ICLR 2025 Papers

Advancements in neural networks have significantly enhanced the performance of classification models, achieving remarkable accuracy across diverse datasets. However, these models often lack transparency and do not support interactive reasoning with human users, which are essential attributes for applications that require trust and user engagement. To overcome these limitations, we introduce an innovative framework, Neural Causal Graph (NCG), that integrates causal inference with neural networks to enable interpretable and intervenable reasoning. We then propose an intervention training method to model the intervention probability of the prediction, serving as a contextual prompt to facilitate the fine-grained reasoning and human-AI interaction abilities of NCG. Our experiments show that the proposed framework significantly enhances the performance of traditional classification baselines. Furthermore, NCG achieves nearly 95\% top-1 accuracy on the ImageNet dataset by employing a test-time intervention method. This framework not only supports sophisticated post-hoc interpretation but also enables dynamic human-AI interactions, significantly improving the model's transparency and applicability in real-world scenarios.

In time-series classification, interpretable models can bring additional insights but be outperformed by deep models since human-understandable features have limited expressivity and flexibility. In this work, we present InterpGN, a framework that integrates an interpretable model and a deep neural network. Within this framework, we introduce a novel gating function design based on the confidence of the interpretable expert, preserving interpretability for samples where interpretable features are significant while also identifying samples that require additional expertise. For the interpretable expert, we incorporate shapelets to effectively model shape-level features for time-series data. We introduce a variant of Shapelet Transforms to build logical predicates using shapelets. Our proposed model achieves comparable performance with state-of-the-art deep learning models while additionally providing interpretable classifiers for various benchmark datasets. We further show that our models improve on quantitative shapelet quality and interpretability metrics over existing shapelet-learning formulations. Finally, we show that our models can integrate additional advanced architectures and be applied to real-world tasks beyond standard benchmarks such as the MIMIC-III and time series extrinsic regression datasets.

Powder X-ray diffraction (XRD) patterns are highly effective for crystal identification and play a pivotal role in materials discovery. While machine learning (ML) has advanced the analysis of powder XRD patterns, progress has been constrained by the limited availability of training data and established benchmarks. To address this, we introduce SimXRD, the largest open-source simulated XRD pattern dataset to date, aimed at accelerating the development of crystallographic informatics. We developed a novel XRD simulation method that incorporates comprehensive physical interactions, resulting in a high-fidelity database. SimXRD comprises 4,065,346 simulated powder XRD patterns, representing 119,569 unique crystal structures under 33 simulated conditions that reflect real-world variations. We benchmark 21 sequence models in both in-library and out-of-library scenarios and analyze the impact of class imbalance in long-tailed crystal label distributions. Remarkably, we find that: (1) current neural networks struggle with classifying low-frequency crystals, particularly in out-of-library situations; (2) models trained on SimXRD can generalize to real experimental data.

Pretrained masked language models (MLMs) have demonstrated an impressive capability to comprehend and encode conceptual knowledge, revealing a lattice structure among concepts. This raises a critical question: how does this conceptualization emerge from MLM pretraining? In this paper, we explore this problem from the perspective of Formal Concept Analysis (FCA), a mathematical framework that derives concept lattices from the observations of object-attribute relationships. We show that the MLM's objective implicitly learns a formal context that describes objects, attributes, and their dependencies, which enables the reconstruction of a concept lattice through FCA. We propose a novel framework for concept lattice construction from pretrained MLMs and investigate the origin of the inductive biases of MLMs in lattice structure learning. Our framework differs from previous work because it does not rely on human-defined concepts and allows for discovering "latent" concepts that extend beyond human definitions. We create three datasets for evaluation, and the empirical results verify our hypothesis.

Causal opacity denotes the difficulty in understanding the "hidden" causal structure underlying the decisions of deep neural network (DNN) models. This leads to the inability to rely on and verify state-of-the-art DNN-based systems, especially in high-stakes scenarios. For this reason, circumventing causal opacity in DNNs represents a key open challenge at the intersection of deep learning, interpretability, and causality. This work addresses this gap by introducing Causal Concept Graph Models (Causal CGMs), a class of interpretable models whose decision-making process is causally transparent by design. Our experiments show that Causal CGMs can: (i) match the generalisation performance of causally opaque models, (ii) enable human-in-the-loop corrections to mispredicted intermediate reasoning steps, boosting not just downstream accuracy after corrections but also the reliability of the explanations provided for specific instances, and (iii) support the analysis of interventional and counterfactual scenarios, thereby improving the model's causal interpretability and supporting the effective verification of its reliability and fairness.

Recent alignment algorithms such as direct preference optimization (DPO) have been developed to improve the safety of large language models (LLMs) by training these models to match human behaviors exemplified by preference data. However, these methods are both computationally intensive and lacking in controllability and transparency, inhibiting their widespread use. Furthermore, these tuning-based methods require large-scale preference data for training and are susceptible to noisy preference data. In this paper, we introduce a tuning-free alignment alternative, ProFS (Projection Filter for Subspaces), and demonstrate its effectiveness under the use case of toxicity reduction. Grounded on theory from factor analysis, ProFS is a sample-efficient model editing approach that identifies a toxic subspace in the model parameter space and reduces model toxicity by projecting away the detected subspace. The toxic subspace is identified by extracting preference data embeddings from the language model, and removing non-toxic information from these embeddings. We show that ProFS is more sample-efficient than DPO, further showcasing greater robustness to noisy data. Finally, we attempt to connect tuning based alignment with editing, by establishing both theoretical and empirical connections between ProFS and DPO, showing that ProFS can be interpreted as a denoised version of a single DPO step.

Data valuation has found many real-world applications, e.g., data pricing and data selection. However, the most adopted approach -- Shapley value (SV) -- is computationally expensive due to the large number of model trainings required. Fortunately, most applications (e.g., data selection) require only knowing the $m$ data points with the highest data values (i.e., top-$m$ data values), which implies the potential for fewer model trainings as exact data values are not required. Existing work formulates top-$m$ Shapley value identification as top-$m$ arms identification in multi-armed bandits (MAB). However, the proposed approach falls short because it does not utilize data features to predict data values, a method that has been shown empirically to be effective. A recent top-$m$ arms identification work does consider the use of arm features while assuming a linear relationship between arm features and rewards, which is often not satisfied in data valuation. To this end, we propose the GPGapE algorithm that uses the Gaussian process to model the \emph{non-linear} mapping from data features to data values, removing the linear assumption. We theoretically analyze the correctness and stopping iteration of GPGapE in finding an $(\epsilon, \delta)$-approximation to the top-$m$ data values. We further improve the computational efficiency, by calculating data values using small data subsets to reduce the computation cost of model training. We empirically demonstrate that GPGapE outperforms other baselines in top-$m$ data values identification, noisy data detection, and data subset selection on real-world datasets. We also demonstrate the efficiency of our GPGapE in data selection for large language model fine-tuning.

Vision and language model (VLM) decoders are currently the best-performing architectures on multimodal tasks. Next to answers, they are able to produce natural language explanations, either in post-hoc or CoT settings. However, it is not clear to what extent they are using the input vision and text modalities when generating answers or explanations. In this work, we investigate if VLMs rely on their input modalities differently when they produce explanations as opposed to answers. We also evaluate the self-consistency of VLM decoders in both post-hoc and CoT explanation settings, by extending existing unimodal tests and measures to VLM decoders. We find that most tested VLMs are less self-consistent than LLMs. Text contributions in all tested VL decoders are more important than image contributions in all examined tasks. However, when comparing explanation generation to answer generation, the contributions of images are significantly stronger for generating explanations compared to answers. This difference is even larger in CoT compared to post-hoc explanations. Lastly, we provide an up-to-date benchmarking of state-of-the-art VL decoders on the VALSE benchmark, which before was restricted to VL encoders. We find that the tested VL decoders still struggle with most phenomena tested by VALSE. We will make our code publicly available.

Despite the significant success of deep learning models in computer vision, they often exhibit systematic failures on specific data subsets, known as error slices. Identifying and mitigating these error slices is crucial to enhancing model robustness and reliability in real-world scenarios. In this paper, we introduce HiBug2, an automated framework for error slice discovery and model repair. HiBug2 first generates task-specific visual attributes to highlight instances prone to errors through an interpretable and structured process. It then employs an efficient slice enumeration algorithm to systematically identify error slices, overcoming the combinatorial challenges that arise during slice exploration. Additionally, HiBug2 extends its capabilities by predicting error slices beyond the validation set, addressing a key limitation of prior approaches. Extensive experiments across multiple domains — including image classification, pose estimation, and object detection — show that HiBug2 not only improves the coherence and precision of identified error slices but also significantly enhances the model repair capabilities.

The increasing capabilities of large generative models and their ever more widespread deployment have raised concerns about their reliability, safety, and potential misuse. To address these issues, recent works have proposed to control model generation by steering model activations in order to effectively induce or prevent the emergence of concepts or behaviors in the generated output. In this paper we introduce Activation Transport (AcT), a general framework to steer activations guided by optimal transport theory that generalizes many previous activation-steering works. AcT is modality-agnostic and provides fine-grained control over the model behavior with negligible computational overhead, while minimally impacting model abilities. We experimentally show the effectiveness and versatility of our approach by addressing key challenges in large language models (LLMs) and text-to-image diffusion models (T2Is). For LLMs, we show that AcT can effectively mitigate toxicity, induce arbitrary concepts, and increase their truthfulness. In T2Is, we show how AcT enables fine-grained style control and concept negation.

In recent years, the Shapley value and SHAP explanations have emerged as one of the most dominant paradigms for providing post-hoc explanations of blackbox models. Despite their well-founded theoretical properties, many recent works have focused on the limitations in both their computational efficiency and their representation power. The underlying connection with additive models, however, is left critically under-emphasized in the current literature. In this work, we find that a variational perspective linking GAM models and SHAP explanations is able to provide deep insights into nearly all recent developments. In light of this connection, we borrow in the other direction to develop a new method to train interpretable GAM models which are automatically purified to compute the Shapley value in a single forward pass. Finally, we provide theoretical results showing the limited representation power of GAM models is the same Achilles’ heel existing in SHAP and discuss the implications for SHAP’s modern usage in CV and NLP.

Large Language Models (LLMs) have made significant strides in natural language processing, and a precise understanding of the internal mechanisms driving their success is essential. In this work, we analyze the trajectories of token embeddings as they pass through transformer blocks, linearizing the system along these trajectories through their Jacobian matrices. By examining the relationships between these block Jacobians, we uncover the phenomenon of **transformer block coupling** in a multitude of LLMs, characterized by the coupling of their top singular vectors across tokens and depth. Our findings reveal that coupling *positively correlates* with model performance, and that this relationship is stronger than with other hyperparameters such as parameter count, model depth, and embedding dimension. We further investigate how these properties emerge during training, observing a progressive development of coupling, increased linearity, and layer-wise exponential growth in token trajectories. Additionally, experiments with Vision Transformers (ViTs) corroborate the emergence of coupling and its relationship with generalization, reinforcing our findings in LLMs. Collectively, these insights offer a novel perspective on token interactions in transformers, opening new directions for studying their mechanisms as well as improving training and generalization.

As Graph Neural Networks (GNNs) become more pervasive, it becomes paramount to build reliable tools for explaining their predictions. A core desideratum is that explanations are *faithful*, i.e., that they portray an accurate picture of the GNN's reasoning process. However, a number of different faithfulness metrics exist, begging the question of what is faithfulness exactly and how to achieve it. We make three key contributions. We begin by showing that *existing metrics are not interchangeable* -- i.e., explanations attaining high faithfulness according to one metric may be unfaithful according to others -- and can *systematically ignore important properties of explanations*. We proceed to show that, surprisingly, *optimizing for faithfulness is not always a sensible design goal*. Specifically, we prove that for injective regular GNN architectures, perfectly faithful explanations are completely uninformative. This does not apply to modular GNNs, such as self-explainable and domain-invariant architectures, prompting us to study the relationship between architectural choices and faithfulness. Finally, we show that *faithfulness is tightly linked to out-of-distribution generalization*, in that simply ensuring that a GNN can correctly recognize the domain-invariant subgraph, as prescribed by the literature, does not guarantee that it is invariant unless this subgraph is also faithful. All our code can be found in the supplementary material.

CLIP has achieved great success in visual representation learning and is becoming an important plug-in component for many large multi-modal models like LLaVA and DALL-E. However, the lack of interpretability caused by the intricate image encoder architecture and training process restricts its wider use in high-stake decision making applications. In this work, we propose an unsupervised adversarial fine-tuning (AFT) with norm-regularization to enhance the visual interpretability of CLIP. We provide theoretical analysis showing that AFT has implicit regularization that enforces the image encoder to encode the input features sparsely, directing the network's focus towards meaningful features. Evaluations by both feature attribution techniques and network dissection offer convincing evidence that the visual interpretability of CLIP has significant improvements. With AFT, the image encoder prioritizes pertinent input features, and the neuron within the encoder exhibits better alignment with human-understandable concepts. Moreover, these effects are generalizable to out-of-distribution datasets and can be transferred to downstream tasks. Additionally, AFT enhances the visual interpretability of derived large vision-language models that incorporate the pre-trained CLIP an integral component. The code of this paper is available at [the CLIP_AFT GitHub repository](https://github.com/peterant330/CLIP_AFT).

Data-driven modeling of dynamical systems is a crucial area of machine learning. In many scenarios, a thorough understanding of the model’s behavior becomes essential for practical applications. For instance, understanding the behavior of a pharmacokinetic model, constructed as part of drug development, may allow us to both verify its biological plausibility (e.g., the drug concentration curve is non-negative and decays to zero in the long term) and to design dosing guidelines (e.g., by looking at the peak concentration and its timing). Discovery of closed-form ordinary differential equations (ODEs) can be employed to obtain such insights by finding a compact mathematical equation and then analyzing it (a two-step approach). However, its widespread use is currently hindered because the analysis process may be time-consuming, requiring substantial mathematical expertise, or even impossible if the equation is too complex. Moreover, if the found equation's behavior does not satisfy the requirements, editing it or influencing the discovery algorithms to rectify it is challenging as the link between the symbolic form of an ODE and its behavior can be elusive. This paper proposes a conceptual shift to modeling low-dimensional dynamical systems by departing from the traditional two-step modeling process. Instead of first discovering a closed-form equation and then analyzing it, our approach, direct semantic modeling, predicts the semantic representation of the dynamical system (i.e., description of its behavior) directly from data, bypassing the need for complex post-hoc analysis. This direct approach also allows the incorporation of intuitive inductive biases into the optimization algorithm and editing the model's behavior directly, ensuring that the model meets the desired specifications. Our approach not only simplifies the modeling pipeline but also enhances the transparency and flexibility of the resulting models compared to traditional closed-form ODEs.

This paper considers an important Graph Anomaly Detection (GAD) task, namely open-set GAD, which aims to train a detection model using a small number of normal and anomaly nodes (referred to as *seen anomalies*) to detect both seen anomalies and *unseen anomalies* (*i.e*., anomalies that cannot be illustrated the training anomalies). Those labelled training data provide crucial prior knowledge about abnormalities for GAD models, enabling substantially reduced detection errors. However, current supervised GAD methods tend to over-emphasise fitting the seen anomalies, leading to many errors of detecting the unseen anomalies as normal nodes. Further, existing open-set AD models were introduced to handle Euclidean data, failing to effectively capture discriminative features from graph structure and node attributes for GAD. In this work, we propose a novel open-set GAD approach, namely $\underline{n}ormal$ $\underline{s}tructure$ $\underline{reg}ularisation$ (**NSReg**), to achieve generalised detection ability to unseen anomalies, while maintaining its effectiveness on detecting seen anomalies. The key idea in NSReg is to introduce a regularisation term that enforces the learning of compact, semantically-rich representations of normal nodes based on their structural relations to other nodes. When being optimised with supervised anomaly detection losses, the regularisation term helps incorporate strong normality into the modelling, and thus, it effectively avoids over-fitting the seen anomalies and learns a better normality decision boundary, largely reducing the false negatives of detecting unseen anomalies as normal. Extensive empirical results on seven real-world datasets show that NSReg significantly outperforms state-of-the-art competing methods by at least 14% AUC-ROC on the unseen anomaly classes and by 10% AUC-ROC on all anomaly classes. Code and datasets are available at https://github.com/mala-lab/NSReg.

With models getting stronger, evaluations have grown more complex, testing multiple skills in one benchmark and even in the same instance at once. However, skill-wise performance is obscured when inspecting aggregate accuracy, under-utilizing the rich signal modern benchmarks contain. We propose an automatic approach to recover the underlying skills relevant for any evaluation instance, by way of inspecting model-generated {\em rationales}. After validating the relevance of rationale-parsed skills and inferring skills for $46$k instances over $12$ benchmarks, we observe many skills to be common across benchmarks, resulting in the curation of hundreds of \emph{skill-slices} (i.e. sets of instances testing a common skill). Inspecting accuracy over these slices yields novel insights on model trade-offs: e.g., compared to GPT-4o and Claude 3.5 Sonnet, on average, Gemini 1.5 Pro is $18\%$ more accurate in \emph{computing molar mass}, but $19\\%$ less accurate in \emph{applying constitutional law}, despite the overall accuracies of the three models differing by a mere $0.4\\%$. Furthermore, we demonstrate the practical utility of our approach by showing that insights derived from skill slice analysis can generalize to held-out instances: when routing each instance to the model strongest on the relevant skills, we see a $3\\%$ accuracy improvement over our $12$ dataset corpus. Our skill-slices and framework open a new avenue in model evaluation, leveraging skill-specific analyses to unlock a more granular and actionable understanding of model capabilities.

Sparse autoencoders (SAEs) are a recent technique for decomposing neural network activations into human-interpretable features. However, in order for SAEs to identify all features represented in frontier models, it will be necessary to scale them up to very high width, posing a computational challenge. In this work, we introduce Switch Sparse Autoencoders, a novel SAE architecture aimed at reducing the compute cost of training SAEs. Inspired by sparse mixture of experts models, Switch SAEs route activation vectors between smaller "expert" SAEs, enabling SAEs to efficiently scale to many more features. We present experiments comparing Switch SAEs with other SAE architectures, and find that Switch SAEs deliver a substantial Pareto improvement in the reconstruction vs. sparsity frontier for a given fixed training compute budget. We also study the geometry of features across experts, analyze features duplicated across experts, and verify that Switch SAE features are as interpretable as features found by other SAE architectures.

Transformers have the capacity to act as supervised learning algorithms: by properly encoding a set of labeled training (''in-context'') examples and an unlabeled test example into an input sequence of vectors of the same dimension, the forward pass of the transformer can produce predictions for that unlabeled test example. A line of recent work has shown that when linear transformers are pre-trained on random instances for linear regression tasks, these trained transformers make predictions using an algorithm similar to that of ordinary least squares. In this work, we investigate the behavior of linear transformers trained on random linear classification tasks. Via an analysis of the implicit regularization of gradient descent, we characterize how many pre-training tasks and in-context examples are needed for the trained transformer to generalize well at test-time. We further show that in some settings, these trained transformers can exhibit ''benign overfitting in-context'': when in-context examples are corrupted by label flipping noise, the transformer memorizes all of its in-context examples (including those with noisy labels) yet still generalizes near-optimally for clean test examples.