further demonstrate this unsoundness. We present \emph{cascade attack} (CasA), an adversarial attack against cascading ensembles, and show that: (1) there exists an adversarial input for up to 88\% of the samples where the ensemble claims to be certifiably robust and accurate; and (2) the accuracy of a cascading ensemble under our attack is as low as 11\% when it claims to be certifiably robust and accurate on 97\% of the test set. Our work reveals a critical pitfall of cascading certifiably robust models by showing that the seemingly beneficial strategy of cascading can actually hurt the robustness of the resulting ensemble. Our code is available at \url{https://github.com/TristaChi/ensembleKW}.