Reward models (RMs) are a crucial component in the alignment of large language models’ (LLMs) outputs with human values. RMs approximate human preferences over possible LLM responses to the same prompt by predicting and comparing reward scores. However, as they are typically modified versions of LLMs with scalar output heads, RMs are large black boxes whose predictions are not explainable. More transparent RMs would enable improved trust in the alignment of LLMs. In this work, we propose to use contrastive explanations to explain any binary response comparison made by an RM. Specifically, we generate a diverse set of new comparisons similar to the original one to characterise the RM’s local behaviour. The perturbed responses forming the new comparisons are generated to explicitly modify manually specified high-level evaluation attributes, on which analyses of RM behaviour are grounded. In quantitative experiments, we validate the effectiveness of our method for finding high-quality contrastive explanations. We then showcase the qualitative usefulness of our method for investigating global sensitivity of RMs to each evaluation attribute, and demonstrate how representative examples can be automatically extracted to explain and compare behaviours of different RMs. We see our method as a flexible framework for RM explanation, providing a basis for more interpretable and trustworthy LLM alignment.

Fine-tuning LLMs on tabular classification tasks can lead to the phenomenon of *fine-tuning multiplicity* where equally well-performing models make conflicting predictions on the same input. Fine-tuning multiplicity can arise due to variations in the training process, e.g., seed, weight initialization, minor changes to training data, etc., raising concerns about the reliability of Tabular LLMs in high-stakes applications such as finance, hiring, education, healthcare. Our work formalizes this unique challenge of fine-tuning multiplicity in Tabular LLMs and proposes a novel measure to quantify the consistency of individual predictions without expensive model retraining. Our measure quantifies a prediction's consistency by analyzing (sampling) the model's local behavior around that input in the embedding space. Interestingly, we show that sampling in the local neighborhood can be leveraged to provide probabilistic guarantees on prediction consistency under a broad class of fine-tuned models, i.e., inputs with sufficiently high local stability (as defined by our measure) also remain consistent across several fine-tuned models with high probability. We perform experiments on multiple real-world datasets to show that our local stability measure preemptively captures consistency under actual multiplicity across several fine-tuned models, outperforming competing measures.

We introduce Mechanistic Error Reduction with Abstention (MERA), a principled framework for steering language models (LMs) to mitigate errors through selective, adaptive interventions. Unlike existing methods that rely on fixed, manually tuned steering strengths, often resulting in under or oversteering, MERA addresses these limitations by (i) optimising the intervention direction, and (ii) calibrating when and how much to steer, thereby provably improving performance or abstaining when no confident correction is possible. Experiments across diverse datasets and LM families demonstrate safe, effective, non-degrading error correction and that MERA outperforms existing baselines. Moreover, MERA can be applied on top of existing steering techniques to further enhance their performance, establishing it as a general-purpose and efficient approach to mechanistic activation steering.

Test-time scaling improves large language models' (LLMs) performance by allocating more compute budget during inference. To achieve this, existing methods often require intricate modifications to prompting and sampling strategies. In this work, we introduce representation consistency (RC), a test-time scaling method for aggregating answers drawn from multiple candidate responses of an LLM regardless of how they were generated, including variations in prompt phrasing and sampling strategy. RC enhances answer aggregation by not only considering the number of occurrences of each answer in the candidate response set, but also the consistency of the model's internal activations while generating the set of responses leading to each answer. These activations can be either dense (raw model activations) or sparse (encoded via pretrained sparse autoencoders). Our rationale is that if the model's representations of multiple responses converging on the same answer are highly variable, this answer is more likely to be the result of incoherent reasoning and should be down-weighted during aggregation. Importantly, our method only uses cached activations and lightweight similarity computations and requires no additional model queries. Through experiments with four open-source LLMs and four reasoning datasets, we validate the effectiveness of RC for improving task performance during inference, with consistent accuracy improvements (up to 4\%) over strong test-time scaling baselines. We also show that consistency in the sparse activation signals aligns well with the common notion of coherent reasoning.

Models for natural language and images benefit from data scaling behavior: the more data fed into the model, the better they perform. This 'better with more' phenomenon enables the effectiveness of large-scale pre-training on vast amounts of data. However, current graph pre-training methods struggle to scale up data due to heterogeneity across graphs. To achieve effective data scaling, we aim to develop a general model that is able to capture diverse data patterns of graphs and can be utilized to adaptively help the downstream tasks. To this end, we propose UniAug, a universal graph structure augmentor built on a diffusion model. We first pre-train a discrete diffusion model on thousands of graphs across domains to learn the graph structural patterns. In the downstream phase, we provide adaptive enhancement by conducting graph structure augmentation with the help of the pre-trained diffusion model via guided generation. By leveraging the pre-trained diffusion model for structure augmentation, we consistently achieve performance improvements across various downstream tasks in a plug-and-play manner. To the best of our knowledge, this study represents the first demonstration of a data-scaling graph structure augmentor on graphs across domains.

Polyhedral geometry can be used to shed light on the behaviour of piecewise linear neural networks, such as ReLU-based architectures. Counterfactual explanations are a popular class of methods for examining model behaviour by comparing a query to the closest point with a different label, subject to constraints. We present a new algorithm, Polyhedral-complex Informed Counterfactual Explanations (PICE), which leverages the decomposition of the piecewise linear neural network into a polyhedral complex to find counterfactuals that are provably minimal in the Euclidean norm and exactly on the decision boundary for any given query. Moreover, we develop variants of the algorithm that target popular counterfactual desiderata such as sparsity, robustness, speed, plausibility, and actionability. We empirically show on four publicly available real-world datasets that our method outperforms other popular techniques to find counterfactuals and adversarial attacks by distance to decision boundary and distance to query. Moreover, we successfully improve our baseline method in the dimensions of the desiderata we target, as supported by experimental evaluations.

We introduce T-CREx, a novel model-agnostic method for local and global counterfactual explanation (CE), which summarises recourse options for both individuals and groups in the form of generalised rules. It leverages tree-based surrogate models to learn the counterfactual rules, alongside metarules denoting their regimes of optimality, providing both a global analysis of model behaviour and diverse recourse options for users. Experiments indicate that T-CREx achieves superior aggregate performance over existing rule-based baselines on a range of CE desiderata, while being orders of magnitude faster to run.

This paper proposes Progressive inference–a framework to explain the predictions of decoder-only transformer models trained to perform sequence classification tasks. Our work is based on the insight that the classification head of a decoder-only model can be used to make intermediate predictions by evaluating them at different points in the input sequence. Due to the masked attention mechanism used in decoder-only models, these intermediate predictions only depend on the tokens seen before the inference point, allowing us to obtain the model’s prediction on a masked input sub-sequence, with negligible computational overheads. We develop two methods to provide sub-sequence level attributions using this core insight. First, we propose Single Pass-Progressive Inference (SP-PI) to compute attributions by simply taking the difference between intermediate predictions. Second, we exploit a connection with Kernel SHAP to develop Multi Pass-Progressive Inference (MP-PI); this uses intermediate predictions from multiple masked versions of the input to compute higher-quality attributions that approximate SHAP values. We perform studies on several text classification datasets to demonstrate that our proposal provides better explanations compared to prior work, both in the single-pass and multi-pass settings.

We introduce a novel approach for detecting distribution shifts that negatively impact the performance of machine learning models in continuous production environments, which requires no access to ground truth data labels. It builds upon the work of Podkopaev and Ramdas [2022], who address scenarios where labels are available for tracking model errors over time. Our solution extends this framework to work in the absence of labels, by employing a proxy for the true error. This proxy is derived using the predictions of a trained error estimator. Experiments show that our method has high power and false alarm control under various distribution shifts, including covariate and label shifts and natural shifts over geography and time.

Shape assembly is a ubiquitous task in daily life, integral for constructing complex 3D structures like IKEA furniture. While significant progress has been made in developing autonomous agents for shape assembly, existing datasets have not yet tackled the 4D grounding of assembly instructions in videos, essential for a holistic understanding of assembly in 3D space over time. We introduce IKEA Video Manuals, a dataset that features 3D models of furniture parts, instructional manuals, assembly videos from the Internet, and most importantly, annotations of dense spatio-temporal alignments between these data modalities. To demonstrate the utility of IKEA Video Manuals, we present five applications essential for shape assembly: assembly plan generation, part-conditioned segmentation, part-conditioned pose estimation, video object segmentation, and furniture assembly based on instructional video manuals. For each application, we provide evaluation metrics and baseline methods. Through experiments on our annotated data, we highlight many challenges in grounding assembly instructions in videos to improve shape assembly, including handling occlusions, varying viewpoints, and extended assembly sequences.

One approach for interpreting black-box machine learning models is to find a global approximation of the model using simple interpretable functions, which is called a metamodel (a model of the model). Approximating the black-box with a metamodel can be used to 1) estimate instance-wise feature importance; 2) understand the functional form of the model; 3) analyze feature interactions. In this work, we propose a new method for finding interpretable metamodels. Our approach utilizes Kolmogorov superposition theorem, which expresses multivariate functions as a composition of univariate functions (our primitive parameterized functions). This composition can be represented in the form of a tree. Inspired by symbolic regression, we use a modified form of genetic programming to search over different tree configurations. Gradient descent (GD) is used to optimize the parameters of a given configuration. Our method is a novel memetic algorithm that uses GD not only for training numerical constants but also for the training of building blocks. Using several experiments, we show that our method outperforms recent metamodeling approaches suggested for interpreting black-boxes.

Explainable Artificial Intelligence (XAI) has received widespread interest in recent years, and two of the most popular types of explanations are feature attributions, and counterfactual explanations. These classes of approaches have been largely studied independently and the few attempts at reconciling them have been primarily empirical. This work establishes a clear theoretical connection between game-theoretic feature attributions, focusing on but not limited to SHAP, and counterfactuals explanations. After motivating operative changes to Shapley values based feature attributions and counterfactual explanations, we prove that, under conditions, they are in fact equivalent. We then extend the equivalency result to game-theoretic solution concepts beyond Shapley values. Moreover, through the analysis of the conditions of such equivalence, we shed light on the limitations of naively using counterfactual explanations to provide feature importances. Experiments on three datasets quantitatively show the difference in explanations at every stage of the connection between the two approaches and corroborate the theoretical findings.

There is an emerging interest in generating robust counterfactual explanations that would remain valid if the model is updated or changed even slightly. Towards finding robust counterfactuals, existing literature often assumes that the original model $m$ and the new model $M$ are bounded in the parameter space, i.e., $\|\text{Params}(M){-}\text{Params}(m)\|{<}\Delta$. However, models can often change significantly in the parameter space with little to no change in their predictions or accuracy on the given dataset. In this work, we introduce a mathematical abstraction termed naturally-occurring model change, which allows for arbitrary changes in the parameter space such that the change in predictions on points that lie on the data manifold is limited. Next, we propose a measure – that we call Stability – to quantify the robustness of counterfactuals to potential model changes for differentiable models, e.g., neural networks. Our main contribution is to show that counterfactuals with sufficiently high value of Stability as defined by our measure will remain valid after potential “naturally-occurring” model changes with high probability (leveraging concentration bounds for Lipschitz function of independent Gaussians). Since our quantification depends on the local Lipschitz constant around a data point which is not always available, we also examine practical relaxations of our proposed measure and demonstrate experimentally how they can be incorporated to find robust counterfactuals for neural networks that are close, realistic, and remain valid after potential model changes.

Counterfactual explanations have been widely studied in explainability, with a range of application dependent methods prominent in fairness, recourse and model understanding. The major shortcoming associated with these methods, however, is their inability to provide explanations beyond the local or instance-level. While many works touch upon the notion of a global explanation, typically suggesting to aggregate masses of local explanations in the hope of ascertaining global properties, few provide frameworks that are both reliable and computationally tractable. Meanwhile, practitioners are requesting more efficient and interactive explainability tools. We take this opportunity to propose Global & Efficient Counterfactual Explanations (GLOBE-CE), a flexible framework that tackles the reliability and scalability issues associated with current state-of-the-art, particularly on higher dimensional datasets and in the presence of continuous features. Furthermore, we provide a unique mathematical analysis of categorical feature translations, utilising it in our method. Experimental evaluation with publicly available datasets and user studies demonstrate that GLOBE-CE performs significantly better than the current state-of-the-art across multiple metrics (e.g., speed, reliability).

Counterfactual explanations inform ways to achieve a desired outcome from a machine learning model. However, such explanations are not robust to certain real-world changes in the underlying model (e.g., retraining the model, changing hyperparameters, etc.), questioning their reliability in several applications, e.g., credit lending. In this work, we propose a novel strategy - that we call RobX - to generate robust counterfactuals for tree-based ensembles, e.g., XGBoost. Tree-based ensembles pose additional challenges in robust counterfactual generation, e.g., they have a non-smooth and non-differentiable objective function, and they can change a lot in the parameter space under retraining on very similar data. We first introduce a novel metric - that we call Counterfactual Stability - that attempts to quantify how robust a counterfactual is going to be to model changes under retraining, and comes with desirable theoretical properties. Our proposed strategy RobX works with any counterfactual generation method (base method) and searches for robust counterfactuals by iteratively refining the counterfactual generated by the base method using our metric Counterfactual Stability. We compare the performance of RobX with popular counterfactual generation methods (for tree-based ensembles) across benchmark datasets. The results demonstrate that our strategy generates counterfactuals that are significantly more robust (nearly 100% validity after actual model changes) and also realistic (in terms of local outlier factor) over existing state-of-the-art methods.

Counterfactual explanations promote explainability in machine learning models by answering the question “how should the input instance be altered to obtain a desired predicted label?". The comparison of this instance before and after perturbation can enhance human interpretation. Most existing studies on counterfactual explanations are limited in tabular data or image data. In this paper, we study the problem of counterfactual explanation generation on graphs. A few studies have explored to generate counterfactual explanations on graphs, but many challenges of this problem are still not well-addressed: 1) optimizing in the discrete and disorganized space of graphs; 2) generalizing on unseen graphs; 3) maintaining the causality in the generated counterfactuals without prior knowledge of the causal model. To tackle these challenges, we propose a novel framework CLEAR which aims to generate counterfactual explanations on graphs for graph-level prediction models. Specifically, CLEAR leverages a graph variational autoencoder based mechanism to facilitate its optimization and generalization, and promotes causality by leveraging an auxiliary variable to better identify the causal model. Extensive experiments on both synthetic and real-world graphs validate the superiority of CLEAR over state-of-the-art counterfactual explanation methods on graphs in different aspects.

One way to analyse the behaviour of machine learning models is through local explanations that highlight input features that maximally influence model predictions. Sensitivity analysis, which involves analysing the effect of input perturbations on model predictions, is one of the methods to generate local explanations. Meaningful input perturbations are essential for generating reliable explanations, but there exists limited work on what such perturbations are and how to perform them. This work investigates these questions in the context of machine listening models that analyse audio. Specifically, we use a state-of-the-art deep singing voice detection (SVD) model to analyse whether explanations from SoundLIME (a local explanation method) are sensitive to how the method perturbs model inputs. The results demonstrate that SoundLIME explanations are sensitive to the content in the occluded input regions. We further propose and demonstrate a novel method for quantitatively identifying suitable content type(s) for reliably occluding inputs of machine listening models. The results for the SVD model suggest that the average magnitude of input mel-spectrogram bins is the most suitable content type for temporal explanations.

Playing recorded speech samples of an enrolled speaker - “replay attack” - is a simple approach to bypass an automatic speaker verification (ASV) system. The vulnerability of ASV systems to such attacks has been acknowledged and studied, but there has been no research into what spoofing detection systems are actually learning to discriminate. In this paper, we analyse the local behaviour of a replay spoofing detection system based on convolutional neural networks (CNNs) adapted from a state-of-the-art CNN (LCNNFFT) submitted at the ASVspoof 2017 challenge. We generate temporal and spectral explanations for predictions of the model using the SLIME algorithm. Our findings suggest that in most instances of spoofing the model is using information in the first 400 milliseconds of each audio instance to make the class prediction. Knowledge of the characteristics that spoofing detection systems are exploiting can help build less vulnerable ASV systems, other spoofing detection systems, as well as better evaluation databases1.