Raad Khraishi

Yifan Bao, Xinyu Xi, Xinyu Liu et al.

Automated data science is a structured model-selection problem. A solution must choose data transformations, feature representations, architecture, training procedure, evaluation protocol, and refinement strategy for a task. AutoML systems automate parts of this process, but typically search within predefined pipeline, model, and hyperparameter spaces. LLM-based agents offer greater flexibility through retrieval, code generation, and execution feedback, yet their modelling decisions are often unstructured, difficult to verify, and hard to reuse. We introduce \textsc{MOSAIC} (Modular Orchestration for Structured Agentic Intelligence and Composition), a structured agentic framework for memory-grounded model selection and workflow construction. Given a task and dataset, \textsc{MOSAIC} builds a semantic task profile, retrieves prior cases and source-code modules, and constructs a blueprint: an intermediate representation specifying selected modelling components, composition, interface constraints, and execution requirements. This blueprint turns model selection into a staged, context-grounded search and grounds LLM-based code generation in retrieved evidence rather than unconstrained synthesis. Candidate models are validated by execution and refined using diagnostic feedback, training traces, task metrics, and a failure-aware reinforcement learning policy. We instantiate \textsc{MOSAIC} on financial time-series forecasting and generation, where models must satisfy predictive accuracy, distributional fidelity, execution reliability, and downstream financial criteria such as risk and tail behaviour. Experiments against AutoML and agentic baselines show that \textsc{MOSAIC} improves task performance, execution success, and decision traceability, demonstrating the value of treating automated data science as structured, reusable, and execution-grounded model selection.

Salvatore Mercuri, Raad Khraishi, Ramin Okhrati et al.

Removing the influence of a specified subset of training data from a machine learning model may be required to address issues such as privacy, fairness, and data quality. Retraining the model from scratch on the remaining data after removal of the subset is an effective but often infeasible option, due to its computational expense. The past few years have therefore seen several novel approaches towards efficient removal, forming the field of "machine unlearning", however, many aspects of the literature published thus far are disparate and lack consensus. In this paper, we summarise and compare seven state-of-the-art machine unlearning algorithms, consolidate definitions of core concepts used in the field, reconcile different approaches for evaluating algorithms, and discuss issues related to applying machine unlearning in practice.

Raad Khraishi, Ramin Okhrati

We introduce a method for pricing consumer credit using recent advances in offline deep reinforcement learning. This approach relies on a static dataset and requires no assumptions on the functional form of demand. Using both real and synthetic data on consumer credit applications, we demonstrate that our approach using the conservative Q-Learning algorithm is capable of learning an effective personalized pricing policy without any online interaction or price experimentation.

Greig Cowan, Salvatore Mercuri, Raad Khraishi

Understanding customer lifetime value is key to nurturing long-term customer relationships, however, estimating it is far from straightforward. In the retail banking industry, commonly used approaches rely on simple heuristics and do not take advantage of the high predictive ability of modern machine learning techniques. We present a general framework for modelling customer lifetime value which may be applied to industries with long-lasting contractual and product-centric customer relationships, of which retail banking is an example. This framework is novel in facilitating CLV predictions over arbitrary time horizons and product-based propensity models. We also detail an implementation of this model which is currently in production at a large UK lender. In testing, we estimate an 43% improvement in out-of-time CLV prediction error relative to a popular baseline approach. Propensity models derived from our CLV model have been used to support customer contact marketing campaigns. In testing, we saw that the top 10% of customers ranked by their propensity to take up investment products were 3.2 times more likely to take up an investment product in the next year than a customer chosen at random.

Devesh Batra, Salvatore Mercuri, Raad Khraishi

We introduce Longitudinal Predictive Conformal Inference (LPCI), a novel distribution-free conformal prediction algorithm for longitudinal data. Current conformal prediction approaches for time series data predominantly focus on the univariate setting, and thus lack cross-sectional coverage when applied individually to each time series in a longitudinal dataset. The current state-of-the-art for longitudinal data relies on creating infinitely-wide prediction intervals to guarantee both cross-sectional and asymptotic longitudinal coverage. The proposed LPCI method addresses this by ensuring that both longitudinal and cross-sectional coverages are guaranteed without resorting to infinitely wide intervals. In our approach, we model the residual data as a quantile fixed-effects regression problem, constructing prediction intervals with a trained quantile regressor. Our extensive experiments demonstrate that LPCI achieves valid cross-sectional coverage and outperforms existing benchmarks in terms of longitudinal coverage rates. Theoretically, we establish LPCI's asymptotic coverage guarantees for both dimensions, with finite-width intervals. The robust performance of LPCI in generating reliable prediction intervals for longitudinal data underscores its potential for broad applications, including in medicine, finance, and supply chain management.

Raad Khraishi, Iman Zafar, Katie Myles et al.

Deployed multi-turn LLM systems routinely switch models mid-interaction due to upgrades, cross-provider routing, and fallbacks. Such handoffs create a context mismatch: the model generating later turns must condition on a dialogue prefix authored by a different model, potentially inducing silent performance drift. We introduce a switch-matrix benchmark that measures this effect by running a prefix model for early turns and a suffix model for the final turn, and comparing against the no-switch baseline using paired episode-level bootstrap confidence intervals. Across CoQA conversational QA and Multi-IF benchmarks, even a single-turn handoff yields prevalent and statistically significant, directional effects and may swing outcomes by -8 to +13 percentage points in Multi-IF strict success rate and +/- 4 absolute F1 on CoQA, comparable to the no-switch gap between common model tiers (e.g., GPT-5-nano vs GPT-5-mini). We further find systematic compatibility patterns: some suffix models degrade under nearly any non-self dialogue history, while others improve under nearly any foreign prefix. To enable compressed handoff risk monitoring, we decompose switch-induced drift into per-model prefix influence and suffix susceptibility terms, accounting for ~70% of variance across benchmarks. These results position handoff robustness as an operational reliability dimension that single-model benchmarks miss, motivating explicit monitoring and handoff-aware mitigation in multi-turn systems.

John Hartley, Conor Hamill, Devesh Batra et al.

Large Language Models (LLMs) are increasingly deployed as autonomous agents, necessitating a deeper understanding of their decision-making behaviour under risk. This study investigates the relationship between LLMs' personality traits and risk propensity, employing cumulative prospect theory (CPT) and the Big Five personality framework. We focus on GPT-4o, comparing its behaviour to human baselines and earlier models. Our findings reveal that GPT-4o exhibits higher Conscientiousness and Agreeableness traits compared to human averages, while functioning as a risk-neutral rational agent in prospect selection. Interventions on GPT-4o's Big Five traits, particularly Openness, significantly influence its risk propensity, mirroring patterns observed in human studies. Notably, Openness emerges as the most influential factor in GPT-4o's risk propensity, aligning with human findings. In contrast, legacy models like GPT-4-Turbo demonstrate inconsistent generalization of the personality-risk relationship. This research advances our understanding of LLM behaviour under risk and elucidates the potential and limitations of personality-based interventions in shaping LLM decision-making. Our findings have implications for the development of more robust and predictable AI systems such as financial modelling.

Robert Hankache, Kingsley Nketia Acheampong, Liang Song et al.

As large language models (LLMs) are increasingly deployed in multi-turn dialogue and other sustained interactive scenarios, it is essential to understand how extended context affects their performance. Popular benchmarks, focusing primarily on single-turn question answering (QA) tasks, fail to capture the effects of multi-turn exchanges. To address this gap, we introduce a novel set of benchmarks that systematically vary the volume and nature of prior context. We evaluate multiple conventional LLMs, including GPT, Claude, and Gemini, across these benchmarks to measure their sensitivity to contextual variations. Our findings reveal that LLM performance on multiple-choice questions can degrade dramatically in multi-turn interactions, with performance drops as large as 73% for certain models. Even highly capable models such as GPT-4o exhibit up to a 32% decrease in accuracy. Notably, the relative performance of larger versus smaller models is not always predictable. Moreover, the strategic placement of the task description within the context can substantially mitigate performance drops, improving the accuracy by as much as a factor of 3.5. These findings underscore the need for robust strategies to design, evaluate, and mitigate context-related sensitivity in LLMs.

Alankar Atreya, Stefan Sylvius Wanger, Devesh Batra et al.

Banks receive millions of reports of fraud, scams, and disputed transactions every year, making it challenging to accurately direct customers to the appropriate specialist teams for assistance. The existing manual process driven by humans is slow and stressful for both customers and staff. To address this, we develop a customer-facing AI powered triaging agent that leverages large language models (LLMs) to conduct multi-turn conversations, ask relevant questions, and classify cases for accurate, policy-guided routing, making it embedded in the customer journey. To evaluate and continuously improve the agent, synthetic digital twins of real customers were simulated, generating realistic, labelled dialogues based on historical data to test a wide range of real-world scenarios. This work details the triage agent's modelling approach, integration with policy, safety guardrails and reasoning frameworks, the use of the synthetic agent for scalable evaluation, and findings on the AI system's accuracy, robustness, and compliance. Results show that the agent successfully improves triaging of historical cases, achieving a 30.6% increase in classification accuracy, with high satisfaction levels reported by our subject-matter experts, highlighting how targeted probing can lead to more effective triage in banking operations at scale.

Giulio Pelosio, Devesh Batra, Noémie Bovey et al.

Large Language Models (LLMs) can exhibit latent biases towards specific nationalities even when explicit demographic markers are not present. In this work, we introduce a novel name-based benchmarking approach derived from the Bias Benchmark for QA (BBQ) dataset to investigate the impact of substituting explicit nationality labels with culturally indicative names, a scenario more reflective of real-world LLM applications. Our novel approach examines how this substitution affects both bias magnitude and accuracy across a spectrum of LLMs from industry leaders such as OpenAI, Google, and Anthropic. Our experiments show that small models are less accurate and exhibit more bias compared to their larger counterparts. For instance, on our name-based dataset and in the ambiguous context (where the correct choice is not revealed), Claude Haiku exhibited the worst stereotypical bias scores of 9%, compared to only 3.5% for its larger counterpart, Claude Sonnet, where the latter also outperformed it by 117.7% in accuracy. Additionally, we find that small models retain a larger portion of existing errors in these ambiguous contexts. For example, after substituting names for explicit nationality references, GPT-4o retains 68% of the error rate versus 76% for GPT-4o-mini, with similar findings for other model providers, in the ambiguous context. Our research highlights the stubborn resilience of biases in LLMs, underscoring their profound implications for the development and deployment of AI systems in diverse, global contexts.

Raad Khraishi, Ramin Okhrati

Data augmentation is a widely used technique for improving model performance in machine learning, particularly in computer vision and natural language processing. Recently, there has been increasing interest in applying augmentation techniques to reinforcement learning (RL) problems, with a focus on image-based augmentation. In this paper, we explore a set of generic wrappers designed to augment RL environments with noise and encourage agent exploration and improve training data diversity which are applicable to a broad spectrum of RL algorithms and environments. Specifically, we concentrate on augmentations concerning states, rewards, and transition dynamics and introduce two novel augmentation techniques. In addition, we introduce a noise rate hyperparameter for control over the frequency of noise injection. We present experimental results on the impact of these wrappers on return using three popular RL algorithms, Soft Actor-Critic (SAC), Twin Delayed DDPG (TD3), and Proximal Policy Optimization (PPO), across five MuJoCo environments. To support the choice of augmentation technique in practice, we also present analysis that explores the performance these techniques across environments. Lastly, we publish the wrappers in our noisyenv repository for use with gym environments.