Elizabeth Daly

Svetoslav Nizhnichenkov, Rahul Nair, Elizabeth Daly et al.

We investigate how successful bias mitigation reshapes the embedding space of encoder-only and decoder-only foundation models, offering an internal audit of model behaviour through representational analysis. Using BERT and Llama2 as representative architectures, we assess the shifts in associations between gender and occupation terms by comparing baseline and bias-mitigated variants of the models. Our findings show that bias mitigation reduces gender-occupation disparities in the embedding space, leading to more neutral and balanced internal representations. These representational shifts are consistent across both model types, suggesting that fairness improvements can manifest as interpretable and geometric transformations. These results position embedding analysis as a valuable tool for understanding and validating the effectiveness of debiasing methods in foundation models. To further promote the assessment of decoder-only models, we introduce WinoDec, a dataset consisting of 4,000 sequences with gender and occupation terms, and release it to the general public. (https://github.com/winodec/wino-dec)

Stefano Teso, Öznur Alkan, Wolfang Stammer et al.

Explanations have gained an increasing level of interest in the AI and Machine Learning (ML) communities in order to improve model transparency and allow users to form a mental model of a trained ML model. However, explanations can go beyond this one way communication as a mechanism to elicit user control, because once users understand, they can then provide feedback. The goal of this paper is to present an overview of research where explanations are combined with interactive capabilities as a mean to learn new models from scratch and to edit and debug existing ones. To this end, we draw a conceptual map of the state-of-the-art, grouping relevant approaches based on their intended purpose and on how they structure the interaction, highlighting similarities and differences between them. We also discuss open research issues and outline possible directions forward, with the hope of spurring further research on this blooming research topic.

Jasmina Gajcin, James McCarthy, Rahul Nair et al.

A well-defined reward function is crucial for successful training of an reinforcement learning (RL) agent. However, defining a suitable reward function is a notoriously challenging task, especially in complex, multi-objective environments. Developers often have to resort to starting with an initial, potentially misspecified reward function, and iteratively adjusting its parameters, based on observed learned behavior. In this work, we aim to automate this process by proposing ITERS, an iterative reward shaping approach using human feedback for mitigating the effects of a misspecified reward function. Our approach allows the user to provide trajectory-level feedback on agent's behavior during training, which can be integrated as a reward shaping signal in the following training iteration. We also allow the user to provide explanations of their feedback, which are used to augment the feedback and reduce user effort and feedback frequency. We evaluate ITERS in three environments and show that it can successfully correct misspecified reward functions.

James McCarthy, Rahul Nair, Elizabeth Daly et al.

Explainability of Reinforcement Learning (RL) policies remains a challenging research problem, particularly when considering RL in a safety context. Understanding the decisions and intentions of an RL policy offer avenues to incorporate safety into the policy by limiting undesirable actions. We propose the use of a Boolean Decision Rules model to create a post-hoc rule-based summary of an agent's policy. We evaluate our proposed approach using a DQN agent trained on an implementation of a lava gridworld and show that, given a hand-crafted feature representation of this gridworld, simple generalised rules can be created, giving a post-hoc explainable summary of the agent's policy. We discuss possible avenues to introduce safety into a RL agent's policy by using rules generated by this rule-based model as constraints imposed on the agent's policy, as well as discuss how creating simple rule summaries of an agent's policy may help in the debugging process of RL agents.

Javier Carnerero-Cano, Massimiliano Pronesti, Radu Marinescu et al.

Large language models (LLMs) are widely used in knowledge-intensive applications but often generate factually incorrect responses. A promising approach to rectify these flaws is correcting LLMs using feedback. Therefore, in this paper, we introduce FactCorrector, a new post-hoc correction method that adapts across domains without retraining and leverages structured feedback about the factuality of the original response to generate a correction. To support rigorous evaluations of factuality correction methods, we also develop the VELI5 benchmark, a novel dataset containing systematically injected factual errors and ground-truth corrections. Experiments on VELI5 and several popular long-form factuality datasets show that the FactCorrector approach significantly improves factual precision while preserving relevance, outperforming strong baselines. We release our code at https://ibm.biz/factcorrector.

Shivani Tomar, Seshu Tirupathi, Elizabeth Daly et al.

Time series foundation models have recently gained a lot of attention due to their ability to model complex time series data encompassing different domains including traffic, energy, and weather. Although they exhibit strong average zero-shot performance on forecasting tasks, their predictions on certain critical regions of the data are not always reliable, limiting their usability in real-world applications, especially when data exhibits unique trends. In this paper, we propose a selective forecasting framework to identify these critical segments of time series using shapelets. We learn shapelets using shift-invariant dictionary learning on the validation split of the target domain dataset. Utilizing distance-based similarity to these shapelets, we facilitate the user to selectively discard unreliable predictions and be informed of the model's realistic capabilities. Empirical results on diverse benchmark time series datasets demonstrate that our approach leveraging both zero-shot and full-shot fine-tuned models reduces the overall error by an average of 22.17% for zero-shot and 22.62% for full-shot fine-tuned model. Furthermore, our approach using zero-shot and full-shot fine-tuned models, also outperforms its random selection counterparts by up to 21.41% and 21.43% on one of the datasets.

Radu Marinescu, Debarun Bhattacharjya, Junkyu Lee et al.

Large language models (LLMs) have achieved remarkable success in generative tasks, yet they often fall short in ensuring the factual accuracy of their outputs, thus limiting their reliability in real-world applications where correctness is critical. In this paper, we present FactReasoner, a novel neuro-symbolic based factuality assessment framework that employs probabilistic reasoning to evaluate the truthfulness of long-form generated responses. FactReasoner decomposes a response into atomic units, retrieves relevant contextual information from external knowledge sources, and models the logical relationships (e.g., entailment, contradiction) between these units and their contexts using probabilistic encodings. It then estimates the posterior probability that each atomic unit is supported by the retrieved evidence. Our experiments on both labeled and unlabeled benchmark datasets demonstrate that FactReasoner often outperforms state-of-the-art prompt-based methods in terms of factual precision and recall. Our open-source implementation is publicly available at: https://github.com/IBM/FactReasoner.

Anna Sokol, Elizabeth Daly, Michael Hind et al.

Large language models (LLMs) are powerful tools capable of handling diverse tasks. Comparing and selecting appropriate LLMs for specific tasks requires systematic evaluation methods, as models exhibit varying capabilities across different domains. However, finding suitable benchmarks is difficult given the many available options. This complexity not only increases the risk of benchmark misuse and misinterpretation but also demands substantial effort from LLM users, seeking the most suitable benchmarks for their specific needs. To address these issues, we introduce \texttt{BenchmarkCards}, an intuitive and validated documentation framework that standardizes critical benchmark attributes such as objectives, methodologies, data sources, and limitations. Through user studies involving benchmark creators and users, we show that \texttt{BenchmarkCards} can simplify benchmark selection and enhance transparency, facilitating informed decision-making in evaluating LLMs. Data & Code: https://github.com/SokolAnn/BenchmarkCards

Srishti Yadav, Jasmina Gajcin, Erik Miehling et al.

Understanding how different stakeholders perceive risks in AI systems is essential for their responsible deployment. This paper presents a framework for stakeholder-grounded risk assessment by using LLMs, acting as judges to predict and explain risks. Using the Risk Atlas Nexus and GloVE explanation method, our framework generates stakeholder-specific, interpretable policies that shows how different stakeholders agree or disagree about the same risks. We demonstrate our method using three real-world AI use cases of medical AI, autonomous vehicles, and fraud detection domain. We further propose an interactive visualization that reveals how and why conflicts emerge across stakeholder perspectives, enhancing transparency in conflict reasoning. Our results show that stakeholder perspectives significantly influence risk perception and conflict patterns. Our work emphasizes the importance of these stakeholder-aware explanations needed to make LLM-based evaluations more transparent, interpretable, and aligned with human-centered AI governance goals.

Seshu Tirupathi, Dhaval Salwala, Elizabeth Daly et al.

As Large Language Models (LLMs) continue to be increasingly applied across various domains, their widespread adoption necessitates rigorous monitoring to prevent unintended negative consequences and ensure robustness. Furthermore, LLMs must be designed to align with human values, like preventing harmful content and ensuring responsible usage. The current automated systems and solutions for monitoring LLMs in production are primarily centered on LLM-specific concerns like hallucination etc, with little consideration given to the requirements of specific use-cases and user preferences. This paper introduces GAF-Guard, a novel agentic framework for LLM governance that places the user, the use-case, and the model itself at the center. The framework is designed to detect and monitor risks associated with the deployment of LLM based applications. The approach models autonomous agents that identify risks, activate risk detection tools, within specific use-cases and facilitate continuous monitoring and reporting to enhance AI safety, and user expectations. The code is available at https://github.com/IBM/risk-atlas-nexus-demos/tree/main/gaf-guard.

Yufang Hou, Alessandra Pascale, Javier Carnerero-Cano et al.

Retrieval-augmented generation (RAG) has emerged as a promising solution to mitigate the limitations of large language models (LLMs), such as hallucinations and outdated information. However, it remains unclear how LLMs handle knowledge conflicts arising from different augmented retrieved passages, especially when these passages originate from the same source and have equal trustworthiness. In this work, we conduct a comprehensive evaluation of LLM-generated answers to questions that have varying answers based on contradictory passages from Wikipedia, a dataset widely regarded as a high-quality pre-training resource for most LLMs. Specifically, we introduce WikiContradict, a benchmark consisting of 253 high-quality, human-annotated instances designed to assess LLM performance when augmented with retrieved passages containing real-world knowledge conflicts. We benchmark a diverse range of both closed and open-source LLMs under different QA scenarios, including RAG with a single passage, and RAG with 2 contradictory passages. Through rigorous human evaluations on a subset of WikiContradict instances involving 5 LLMs and over 3,500 judgements, we shed light on the behaviour and limitations of these models. For instance, when provided with two passages containing contradictory facts, all models struggle to generate answers that accurately reflect the conflicting nature of the context, especially for implicit conflicts requiring reasoning. Since human evaluation is costly, we also introduce an automated model that estimates LLM performance using a strong open-source language model, achieving an F-score of 0.8. Using this automated metric, we evaluate more than 1,500 answers from seven LLMs across all WikiContradict instances. To facilitate future work, we release WikiContradict on: https://ibm.biz/wikicontradict.

Amit Dhurandhar, Rahul Nair, Moninder Singh et al.

Evaluation and ranking of large language models (LLMs) has become an important problem with the proliferation of these models and their impact. Evaluation methods either require human responses which are expensive to acquire or use pairs of LLMs to evaluate each other which can be unreliable. In this paper, we provide a novel perspective where, given a dataset of prompts (viz. questions, instructions, etc.) and a set of LLMs, we rank them without access to any ground truth or reference responses. Inspired by real life where both an expert and a knowledgeable person can identify a novice our main idea is to consider triplets of models, where each one of them evaluates the other two, correctly identifying the worst model in the triplet with high probability. We also analyze our idea and provide sufficient conditions for it to succeed. Applying this idea repeatedly, we propose two methods to rank LLMs. In experiments on different generative tasks (summarization, multiple-choice, and dialog), our methods reliably recover close to true rankings without reference data. This points to a viable low-resource mechanism for practical use.

Jasmina Gajcin, Erik Miehling, Rahul Nair et al.

Using LLMs to evaluate text, that is, LLM-as-a-judge, is increasingly being used at scale to augment or even replace human annotations. As such, it is imperative that we understand the potential biases and risks of doing so. In this work, we propose an approach for extracting high-level concept-based global policies from LLM-as-a-Judge. Our approach consists of two algorithms: 1) CLoVE (Contrastive Local Verifiable Explanations), which generates verifiable, concept-based, contrastive local explanations and 2) GloVE (Global Verifiable Explanations), which uses iterative clustering, summarization and verification to condense local rules into a global policy. We evaluate GloVE on seven standard benchmarking datasets for content harm detection. We find that the extracted global policies are highly faithful to decisions of the LLM-as-a-Judge. Additionally, we evaluated the robustness of global policies to text perturbations and adversarial attacks. Finally, we conducted a user study to evaluate user understanding and satisfaction with global policies.

James McCarthy, Radu Marinescu, Elizabeth Daly et al.

Risk-averse Constrained Reinforcement Learning (RaCRL) aims to learn policies that minimise the likelihood of rare and catastrophic constraint violations caused by an environment's inherent randomness. In general, risk-aversion leads to conservative exploration of the environment which typically results in converging to sub-optimal policies that fail to adequately maximise reward or, in some cases, fail to achieve the goal. In this paper, we propose an exploration-based approach for RaCRL called Optimistic Risk-averse Actor Critic (ORAC), which constructs an exploratory policy by maximising a local upper confidence bound of the state-action reward value function whilst minimising a local lower confidence bound of the risk-averse state-action cost value function. Specifically, at each step, the weighting assigned to the cost value is increased or decreased if it exceeds or falls below the safety constraint value. This way the policy is encouraged to explore uncertain regions of the environment to discover high reward states whilst still satisfying the safety constraints. Our experimental results demonstrate that the ORAC approach prevents convergence to sub-optimal policies and improves significantly the reward-cost trade-off in various continuous control tasks such as Safety-Gymnasium and a complex building energy management environment CityLearn.

Celia Cintas, Miriam Rateike, Erik Miehling et al.

We present a study on how and where personas -- defined by distinct sets of human characteristics, values, and beliefs -- are encoded in the representation space of large language models (LLMs). Using a range of dimension reduction and pattern recognition methods, we first identify the model layers that show the greatest divergence in encoding these representations. We then analyze the activations within a selected layer to examine how specific personas are encoded relative to others, including their shared and distinct embedding spaces. We find that, across multiple pre-trained decoder-only LLMs, the analyzed personas show large differences in representation space only within the final third of the decoder layers. We observe overlapping activations for specific ethical perspectives -- such as moral nihilism and utilitarianism -- suggesting a degree of polysemy. In contrast, political ideologies like conservatism and liberalism appear to be represented in more distinct regions. These findings help to improve our understanding of how LLMs internally represent information and can inform future efforts in refining the modulation of specific human traits in LLM outputs. Warning: This paper includes potentially offensive sample statements.

Rahul Nair, Inge Vejsbjerg, Elizabeth Daly et al.

Humble AI (Knowles et al., 2023) argues for cautiousness in AI development and deployments through scepticism (accounting for limitations of statistical learning), curiosity (accounting for unexpected outcomes), and commitment (accounting for multifaceted values beyond performance). We present a real-world case study for humble AI in the domain of algorithmic hiring. Specifically, we evaluate virtual screening algorithms in a widely used hiring platform that matches candidates to job openings. There are several challenges in misrecognition and stereotyping in such contexts that are difficult to assess through standard fairness and trust frameworks; e.g., someone with a non-traditional background is less likely to rank highly. We demonstrate technical feasibility of how humble AI principles can be translated to practice through uncertainty quantification of ranks, entropy estimates, and a user experience that highlights algorithmic unknowns. We describe preliminary discussions with focus groups made up of recruiters. Future user studies seek to evaluate whether the higher cognitive load of a humble AI system fosters a climate of trust in its outcomes.

Jasmina Gajcin, Rahul Nair, Tejaswini Pedapati et al.

In complex tasks where the reward function is not straightforward and consists of a set of objectives, multiple reinforcement learning (RL) policies that perform task adequately, but employ different strategies can be trained by adjusting the impact of individual objectives on reward function. Understanding the differences in strategies between policies is necessary to enable users to choose between offered policies, and can help developers understand different behaviors that emerge from various reward functions and training hyperparameters in RL systems. In this work we compare behavior of two policies trained on the same task, but with different preferences in objectives. We propose a method for distinguishing between differences in behavior that stem from different abilities from those that are a consequence of opposing preferences of two RL agents. Furthermore, we use only data on preference-based differences in order to generate contrasting explanations about agents' preferences. Finally, we test and evaluate our approach on an autonomous driving task and compare the behavior of a safety-oriented policy and one that prefers speed.

Paulito P. Palmes, Akihiro Kishimoto, Radu Marinescu et al.

The pipeline optimization problem in machine learning requires simultaneous optimization of pipeline structures and parameter adaptation of their elements. Having an elegant way to express these structures can help lessen the complexity in the management and analysis of their performances together with the different choices of optimization strategies. With these issues in mind, we created the AutoMLPipeline (AMLP) toolkit which facilitates the creation and evaluation of complex machine learning pipeline structures using simple expressions. We use AMLP to find optimal pipeline signatures, datamine them, and use these datamined features to speed-up learning and prediction. We formulated a two-stage pipeline optimization with surrogate modeling in AMLP which outperforms other AutoML approaches with a 4-hour time budget in less than 5 minutes of AMLP computation time.

Oznur Alkan, Elizabeth Daly

Recommender systems are a valuable way to engage users in a system, increase participation and show them resources they may not have found otherwise. One significant challenge is that user interests may change over time and certain items have an inherently temporal aspect. As a result, a recommender system should try and take into account the time-dependant user-item relationships. However, temporal aspects of a user profile may not always be explicitly available and so we may need to infer this information from available resources. Product reviews on sites, such as Amazon, represent a valuable data source to understand why someone bought an item and potentially who the item is for. This information can then be used to construct a dynamic user profile. In this paper, we demonstrate utilising reviews to extract temporal information to infer the \textit{age category preference} of users, and leverage this feature to generate time-dependent recommendations. Given the predictable and yet shifting nature of age and time, we show that, recommendations generated using this dynamic aspect lead to higher accuracy compared with techniques from state of art. Mining temporally related content in reviews can enable the recommender to go beyond finding similar items or users to potentially predict a future need of a user.

Adi Botea, Christian Muise, Shubham Agarwal et al.

Dialogue systems have many applications such as customer support or question answering. Typically they have been limited to shallow single turn interactions. However more advanced applications such as career coaching or planning a trip require a much more complex multi-turn dialogue. Current limitations of conversational systems have made it difficult to support applications that require personalization, customization and context dependent interactions. We tackle this challenging problem by using domain-independent AI planning to automatically create dialogue plans, customized to guide a dialogue towards achieving a given goal. The input includes a library of atomic dialogue actions, an initial state of the dialogue, and a goal. Dialogue plans are plugged into a dialogue system capable to orchestrate their execution. Use cases demonstrate the viability of the approach. Our work on dialogue planning has been integrated into a product, and it is in the process of being deployed into another.