Chris Russell

Florian Wenzel, Andrea Dittadi, Peter Vincent Gehler et al. · eth-zurich

Since out-of-distribution generalization is a generally ill-posed problem, various proxy targets (e.g., calibration, adversarial robustness, algorithmic corruptions, invariance across shifts) were studied across different research programs resulting in different recommendations. While sharing the same aspirational goal, these approaches have never been tested under the same experimental conditions on real data. In this paper, we take a unified view of previous work, highlighting message discrepancies that we address empirically, and providing recommendations on how to measure the robustness of a model and how to improve it. To this end, we collect 172 publicly available dataset pairs for training and out-of-distribution evaluation of accuracy, calibration error, adversarial attacks, environment invariance, and synthetic corruptions. We fine-tune over 31k networks, from nine different architectures in the many- and few-shot setting. Our findings confirm that in- and out-of-distribution accuracies tend to increase jointly, but show that their relation is largely dataset-dependent, and in general more nuanced and more complex than posited by previous, smaller scale studies.

Jaime Spencer, Chris Russell, Simon Hadfield et al.

This paper presents an open and comprehensive framework to systematically evaluate state-of-the-art contributions to self-supervised monocular depth estimation. This includes pretraining, backbone, architectural design choices and loss functions. Many papers in this field claim novelty in either architecture design or loss formulation. However, simply updating the backbone of historical systems results in relative improvements of 25%, allowing them to outperform the majority of existing systems. A systematic evaluation of papers in this field was not straightforward. The need to compare like-with-like in previous papers means that longstanding errors in the evaluation protocol are ubiquitous in the field. It is likely that many papers were not only optimized for particular datasets, but also for errors in the data and evaluation criteria. To aid future research in this area, we release a modular codebase (https://github.com/jspenmar/monodepth_benchmark), allowing for easy evaluation of alternate design decisions against corrected data and evaluation criteria. We re-implement, validate and re-evaluate 16 state-of-the-art contributions and introduce a new dataset (SYNS-Patches) containing dense outdoor depth maps in a variety of both natural and urban scenes. This allows for the computation of informative metrics in complex regions such as depth boundaries.

Jaime Spencer, Chris Russell, Simon Hadfield et al.

Self-supervised monocular depth estimation (SS-MDE) has the potential to scale to vast quantities of data. Unfortunately, existing approaches limit themselves to the automotive domain, resulting in models incapable of generalizing to complex environments such as natural or indoor settings. To address this, we propose a large-scale SlowTV dataset curated from YouTube, containing an order of magnitude more data than existing automotive datasets. SlowTV contains 1.7M images from a rich diversity of environments, such as worldwide seasonal hiking, scenic driving and scuba diving. Using this dataset, we train an SS-MDE model that provides zero-shot generalization to a large collection of indoor/outdoor datasets. The resulting model outperforms all existing SSL approaches and closes the gap on supervised SoTA, despite using a more efficient architecture. We additionally introduce a collection of best-practices to further maximize performance and zero-shot generalization. This includes 1) aspect ratio augmentation, 2) camera intrinsic estimation, 3) support frame randomization and 4) flexible motion estimation. Code is available at https://github.com/jspenmar/slowtv_monodepth.

Dominik Zietlow, Michael Lohaus, Guha Balakrishnan et al.

Algorithmic fairness is frequently motivated in terms of a trade-off in which overall performance is decreased so as to improve performance on disadvantaged groups where the algorithm would otherwise be less accurate. Contrary to this, we find that applying existing fairness approaches to computer vision improve fairness by degrading the performance of classifiers across all groups (with increased degradation on the best performing groups). Extending the bias-variance decomposition for classification to fairness, we theoretically explain why the majority of fairness classifiers designed for low capacity models should not be used in settings involving high-capacity models, a scenario common to computer vision. We corroborate this analysis with extensive experimental support that shows that many of the fairness heuristics used in computer vision also degrade performance on the most disadvantaged groups. Building on these insights, we propose an adaptive augmentation strategy that, uniquely, of all methods tested, improves performance for the disadvantaged groups.

Pengyun Wang, Yadi Cao, Chris Russell et al. · pku

Graph domain adaptation has recently enabled knowledge transfer across different graphs. However, without the semantic information on target graphs, the performance on target graphs is still far from satisfactory. To address the issue, we study the problem of active graph domain adaptation, which selects a small quantitative of informative nodes on the target graph for extra annotation. This problem is highly challenging due to the complicated topological relationships and the distribution discrepancy across graphs. In this paper, we propose a novel approach named Dual Consistency Delving with Topological Uncertainty (DELTA) for active graph domain adaptation. Our DELTA consists of an edge-oriented graph subnetwork and a path-oriented graph subnetwork, which can explore topological semantics from complementary perspectives. In particular, our edge-oriented graph subnetwork utilizes the message passing mechanism to learn neighborhood information, while our path-oriented graph subnetwork explores high-order relationships from sub-structures. To jointly learn from two subnetworks, we roughly select informative candidate nodes with the consideration of consistency across two subnetworks. Then, we aggregate local semantics from its K-hop subgraph based on node degrees for topological uncertainty estimation. To overcome potential distribution shifts, we compare target nodes and their corresponding source nodes for discrepancy scores as an additional component for fine selection. Extensive experiments on benchmark datasets demonstrate that DELTA outperforms various state-of-the-art approaches. The code implementation of DELTA is available at https://github.com/goose315/DELTA.

Yuejiang Liu, Alexandre Alahi, Chris Russell et al.

Recent years have seen a surge of interest in learning high-level causal representations from low-level image pairs under interventions. Yet, existing efforts are largely limited to simple synthetic settings that are far away from real-world problems. In this paper, we present Causal Triplet, a causal representation learning benchmark featuring not only visually more complex scenes, but also two crucial desiderata commonly overlooked in previous works: (i) an actionable counterfactual setting, where only certain object-level variables allow for counterfactual observations whereas others do not; (ii) an interventional downstream task with an emphasis on out-of-distribution robustness from the independent causal mechanisms principle. Through extensive experiments, we find that models built with the knowledge of disentangled or object-centric representations significantly outperform their distributed counterparts. However, recent causal representation learning methods still struggle to identify such latent structures, indicating substantial challenges and opportunities for future work. Our code and datasets will be available at https://sites.google.com/view/causaltriplet.

Ryan Brown, Chris Russell

We introduce Adaptive Calibration (AC), a novel calibration strategy for facial recognition that maps cosine similarity between normalized embeddings to well-calibrated probabilities. By incorporating local context into calibration, Adaptive Calibration corrects for a fundamental mismatch in cosine similarity, whereby the same distance can correspond to different match probabilities in different embedding regions. Our approach improves both overall performance and results in a fairer calibration without requiring demographic metadata. Our approach consistently dominates existing methods both on accuracy and fairness metrics across a variety of pretrained models and standard benchmarks. AC provides a practical solution for equitable facial recognition, without requiring demographic group annotations, and while improving overall performance. Unlike existing approaches, our method provides continuous, region-specific calibration that avoids "leveling down" where fairness comes at the cost of degraded performance for some groups.

Guillaume Rochette, Chris Russell, Richard Bowden

We present a new approach for synthesizing novel views of people in new poses. Our novel differentiable renderer enables the synthesis of highly realistic images from any viewpoint. Rather than operating over mesh-based structures, our renderer makes use of diffuse Gaussian primitives that directly represent the underlying skeletal structure of a human. Rendering these primitives gives results in a high-dimensional latent image, which is then transformed into an RGB image by a decoder network. The formulation gives rise to a fully differentiable framework that can be trained end-to-end. We demonstrate the effectiveness of our approach to image reconstruction on both the Human3.6M and Panoptic Studio datasets. We show how our approach can be used for motion transfer between individuals; novel view synthesis of individuals captured from just a single camera; to synthesize individuals from any virtual viewpoint; and to re-render people in novel poses. Code and video results are available at https://github.com/GuillaumeRochette/HumanViewSynthesis.

Matthäus Kleindessner, Michele Donini, Chris Russell et al. · amazon-science

We revisit the problem of fair principal component analysis (PCA), where the goal is to learn the best low-rank linear approximation of the data that obfuscates demographic information. We propose a conceptually simple approach that allows for an analytic solution similar to standard PCA and can be kernelized. Our methods have the same complexity as standard PCA, or kernel PCA, and run much faster than existing methods for fair PCA based on semidefinite programming or manifold optimization, while achieving similar results.

Yash Sharma, Yi Zhu, Chris Russell et al.

While self-supervised learning has enabled effective representation learning in the absence of labels, for vision, video remains a relatively untapped source of supervision. To address this, we propose Pixel-level Correspondence (PiCo), a method for dense contrastive learning from video. By tracking points with optical flow, we obtain a correspondence map which can be used to match local features at different points in time. We validate PiCo on standard benchmarks, outperforming self-supervised baselines on multiple dense prediction tasks, without compromising performance on image classification.

Jaime Spencer, C. Stella Qian, Michaela Trescakova et al.

This paper discusses the results for the second edition of the Monocular Depth Estimation Challenge (MDEC). This edition was open to methods using any form of supervision, including fully-supervised, self-supervised, multi-task or proxy depth. The challenge was based around the SYNS-Patches dataset, which features a wide diversity of environments with high-quality dense ground-truth. This includes complex natural environments, e.g. forests or fields, which are greatly underrepresented in current benchmarks. The challenge received eight unique submissions that outperformed the provided SotA baseline on any of the pointcloud- or image-based metrics. The top supervised submission improved relative F-Score by 27.62%, while the top self-supervised improved it by 16.61%. Supervised submissions generally leveraged large collections of datasets to improve data diversity. Self-supervised submissions instead updated the network architecture and pretrained backbones. These results represent a significant progress in the field, while highlighting avenues for future research, such as reducing interpolation artifacts at depth boundaries, improving self-supervised indoor performance and overall natural image accuracy.

Jaime Spencer, C. Stella Qian, Chris Russell et al.

This paper summarizes the results of the first Monocular Depth Estimation Challenge (MDEC) organized at WACV2023. This challenge evaluated the progress of self-supervised monocular depth estimation on the challenging SYNS-Patches dataset. The challenge was organized on CodaLab and received submissions from 4 valid teams. Participants were provided a devkit containing updated reference implementations for 16 State-of-the-Art algorithms and 4 novel techniques. The threshold for acceptance for novel techniques was to outperform every one of the 16 SotA baselines. All participants outperformed the baseline in traditional metrics such as MAE or AbsRel. However, pointcloud reconstruction metrics were challenging to improve upon. We found predictions were characterized by interpolation artefacts at object boundaries and errors in relative object positioning. We hope this challenge is a valuable contribution to the community and encourage authors to participate in future editions.

Junaid Ali, Matthaeus Kleindessner, Florian Wenzel et al.

We propose a novel taxonomy for bias evaluation of discriminative foundation models, such as Contrastive Language-Pretraining (CLIP), that are used for labeling tasks. We then systematically evaluate existing methods for mitigating bias in these models with respect to our taxonomy. Specifically, we evaluate OpenAI's CLIP and OpenCLIP models for key applications, such as zero-shot classification, image retrieval and image captioning. We categorize desired behaviors based around three axes: (i) if the task concerns humans; (ii) how subjective the task is (i.e., how likely it is that people from a diverse range of backgrounds would agree on a labeling); and (iii) the intended purpose of the task and if fairness is better served by impartiality (i.e., making decisions independent of the protected attributes) or representation (i.e., making decisions to maximize diversity). Finally, we provide quantitative fairness evaluations for both binary-valued and multi-valued protected attributes over ten diverse datasets. We find that fair PCA, a post-processing method for fair representations, works very well for debiasing in most of the aforementioned tasks while incurring only minor loss of performance. However, different debiasing approaches vary in their effectiveness depending on the task. Hence, one should choose the debiasing approach depending on the specific use case.

Brent Mittelstadt, Sandra Wachter, Chris Russell

In recent years fairness in machine learning (ML) has emerged as a highly active area of research and development. Most define fairness in simple terms, where fairness means reducing gaps in performance or outcomes between demographic groups while preserving as much of the accuracy of the original system as possible. This oversimplification of equality through fairness measures is troubling. Many current fairness measures suffer from both fairness and performance degradation, or "levelling down," where fairness is achieved by making every group worse off, or by bringing better performing groups down to the level of the worst off. When fairness can only be achieved by making everyone worse off in material or relational terms through injuries of stigma, loss of solidarity, unequal concern, and missed opportunities for substantive equality, something would appear to have gone wrong in translating the vague concept of 'fairness' into practice. This paper examines the causes and prevalence of levelling down across fairML, and explore possible justifications and criticisms based on philosophical and legal theories of equality and distributive justice, as well as equality law jurisprudence. We find that fairML does not currently engage in the type of measurement, reporting, or analysis necessary to justify levelling down in practice. We propose a first step towards substantive equality in fairML: "levelling up" systems by design through enforcement of minimum acceptable harm thresholds, or "minimum rate constraints," as fairness constraints. We likewise propose an alternative harms-based framework to counter the oversimplified egalitarian framing currently dominant in the field and push future discussion more towards substantive equality opportunities and away from strict egalitarianism by default. N.B. Shortened abstract, see paper for full abstract.

Avishkar Saha, Oscar Mendez, Chris Russell et al.

Graph convolutional networks (GCNs) enable end-to-end learning on graph structured data. However, many works assume a given graph structure. When the input graph is noisy or unavailable, one approach is to construct or learn a latent graph structure. These methods typically fix the choice of node degree for the entire graph, which is suboptimal. Instead, we propose a novel end-to-end differentiable graph generator which builds graph topologies where each node selects both its neighborhood and its size. Our module can be readily integrated into existing pipelines involving graph convolution operations, replacing the predetermined or existing adjacency matrix with one that is learned, and optimized, as part of the general objective. As such it is applicable to any GCN. We integrate our module into trajectory prediction, point cloud classification and node classification pipelines resulting in improved accuracy over other structure-learning methods across a wide range of datasets and GCN backbones.

Avishkar Saha, Oscar Mendez, Chris Russell et al.

Estimating a semantically segmented bird's-eye-view (BEV) map from a single image has become a popular technique for autonomous control and navigation. However, they show an increase in localization error with distance from the camera. While such an increase in error is entirely expected - localization is harder at distance - much of the drop in performance can be attributed to the cues used by current texture-based models, in particular, they make heavy use of object-ground intersections (such as shadows), which become increasingly sparse and uncertain for distant objects. In this work, we address these shortcomings in BEV-mapping by learning the spatial relationship between objects in a scene. We propose a graph neural network which predicts BEV objects from a monocular image by spatially reasoning about an object within the context of other objects. Our approach sets a new state-of-the-art in BEV estimation from monocular images across three large-scale datasets, including a 50% relative improvement for objects on nuScenes.

Michael Lohaus, Matthäus Kleindessner, Krishnaram Kenthapadi et al.

We show that deep networks trained to satisfy demographic parity often do so through a form of race or gender awareness, and that the more we force a network to be fair, the more accurately we can recover race or gender from the internal state of the network. Based on this observation, we investigate an alternative fairness approach: we add a second classification head to the network to explicitly predict the protected attribute (such as race or gender) alongside the original task. After training the two-headed network, we enforce demographic parity by merging the two heads, creating a network with the same architecture as the original network. We establish a close relationship between existing approaches and our approach by showing (1) that the decisions of a fair classifier are well-approximated by our approach, and (2) that an unfair and optimally accurate classifier can be recovered from a fair classifier and our second head predicting the protected attribute. We use our explicit formulation to argue that the existing fairness approaches, just as ours, demonstrate disparate treatment and that they are likely to be unlawful in a wide range of scenarios under US law.

Max F. Burg, Florian Wenzel, Dominik Zietlow et al.

Many approaches have been proposed to use diffusion models to augment training datasets for downstream tasks, such as classification. However, diffusion models are themselves trained on large datasets, often with noisy annotations, and it remains an open question to which extent these models contribute to downstream classification performance. In particular, it remains unclear if they generalize enough to improve over directly using the additional data of their pre-training process for augmentation. We systematically evaluate a range of existing methods to generate images from diffusion models and study new extensions to assess their benefit for data augmentation. Personalizing diffusion models towards the target data outperforms simpler prompting strategies. However, using the pre-training data of the diffusion model alone, via a simple nearest-neighbor retrieval procedure, leads to even stronger downstream performance. Our study explores the potential of diffusion models in generating new training data, and surprisingly finds that these sophisticated models are not yet able to beat a simple and strong image retrieval baseline on simple downstream vision tasks.

William Lugoloobi, Samuelle Marro, Jabez Magomere et al.

As LLM-based agents increasingly browse the web on users' behalf, a natural question arises: can websites passively identify which underlying model powers an agent? Doing so would represent a significant security risk, enabling targeted attacks tailored to known model vulnerabilities. Across 14 frontier LLMs and four web environments spanning information retrieval and shopping tasks, we show that an agent's actions and interaction timings, captured via a passive JavaScript tracker, are sufficient to identify the underlying model with up to 96\% F1. We formalise this attack surface by demonstrating that classifiers trained on agent actions generalise across model sizes and families. We further show that strong classifiers can be trained from few interaction traces and that agent identity can be inferred early within an episode. Injecting randomised timing delays between actions substantially degrades classifier performance, but does not provide robust protection: a classifier retrained on delayed traces largely recovers performance. We release our harness and a labelled corpus of agent traces \href{https://github.com/KabakaWilliam/known_actions}{here}.

Jonathan Rystrøm, Sofie Burgos-Thorsen, Zihao Fu et al.

Text embeddings are widely used to analyse large corpora of complex texts. However, it is unclear whether the embeddings capture the same semantic distances as the human experts using them. Ensuring alignment between embedding representations and human intentions is essential for valid analyses. We present the Stakeholder Grounding Exercise, a method for making expert associations explicit and grounding embedding model results in human understanding. In our primary case study on Danish policy issues, we find that neural text embeddings are substantially less reliable than human experts (19-26 pp gap), and that this misalignment propagates to downstream clustering performance (Spearman $ρ=0.9$ between exercise ranking and cluster quality). A secondary study on US Federal AI use cases replicates the gap (16pp) in English, using a digital protocol and a different community of experts -- demonstrating that the gap is not an artefact of a single instrument or domain. The Stakeholder Grounding Exercise offers a practical method for assessing whether embedding models capture the semantic distinctions that matter most to domain experts.

William Lugoloobi, Thomas Foster, William Bankes et al.

Running LLMs with extended reasoning on every problem is expensive, but determining which inputs actually require additional compute remains challenging. We investigate whether their own likelihood of success is recoverable from their internal representations before generation, and if this signal can guide more efficient inference. We train linear probes on pre-generation activations to predict policy-specific success on math and coding tasks, substantially outperforming surface features such as question length and TF-IDF. Using E2H-AMC, which provides both human and model performance on identical problems, we show that models encode a model-specific notion of difficulty that is distinct from human difficulty, and that this distinction increases with extended reasoning. Leveraging these probes, we demonstrate that routing queries across a pool of models can exceed the best-performing model whilst reducing inference cost by up to 70\% on MATH, showing that internal representations enable practical efficiency gains even when they diverge from human intuitions about difficulty. Our code is available at: https://github.com/KabakaWilliam/llms_know_difficulty

Jaime Spencer, Chris Russell, Simon Hadfield et al.

Self-supervised learning is the key to unlocking generic computer vision systems. By eliminating the reliance on ground-truth annotations, it allows scaling to much larger data quantities. Unfortunately, self-supervised monocular depth estimation (SS-MDE) has been limited by the absence of diverse training data. Existing datasets have focused exclusively on urban driving in densely populated cities, resulting in models that fail to generalize beyond this domain. To address these limitations, this paper proposes two novel datasets: SlowTV and CribsTV. These are large-scale datasets curated from publicly available YouTube videos, containing a total of 2M training frames. They offer an incredibly diverse set of environments, ranging from snowy forests to coastal roads, luxury mansions and even underwater coral reefs. We leverage these datasets to tackle the challenging task of zero-shot generalization, outperforming every existing SS-MDE approach and even some state-of-the-art supervised methods. The generalization capabilities of our models are further enhanced by a range of components and contributions: 1) learning the camera intrinsics, 2) a stronger augmentation regime targeting aspect ratio changes, 3) support frame randomization, 4) flexible motion estimation, 5) a modern transformer-based architecture. We demonstrate the effectiveness of each component in extensive ablation experiments. To facilitate the development of future research, we make the datasets, code and pretrained models available to the public at https://github.com/jspenmar/slowtv_monodepth.

Karolina Korgul, Yushi Yang, Arkadiusz Drohomirecki et al.

Web-based agents powered by large language models are increasingly used for tasks such as email management or professional networking. Their reliance on dynamic web content, however, makes them vulnerable to prompt injection attacks: adversarial instructions hidden in interface elements that persuade the agent to divert from its original task. We introduce the Task-Redirecting Agent Persuasion Benchmark (TRAP), an evaluation for studying how persuasion techniques misguide autonomous web agents on realistic tasks. Across six frontier models, agents are susceptible to prompt injection in 25\% of tasks on average (13\% for GPT-5 to 43\% for DeepSeek-R1), with small interface or contextual changes often doubling success rates and revealing systemic, psychologically driven vulnerabilities in web-based agents. We also provide a modular social-engineering injection framework with controlled experiments on high-fidelity website clones, allowing for further benchmark expansion.

Stratis Tsirtsis, Kai Rawal, Chris Russell et al.

Generative artificial intelligence (AI) is increasingly integrated into the online platforms where humans exchange opinions; large language models (LLMs) now polish users' posts on LinkedIn and provide context for content shared on X. While prior work has shown that AI can express biased opinions and shape individuals' opinions during human-AI interactions, less attention has been paid to its influence on collective opinion formation when mediating human-to-human communication. We address this gap via a combination of empirical and theoretical analyses. We show empirically that LLMs from multiple popular families introduce directional biases when instructed to edit human-written texts on contested topics, for example, nudging texts in favor of gun control and against atheism. Building on this observation, we introduce a mathematical model of opinion dynamics in which an AI system sits between users on a social network, transforming the opinions they express and perceive. By analytically characterizing the equilibrium of this model and performing simulations on real social network data, we show that biases introduced by AI in human-to-human communication can be amplified through the network and shift collective opinion in their direction. In light of these findings, we investigate whether such biases are controllable by online platforms. We audit the "Explain this post" feature on X and find evidence of pro-life bias in Grok's outputs on abortion-related content, which we trace back to specific design choices. We conclude with a discussion of the broader implications of our findings in relation to ongoing legislative efforts in the European Union.

Harry Mayne, Ryan Othniel Kearns, Yushi Yang et al.

To collaborate effectively with humans, language models must be able to explain their decisions in natural language. We study a specific type of self-explanation: self-generated counterfactual explanations (SCEs), where a model explains its prediction by modifying the input such that it would have predicted a different outcome. We evaluate whether LLMs can produce SCEs that are valid, achieving the intended outcome, and minimal, modifying the input no more than necessary. When asked to generate counterfactuals, we find that LLMs typically produce SCEs that are valid, but far from minimal, offering little insight into their decision-making behaviour. Worryingly, when asked to generate minimal counterfactuals, LLMs typically make excessively small edits that fail to change predictions. The observed validity-minimality trade-off is consistent across several LLMs, datasets, and evaluation settings. Our findings suggest that SCEs are, at best, an ineffective explainability tool and, at worst, can provide misleading insights into model behaviour. Proposals to deploy LLMs in high-stakes settings must consider the impact of unreliable self-explanations on downstream decision-making. Our code is available at https://github.com/HarryMayne/SCEs.

Kaivalya Rawal, Zihao Fu, Eoin Delaney et al.

There can be many competing and contradictory explanations for a single model prediction, making it difficult to select which one to use. Current explanation evaluation frameworks measure quality by comparing against ideal "ground-truth" explanations, or by verifying model sensitivity to important inputs. We outline the limitations of these approaches, and propose three desirable principles to ground the future development of explanation evaluation strategies for local feature importance explanations. We propose a ground-truth Agnostic eXplanation Evaluation framework (AXE) for evaluating and comparing model explanations that satisfies these principles. Unlike prior approaches, AXE does not require access to ideal ground-truth explanations for comparison, or rely on model sensitivity - providing an independent measure of explanation quality. We verify AXE by comparing with baselines, and show how it can be used to detect explanation fairwashing. Our code is available at https://github.com/KaiRawal/Evaluating-Model-Explanations-without-Ground-Truth.

Jonathan Rystrøm, Zihao Fu, Chris Russell

We address the problem of fair classification in settings where data is scarce and unbalanced across demographic groups. Such low-data regimes are common in domains like medical imaging, where false negatives can have fatal consequences. We propose a novel approach \emph{OxEnsemble} for efficiently training ensembles and enforcing fairness in these low-data regimes. Unlike other approaches, we aggregate predictions across ensemble members, each trained to satisfy fairness constraints. By construction, \emph{OxEnsemble} is both data-efficient, carefully reusing held-out data to enforce fairness reliably, and compute-efficient, requiring little more compute than used to fine-tune or evaluate an existing model. We validate this approach with new theoretical guarantees. Experimentally, our approach yields more consistent outcomes and stronger fairness-accuracy trade-offs than existing methods across multiple challenging medical imaging classification datasets.

Jonathan Rystrøm, Chris Schmitz, Karolina Korgul et al.

Deploying Large Language Model-based agents (LLM agents) in the public sector requires assuring that they meet the stringent legal, procedural, and structural requirements of public-sector institutions. Practitioners and researchers often turn to benchmarks for such assessments. However, it remains unclear what criteria benchmarks must meet to ensure they adequately reflect public-sector requirements, or how many existing benchmarks do so. In this paper, we first define such criteria based on a first-principles survey of public administration literature: benchmarks must be \emph{process-based}, \emph{realistic}, \emph{public-sector-specific} and report \emph{metrics} that reflect the unique requirements of the public sector. We analyse more than 1,300 benchmark papers for these criteria using an expert-validated LLM-assisted pipeline. Our results show that no single benchmark meets all of the criteria. Our findings provide a call to action for both researchers to develop public sector-relevant benchmarks and for public-sector officials to apply these criteria when evaluating their own agentic use cases.

Zihao Fu, Ryan Brown, Shun Shao et al.

Text-to-image diffusion models, such as Stable Diffusion, have demonstrated remarkable capabilities in generating high-quality and diverse images from natural language prompts. However, recent studies reveal that these models often replicate and amplify societal biases, particularly along demographic attributes like gender and race. In this paper, we introduce FairImagen (https://github.com/fuzihaofzh/FairImagen), a post-hoc debiasing framework that operates on prompt embeddings to mitigate such biases without retraining or modifying the underlying diffusion model. Our method integrates Fair Principal Component Analysis to project CLIP-based input embeddings into a subspace that minimizes group-specific information while preserving semantic content. We further enhance debiasing effectiveness through empirical noise injection and propose a unified cross-demographic projection method that enables simultaneous debiasing across multiple demographic attributes. Extensive experiments across gender, race, and intersectional settings demonstrate that FairImagen significantly improves fairness with a moderate trade-off in image quality and prompt fidelity. Our framework outperforms existing post-hoc methods and offers a simple, scalable, and model-agnostic solution for equitable text-to-image generation.

Eoin Delaney, Zihao Fu, Sandra Wachter et al.

We present OxonFair, a new open source toolkit for enforcing fairness in binary classification. Compared to existing toolkits: (i) We support NLP and Computer Vision classification as well as standard tabular problems. (ii) We support enforcing fairness on validation data, making us robust to a wide range of overfitting challenges. (iii) Our approach can optimize any measure based on True Positives, False Positive, False Negatives, and True Negatives. This makes it easily extensible and much more expressive than existing toolkits. It supports all 9 and all 10 of the decision-based group metrics of two popular review articles. (iv) We jointly optimize a performance objective alongside fairness constraints. This minimizes degradation while enforcing fairness, and even improves the performance of inadequately tuned unfair baselines. OxonFair is compatible with standard ML toolkits, including sklearn, Autogluon, and PyTorch and is available at https://github.com/oxfordinternetinstitute/oxonfair

Guillaume Rochette, Chris Russell, Richard Bowden

We present a new approach for synthesizing novel views of people in new poses. Our novel differentiable renderer enables the synthesis of highly realistic images from any viewpoint. Rather than operating over mesh-based structures, our renderer makes use of diffuse Gaussian primitives that directly represent the underlying skeletal structure of a human. Rendering these primitives gives results in a high-dimensional latent image, which is then transformed into an RGB image by a decoder network. The formulation gives rise to a fully differentiable framework that can be trained end-to-end. We demonstrate the effectiveness of our approach to image reconstruction on both the Human3.6M and Panoptic Studio datasets. We show how our approach can be used for motion transfer between individuals; novel view synthesis of individuals captured from just a single camera; to synthesize individuals from any virtual viewpoint; and to re-render people in novel poses. Code and video results are available at https://github.com/GuillaumeRochette/HumanViewSynthesis.

Avishkar Saha, Oscar Mendez Maldonado, Chris Russell et al.

We approach instantaneous mapping, converting images to a top-down view of the world, as a translation problem. We show how a novel form of transformer network can be used to map from images and video directly to an overhead map or bird's-eye-view (BEV) of the world, in a single end-to-end network. We assume a 1-1 correspondence between a vertical scanline in the image, and rays passing through the camera location in an overhead map. This lets us formulate map generation from an image as a set of sequence-to-sequence translations. Posing the problem as translation allows the network to use the context of the image when interpreting the role of each pixel. This constrained formulation, based upon a strong physical grounding of the problem, leads to a restricted transformer network that is convolutional in the horizontal direction only. The structure allows us to make efficient use of data when training, and obtains state-of-the-art results for instantaneous mapping of three large-scale datasets, including a 15% and 30% relative gain against existing best performing methods on the nuScenes and Argoverse datasets, respectively. We make our code available on https://github.com/avishkarsaha/translating-images-into-maps.

Kaivalya Rawal, Eoin Delaney, Zihao Fu et al.

Explainable artificial intelligence (XAI) is concerned with producing explanations indicating the inner workings of models. For a Rashomon set of similarly performing models, explanations provide a way of disambiguating the behavior of individual models, helping select models for deployment. However explanations themselves can vary depending on the explainer used, and need to be evaluated. In the paper "Evaluating Model Explanations without Ground Truth", we proposed three principles of explanation evaluation and a new method "AXE" to evaluate the quality of feature-importance explanations. We go on to illustrate how evaluation metrics that rely on comparing model explanations against ideal ground truth explanations obscure behavioral differences within a Rashomon set. Explanation evaluation aligned with our proposed principles would highlight these differences instead, helping select models from the Rashomon set. The selection of alternate models from the Rashomon set can maintain identical predictions but mislead explainers into generating false explanations, and mislead evaluation methods into considering the false explanations to be of high quality. AXE, our proposed explanation evaluation method, can detect this adversarial fairwashing of explanations with a 100% success rate. Unlike prior explanation evaluation strategies such as those based on model sensitivity or ground truth comparison, AXE can determine when protected attributes are used to make predictions.

Jaime Spencer, Fabio Tosi, Matteo Poggi et al.

This paper discusses the results of the third edition of the Monocular Depth Estimation Challenge (MDEC). The challenge focuses on zero-shot generalization to the challenging SYNS-Patches dataset, featuring complex scenes in natural and indoor settings. As with the previous edition, methods can use any form of supervision, i.e. supervised or self-supervised. The challenge received a total of 19 submissions outperforming the baseline on the test set: 10 among them submitted a report describing their approach, highlighting a diffused use of foundational models such as Depth Anything at the core of their method. The challenge winners drastically improved 3D F-Score performance, from 17.51% to 23.72%.

Will Hawkins, Brent Mittelstadt, Chris Russell

Fine-tuning language models has become increasingly popular following the proliferation of open models and improvements in cost-effective parameter efficient fine-tuning. However, fine-tuning can influence model properties such as safety. We assess how fine-tuning can impact different open models' propensity to output toxic content. We assess the impacts of fine-tuning Gemma, Llama, and Phi models on toxicity through three experiments. We compare how toxicity is reduced by model developers during instruction-tuning. We show that small amounts of parameter-efficient fine-tuning on developer-tuned models via low-rank adaptation on a non-adversarial dataset can significantly alter these results across models. Finally, we highlight the impact of this in the wild, demonstrating how toxicity rates of models fine-tuned by community contributors can deviate in hard-to-predict ways.

Will Hawkins, Chris Russell, Brent Mittelstadt

Advances in multimodal machine learning have made text-to-image (T2I) models increasingly accessible and popular. However, T2I models introduce risks such as the generation of non-consensual depictions of identifiable individuals, otherwise known as deepfakes. This paper presents an empirical study exploring the accessibility of deepfake model variants online. Through a metadata analysis of thousands of publicly downloadable model variants on two popular repositories, Hugging Face and Civitai, we demonstrate a huge rise in easily accessible deepfake models. Almost 35,000 examples of publicly downloadable deepfake model variants are identified, primarily hosted on Civitai. These deepfake models have been downloaded almost 15 million times since November 2022, with the models targeting a range of individuals from global celebrities to Instagram users with under 10,000 followers. Both Stable Diffusion and Flux models are used for the creation of deepfake models, with 96% of these targeting women and many signalling intent to generate non-consensual intimate imagery (NCII). Deepfake model variants are often created via the parameter-efficient fine-tuning technique known as low rank adaptation (LoRA), requiring as few as 20 images, 24GB VRAM, and 15 minutes of time, making this process widely accessible via consumer-grade computers. Despite these models violating the Terms of Service of hosting platforms, and regulation seeking to prevent dissemination, these results emphasise the pressing need for greater action to be taken against the creation of deepfakes and NCII.

Anton Obukhov, Matteo Poggi, Fabio Tosi et al.

This paper presents the results of the fourth edition of the Monocular Depth Estimation Challenge (MDEC), which focuses on zero-shot generalization to the SYNS-Patches benchmark, a dataset featuring challenging environments in both natural and indoor settings. In this edition, we revised the evaluation protocol to use least-squares alignment with two degrees of freedom to support disparity and affine-invariant predictions. We also revised the baselines and included popular off-the-shelf methods: Depth Anything v2 and Marigold. The challenge received a total of 24 submissions that outperformed the baselines on the test set; 10 of these included a report describing their approach, with most leading methods relying on affine-invariant predictions. The challenge winners improved the 3D F-Score over the previous edition's best result, raising it from 22.58% to 23.05%.

Ryan Brown, Chris Russell

Knowledge distillation from large language models (LLMs) assumes that the teacher's output distribution is a high-quality training signal. On reasoning tasks, this assumption is frequently violated. A model's intermediate representations may encode the correct answer, yet this information is lost or distorted through the vocabulary projection, where prompt formatting and answer-token choices creates brittle, noisy outputs. We introduce \method{}, a distillation framework that bypasses this bottleneck by training lightweight probes on frozen teacher hidden states and using the probe's predictions, rather than output logits, as supervision for student training. This simple change yields consistent improvements across four reasoning benchmarks (AQuA-RAT, ARC Easy/Challenge, and MMLU), with gains most pronounced under limited data. Probes trained on intermediate representations provide cleaner labels than the teacher's own outputs, effectively denoising the distillation signal. \method{} requires no architectural changes to student or teacher, is architecture-agnostic, and adds minimal compute since probe training is cheap and teacher representations can be cached. By exploiting internal representations, \method{} enables practitioners to extract more value from large teacher models without additional training data or architectural complexity.

William Lugoloobi, Chris Russell

Large language models exhibit a puzzling inconsistency: they solve complex problems yet frequently fail on seemingly simpler ones. We investigate whether LLMs internally encode problem difficulty in a way that aligns with human judgment, and whether this representation tracks generalization during reinforcement learning post-training. We train linear probes across layers and token positions on 60 models, evaluating on mathematical and coding subsets of Easy2HardBench. We find that human-labeled difficulty is strongly linearly decodable (AMC: $ρ\approx 0.88$) and exhibits clear model-size scaling, whereas LLM-derived difficulty is substantially weaker and scales poorly. Steering along the difficulty direction reveals that pushing models toward "easier" representations reduces hallucination and improves accuracy. During GRPO training on Qwen2.5-Math-1.5B, the human-difficulty probe strengthens and positively correlates with test accuracy across training steps, while the LLM-difficulty probe degrades and negatively correlates with performance. These results suggest that human annotations provide a stable difficulty signal that RL amplifies, while automated difficulty estimates derived from model performance become misaligned precisely as models improve. We release probe code and evaluation scripts to facilitate replication.

Zihao Fu, Ming Liao, Chris Russell et al.

Large language models have achieved remarkable success but remain largely black boxes with poorly understood internal mechanisms. To address this limitation, many researchers have proposed various interpretability methods including mechanistic analysis, probing classifiers, and activation visualization, each providing valuable insights from different perspectives. Building upon this rich landscape of complementary approaches, we introduce CAST (Compositional Analysis via Spectral Tracking), a probe-free framework that contributes a novel perspective by analyzing transformer layer functions through direct transformation matrix estimation and comprehensive spectral analysis. CAST offers complementary insights to existing methods by estimating the realized transformation matrices for each layer using Moore-Penrose pseudoinverse and applying spectral analysis with six interpretable metrics characterizing layer behavior. Our analysis reveals distinct behaviors between encoder-only and decoder-only models, with decoder models exhibiting compression-expansion cycles while encoder models maintain consistent high-rank processing. Kernel analysis further demonstrates functional relationship patterns between layers, with CKA similarity matrices clearly partitioning layers into three phases: feature extraction, compression, and specialization.

João Matos, Ben Van Calster, Leo Anthony Celi et al.

Predictive artificial intelligence (AI) offers an opportunity to improve clinical practice and patient outcomes, but risks perpetuating biases if fairness is inadequately addressed. However, the definition of "fairness" remains unclear. We conducted a scoping review to identify and critically appraise fairness metrics for clinical predictive AI. We defined a "fairness metric" as a measure quantifying whether a model discriminates (societally) against individuals or groups defined by sensitive attributes. We searched five databases (2014-2024), screening 820 records, to include 41 studies, and extracted 62 fairness metrics. Metrics were classified by performance-dependency, model output level, and base performance metric, revealing a fragmented landscape with limited clinical validation and overreliance on threshold-dependent measures. Eighteen metrics were explicitly developed for healthcare, including only one clinical utility metric. Our findings highlight conceptual challenges in defining and quantifying fairness and identify gaps in uncertainty quantification, intersectionality, and real-world applicability. Future work should prioritise clinically meaningful metrics.

Zihao Fu, Chris Russell

Digital watermarking is a promising solution for mitigating some of the risks arising from the misuse of automatically generated text. These approaches either embed non-specific watermarks to allow for the detection of any text generated by a particular sampler, or embed specific keys that allow the identification of the LLM user. However, simultaneously using the same embedding for both detection and user identification leads to a false detection problem, whereby, as user capacity grows, unwatermarked text is increasingly likely to be falsely detected as watermarked. Through theoretical analysis, we identify the underlying causes of this phenomenon. Building on these insights, we propose Dual Watermarking which jointly encodes detection and identification watermarks into generated text, significantly reducing false positives while maintaining high detection accuracy. Our experimental results validate our theoretical findings and demonstrate the effectiveness of our approach.

Sofie Goethals, Eoin Delaney, Brent Mittelstadt et al.

Access to resources strongly constrains the decisions we make. While we might wish to offer every student a scholarship, or schedule every patient for follow-up meetings with a specialist, limited resources mean that this is not possible. When deploying machine learning systems, these resource constraints are simply enforced by varying the threshold of a classifier. However, these finite resource limitations are disregarded by most existing tools for fair machine learning, which do not allow the specification of resource limitations and do not remain fair when varying thresholds. This makes them ill-suited for real-world deployment. Our research introduces the concept of "resource-constrained fairness" and quantifies the cost of fairness within this framework. We demonstrate that the level of available resources significantly influences this cost, a factor overlooked in previous evaluations.

Lukas Schott, Julius von Kügelgen, Frederik Träuble et al.

An important component for generalization in machine learning is to uncover underlying latent factors of variation as well as the mechanism through which each factor acts in the world. In this paper, we test whether 17 unsupervised, weakly supervised, and fully supervised representation learning approaches correctly infer the generative factors of variation in simple datasets (dSprites, Shapes3D, MPI3D) from controlled environments, and on our contributed CelebGlow dataset. In contrast to prior robustness work that introduces novel factors of variation during test time, such as blur or other (un)structured noise, we here recompose, interpolate, or extrapolate only existing factors of variation from the training data set (e.g., small and medium-sized objects during training and large objects during testing). Models that learn the correct mechanism should be able to generalize to this benchmark. In total, we train and test 2000+ models and observe that all of them struggle to learn the underlying mechanism regardless of supervision signal and architectural bias. Moreover, the generalization capabilities of all tested models drop significantly as we move from artificial datasets towards more realistic real-world datasets. Despite their inability to identify the correct mechanism, the models are quite modular as their ability to infer other in-distribution factors remains fairly stable, providing only a single factor is out-of-distribution. These results point to an important yet understudied problem of learning mechanistic models of observations that can facilitate generalization.

Matthäus Kleindessner, Samira Samadi, Muhammad Bilal Zafar et al.

We initiate the study of fairness for ordinal regression. We adapt two fairness notions previously considered in fair ranking and propose a strategy for training a predictor that is approximately fair according to either notion. Our predictor has the form of a threshold model, composed of a scoring function and a set of thresholds, and our strategy is based on a reduction to fair binary classification for learning the scoring function and local search for choosing the thresholds. We provide generalization guarantees on the error and fairness violation of our predictor, and we illustrate the effectiveness of our approach in extensive experiments.

Herman Yau, Chris Russell, Simon Hadfield

We present a novel form of explanation for Reinforcement Learning, based around the notion of intended outcome. These explanations describe the outcome an agent is trying to achieve by its actions. We provide a simple proof that general methods for post-hoc explanations of this nature are impossible in traditional reinforcement learning. Rather, the information needed for the explanations must be collected in conjunction with training the agent. We derive approaches designed to extract local explanations based on intention for several variants of Q-function approximation and prove consistency between the explanations and the Q-values learned. We demonstrate our method on multiple reinforcement learning problems, and provide code to help researchers introspecting their RL environments and algorithms.

Jacob Abernethy, Pranjal Awasthi, Matthäus Kleindessner et al.

We propose simple active sampling and reweighting strategies for optimizing min-max fairness that can be applied to any classification or regression model learned via loss minimization. The key intuition behind our approach is to use at each timestep a datapoint from the group that is worst off under the current model for updating the model. The ease of implementation and the generality of our robust formulation make it an attractive option for improving model performance on disadvantaged groups. For convex learning problems, such as linear or logistic regression, we provide a fine-grained analysis, proving the rate of convergence to a min-max fair solution.

Sandra Wachter, Brent Mittelstadt, Chris Russell

This article identifies a critical incompatibility between European notions of discrimination and existing statistical measures of fairness. First, we review the evidential requirements to bring a claim under EU non-discrimination law. Due to the disparate nature of algorithmic and human discrimination, the EU's current requirements are too contextual, reliant on intuition, and open to judicial interpretation to be automated. Second, we show how the legal protection offered by non-discrimination law is challenged when AI, not humans, discriminate. Humans discriminate due to negative attitudes (e.g. stereotypes, prejudice) and unintentional biases (e.g. organisational practices or internalised stereotypes) which can act as a signal to victims that discrimination has occurred. Finally, we examine how existing work on fairness in machine learning lines up with procedures for assessing cases under EU non-discrimination law. We propose "conditional demographic disparity" (CDD) as a standard baseline statistical measurement that aligns with the European Court of Justice's "gold standard." Establishing a standard set of statistical evidence for automated discrimination cases can help ensure consistent procedures for assessment, but not judicial interpretation, of cases involving AI and automated systems. Through this proposal for procedural regularity in the identification and assessment of automated discrimination, we clarify how to build considerations of fairness into automated systems as far as possible while still respecting and enabling the contextual approach to judicial interpretation practiced under EU non-discrimination law. N.B. Abridged abstract

Nicholas Asher, Soumya Paul, Chris Russell

Explaining sophisticated machine-learning based systems is an important issue at the foundations of AI. Recent efforts have shown various methods for providing explanations. These approaches can be broadly divided into two schools: those that provide a local and human interpreatable approximation of a machine learning algorithm, and logical approaches that exactly characterise one aspect of the decision. In this paper we focus upon the second school of exact explanations with a rigorous logical foundation. There is an epistemological problem with these exact methods. While they can furnish complete explanations, such explanations may be too complex for humans to understand or even to write down in human readable form. Interpretability requires epistemically accessible explanations, explanations humans can grasp. Yet what is a sufficiently complete epistemically accessible explanation still needs clarification. We do this here in terms of counterfactuals, following [Wachter et al., 2017]. With counterfactual explanations, many of the assumptions needed to provide a complete explanation are left implicit. To do so, counterfactual explanations exploit the properties of a particular data point or sample, and as such are also local as well as partial explanations. We explore how to move from local partial explanations to what we call complete local explanations and then to global ones. But to preserve accessibility we argue for the need for partiality. This partiality makes it possible to hide explicit biases present in the algorithm that may be injurious or unfair.We investigate how easy it is to uncover these biases in providing complete and fair explanations by exploiting the structure of the set of counterfactuals providing a complete local explanation.

Andrew Elliott, Stephen Law, Chris Russell

We present a simple regularization of adversarial perturbations based upon the perceptual loss. While the resulting perturbations remain imperceptible to the human eye, they differ from existing adversarial perturbations in that they are semi-sparse alterations that highlight objects and regions of interest while leaving the background unaltered. As a semantically meaningful adverse perturbations, it forms a bridge between counterfactual explanations and adversarial perturbations in the space of images. We evaluate our approach on several standard explainability benchmarks, namely, weak localization, insertion deletion, and the pointing game demonstrating that perceptually regularized counterfactuals are an effective explanation for image-based classifiers.