ICLR 2025 Papers

Vision-Language Models (VLMs) are powerful tools for processing and understanding text and images. We study the processing of visual tokens in the language model component of LLaVA, a prominent VLM. Our approach focuses on analyzing the localization of object information, the evolution of visual token representations across layers, and the mechanism of integrating visual information for predictions. Through ablation studies, we demonstrated that object identification accuracy drops by over 70\% when object-specific tokens are removed. We observed that visual token representations become increasingly interpretable in the vocabulary space across layers, suggesting an alignment with textual tokens corresponding to image content. Finally, we found that the model extracts object information from these refined representations at the last token position for prediction, mirroring the process in text-only language models for factual association tasks. These findings provide crucial insights into how VLMs process and integrate visual information, bridging the gap between our understanding of language and vision models, and paving the way for more interpretable and controllable multimodal systems.

We investigate the approximation and estimation rates of conditional diffusion transformers (DiTs) with classifier-free guidance. We present a comprehensive analysis for “in-context” conditional DiTs under various common assumptions: generic and strong Hölder, linear latent (subspace), and Lipschitz score function assumptions. Importantly, we establish minimax optimality of DiTs by leveraging score function regularity. Specifically, we discretize the input domains into infinitesimal grids and then perform term-by-term Taylor expansions on the conditional diffusion score function under the Hölder smooth data assumption. This enables fine-grained use of transformers’ universal approximation through a more detailed piecewise constant approximation, and hence obtains tighter bounds. Additionally, we extend our analysis to latent settings. Our findings establish statistical limits for DiTs and offer practical guidance toward more efficient and accurate designs.

Transformers have exhibited exceptional capabilities in sequence modelling tasks, leveraging self-attention and in-context learning. Critical to this success are induction heads, attention circuits that enable copying tokens based on their previous occurrences. In this work, we introduce a novel synthetic framework designed to enable the theoretical analysis of transformers’ ability to dynamically handle causal structures. Existing works rely on Markov Chains to study the formation of induction heads, revealing how transformers capture causal dependencies and learn transition probabilities in-context. However, they rely on a fixed causal structure that fails to capture the complexity of natural languages, where the relationship between tokens dynamically changes with context. To this end, our framework varies the causal structure through interleaved Markov chains with different lags while keeping the transition probabilities fixed. This setting unveils the formation of Selective Induction Heads, a new circuit that endows transformers with the ability to select the correct causal structure in-context. We empirically demonstrate that attention-only transformers learn this mechanism to predict the next token by identifying the correct lag and copying the corresponding token from the past. We provide a detailed construction of a 3-layer transformer to implement the selective induction head, and a theoretical analysis proving that this mechanism asymptotically converges to the maximum likelihood solution. Our findings advance the theoretical understanding of how transformers select causal structures, providing new insights into their functioning and interpretability.

The stability of language model pre-training and its effects on downstream performance are still understudied. Prior work shows that the training process can yield significantly different results in response to slight variations in initial conditions, e.g., the random seed. Crucially, the research community still lacks sufficient resources and tools to systematically investigate pre-training stability, particularly for decoder-only language models. We introduce the PolyPythias, a set of 45 new training runs for the Pythia model suite: 9 new seeds across 5 model sizes, from 14M to 410M parameters, resulting in about 7k new checkpoints that we release. Using these new 45 training runs, in addition to the 5 already available, we study the effects of different initial conditions determined by the seed---i.e., parameters' initialisation and data order---on (i) downstream performance, (ii) learned linguistic representations, and (iii) emergence of training phases. In addition to common scaling behaviours, our analyses generally reveal highly consistent training dynamics across both model sizes and initial conditions. Further, the new seeds for each model allow us to identify outlier training runs and delineate their characteristics. Our findings show the potential of using these methods to predict training stability.

Neuronal dynamics are highly nonlinear and nonstationary. Traditional methods for extracting the underlying network structure from neuronal activity recordings mainly concentrate on modeling static connectivity, without accounting for key nonstationary aspects of biological neural systems, such as ongoing synaptic plasticity and neuronal modulation. To bridge this gap, we introduce the NetFormer model, an interpretable approach applicable to such systems. In NetFormer, the activity of each neuron across a series of historical time steps is defined as a token. These tokens are then linearly mapped through a query and key mechanism to generate a state- (and hence time-) dependent attention matrix that directly encodes nonstationary connectivity structures. We analyze our formulation from the perspective of nonstationary and nonlinear networked dynamical systems, and show both via an analytical expansion and targeted simulations how it can approximate the underlying ground truth. Next, we demonstrate NetFormer's ability to model a key feature of biological networks, spike-timing-dependent plasticity, whereby connection strengths continually change in response to local activity patterns. We further demonstrate that NetFormer can capture task-induced connectivity patterns on activity generated by task-trained recurrent neural networks. Thus informed, we apply NetFormer to a multi-modal dataset of real neural recordings, which contains neural activity, cell type, and behavioral state information. We show that the NetFormer effectively predicts neural dynamics and identifies cell-type specific, state-dependent dynamic connectivity that matches patterns measured in separate ground-truth physiology experiments, demonstrating its ability to help decode complex neural interactions based on population activity observations alone.

The linear representation hypothesis is the informal idea that semantic concepts are encoded as linear directions in the representation spaces of large language models (LLMs). Previous work has shown how to make this notion precise for representing binary concepts that have natural contrasts (e.g., {male, female}) as _directions_ in representation space. However, many natural concepts do not have natural contrasts (e.g., whether the output is about an animal). In this work, we show how to extend the formalization of the linear representation hypothesis to represent features (e.g., is_animal) as _vectors_. This allows us to immediately formalize the representation of categorical concepts as polytopes in the representation space. Further, we use the formalization to prove a relationship between the hierarchical structure of concepts and the geometry of their representations. We validate these theoretical results on the Gemma and LLaMA-3 large language models, estimating representations for 900+ hierarchically related concepts using data from WordNet.

We study causal interactions for payoff allocation in cooperative game theory, including quantifying feature attribution for deep learning models. Most feature attribution methods mainly stem from the criteria of the Shapley value, which assigns fair payoffs to players based on their expected contribution in a cooperative game. However, interactions between players in the game do not explicitly appear in the original formulation of the Shapley value. In this work, we reformulate the Shapley value to clarify the role of interactions and discuss implicit assumptions from a game-theoretical perspective. Our theoretical analysis demonstrates that when negative interactions exist—common in deep learning models—the efficiency axiom can lead to the undervaluation of attributions or payoffs. We suggest a new allocation rule that decomposes contributions into interactions and aggregates positive parts for non-convex games. Furthermore, we propose an approximation algorithm to reduce the cost of interaction computation which can be applied to differentiable functions such as deep learning models. Our approach mitigates counterintuitive attribution outcomes observed in existing methods, ensuring that features critical to a model’s decision receive appropriate attribution.

Deep learning has achieved advancements across a variety of forefront fields. However, its inherent 'black box' characteristic poses challenges to the comprehension and trustworthiness of the decision-making processes within neural networks. To mitigate these challenges, we introduce InnerSightNet, an inner information analysis algorithm designed to illuminate the inner workings of deep neural networks through the perspectives of community. This approach is aimed at deciphering the intricate patterns of neurons within deep neural networks, thereby shedding light on the networks' information processing and decision-making pathways. InnerSightNet operates in three primary phases, 'neuronization-aggregation-evaluation'. Initially, it transforms learnable units into a structured network of neurons. Subsequently, these neurons are aggregated into distinct communities according to representation attributes. The final phase involves the evaluation of these communities' roles and functionalities, to unpick the information flow and decision-making. By transcending focus on single-layer or individual neuron, InnerSightNet broadens the horizon for deep neural network interpretation. InnerSightNet offers a unique vantage point, enabling insights into the collective behavior of communities within the overarching architecture, thereby enhancing transparency and trust in deep learning systems.

As the modern tools of choice for text understanding and generation, large language models (LLMs) are expected to accurately output answers by leveraging the input context. This requires LLMs to possess both context-faithfulness and factual accuracy. While extensive efforts aim to reduce hallucinations through factuality enhancement methods, they also pose risks of hindering context-faithfulness, as factuality enhancement can lead LLMs to become overly confident in their parametric knowledge, causing them to overlook the relevant input context. In this work, we argue that current factuality enhancement methods can significantly undermine the context-faithfulness of LLMs. We first revisit the current factuality enhancement methods and evaluate their effectiveness in enhancing factual accuracy. Next, we evaluate their performance on knowledge editing tasks to assess the potential impact on context-faithfulness. The experimental results reveal that while these methods may yield inconsistent improvements in factual accuracy, they also cause a more severe decline in context-faithfulness, with the largest decrease reaching a striking 69.7\%. To explain these declines, we analyze the hidden states and logit distributions for the tokens representing new knowledge and parametric knowledge respectively, highlighting the limitations of current approaches. Our finding highlights the complex trade-offs inherent in enhancing LLMs. Therefore, we recommend that more research on LLMs' factuality enhancement make efforts to reduce the sacrifice of context-faithfulness.

We introduce ORC-ManL, a new algorithm to prune spurious edges from nearest neighbor graphs using a criterion based on Ollivier-Ricci curvature and estimated metric distortion. Our motivation comes from manifold learning: we show that when the data generating the nearest-neighbor graph consists of noisy samples from a low-dimensional manifold, edges that shortcut through the ambient space have more negative Ollivier-Ricci curvature than edges that lie along the data manifold. We demonstrate that our method outperforms alternative pruning methods and that it significantly improves performance on many downstream geometric data analysis tasks that use nearest neighbor graphs as input. Specifically, we evaluate on manifold learning, persistent homology, dimension estimation, and others. We also show that ORC-ManL can be used to improve clustering and manifold learning of single-cell RNA sequencing data. Finally, we provide empirical convergence experiments that support our theoretical findings.

We explore a class of adversarial attacks targeting the activations of language models to derive upper-bound scaling laws on their attack susceptibility. By manipulating a relatively small subset of model activations, $a$, we demonstrate the ability to control the exact prediction of a significant number (in some cases up to 1000) of subsequent tokens $t$. We empirically verify a scaling law where the maximum number of target tokens predicted, $t_\mathrm{max}$, depends linearly on the number of tokens $a$ whose activations the attacker controls as $t_\mathrm{max} = \kappa a$. We find that the number of bits the attacker controls on the input to exert a single bit of control on the output (a property we call \textit{attack resistance $\chi$}) is remarkably stable between $\approx 16$ and $\approx 25$ over orders of magnitude of model sizes and between model families. Compared to attacks directly on input tokens, attacks on activations are predictably much stronger, however, we identify a surprising regularity where one bit of input steered either via activations or via tokens is able to exert a surprisingly similar amount of control over the model predictions. This gives support for the hypothesis that adversarial attacks are a consequence of dimensionality mismatch between the input and output spaces. A practical implication of the ease of attacking language model activations instead of tokens is for multi-modal and selected retrieval models. By using language models as a controllable test-bed to study adversarial attacks, we explored input-output dimension regimes that are inaccessible in computer vision and greatly extended the empirical support for the dimensionality theory of adversarial attacks.

We consider function spaces defined by self-attention networks without normalization, and theoretically analyze their geometry. Since these networks are polynomial, we rely on tools from algebraic geometry. In particular, we study the identifiability of deep attention by providing a description of the generic fibers of the parametrization for an arbitrary number of layers and, as a consequence, compute the dimension of the function space. Additionally, for a single-layer model, we characterize the singular and boundary points. Finally, we formulate a conjectural extension of our results to normalized self-attention networks, prove it for a single layer, and numerically verify it in the deep case.

In open-ended generative tasks like narrative writing or dialogue, large language models often exhibit cultural biases, showing limited knowledge and generating templated outputs for less prevalent cultures. Recent works show that these biases may stem from uneven cultural representation in pretraining corpora. This work investigates how pretraining leads to biased culture-conditioned generations by analyzing how models associate entities with cultures based on pretraining data patterns. We propose the MEMOED framework (MEMOrization from prEtraining Document) to determine whether a generation for a culture arises from memorization. Using MEMOED on culture-conditioned generations about food and clothing for 110 cultures, we find that high-frequency cultures in pretraining data yield more generations with memorized symbols, while some low-frequency cultures produce none. Additionally, the model favors generating entities with extraordinarily high frequency regardless of the conditioned culture, reflecting biases toward frequent pretraining terms irrespective of relevance. We hope that the MEMOED framework and our insights will inspire more works on attributing model performance on pretraining data.

In-Context Learning (ICL) has significantly expanded the general-purpose nature of large language models, allowing them to adapt to novel tasks using merely the inputted context. This has motivated a series of papers that analyze tractable synthetic domains and postulate precise mechanisms that may underlie ICL. However, the use of relatively distinct setups that often lack a sequence modeling nature to them makes it unclear how general the reported insights from such studies are. Motivated by this, we propose a synthetic sequence modeling task that involves learning to simulate a finite mixture of Markov chains. As we show, models trained on this task reproduce most well-known results on ICL, hence offering a unified setting for studying the concept. Building on this setup, we demonstrate we can explain a model’s behavior by decomposing it into four broad algorithms that combine a fuzzy retrieval vs. inference approach with either unigram or bigram statistics of the context. These algorithms engage in a competitive dynamics to dominate model behavior, with the precise experimental conditions dictating which algorithm ends up superseding others: e.g., we find merely varying context size or amount of training yields (at times sharp) transitions between which algorithm dictates the model behavior, revealing a mechanism that explains the transient nature of ICL. In this sense, we argue ICL is best thought of as a mixture of different algorithms, each with its own peculiarities, instead of a monolithic capability. This also implies that making general claims about ICL that hold universally across all settings may be infeasible.

Deep ensembles improve the performance of the models by taking the average predictions of a group of ensemble members. However, the origin of these capabilities remains a mystery and deep ensembles are used as a reliable “black box” to improve the performance. Existing studies typically attribute such improvement to Jensen gaps of the deep ensemble method, where the loss of the mean does not exceed the mean of the loss for any convex loss metric. In this work, we demonstrate that Jensen’s inequality is not responsible for the effectiveness of deep ensembles, and convexity is not a necessary condition. Instead, Jensen Gap focuses on the “average loss” of individual models, which provides no practical meaning. Thus it fails to explain the core phenomena of deep ensembles such as the superiority to any single ensemble member, the decreasing loss with the number of ensemble members, etc. Regarding this mystery, we provide theoretical analysis and comprehensive empirical results from a statistical perspective that reveal the true mechanism of deep ensembles. Our results highlight that deep ensembles originate from the homogeneous output distribution across all ensemble members. Specifically, the predictions of homogeneous models (Abe et al., 2022b) have the distributional equivalence property – Although the predictions of independent ensemble members are point-wise different, they form an identical distribution. Such agreement and disagreement contribute to deep ensembles’ “magical power”. Based on this discovery, we provide rigorous proof of the effectiveness of deep ensembles and analytically quantify the extent to which ensembles improve performance. The derivations not only theoretically quantify the effectiveness of deep ensembles for the first time, but also enable estimation schemes that foresee the performance of ensembles with different capacities. Furthermore, different from existing studies, our results also point out that deep ensembles work in a different mechanism from model scaling a single model, even though significant correlations between them have been observed.

The cuneiform writing system served as the medium for transmitting knowledge in the ancient Near East for a period of over three thousand years. Cuneiform signs have a complex internal structure which is the subject of expert paleographic analysis, as variations in sign shapes bear witness to historical developments and transmission of writing and culture over time. However, prior automated techniques mostly treat sign types as categorical and do not explicitly model their highly varied internal configurations. In this work, we present an unsupervised approach for recovering the fine-grained internal configuration of cuneiform signs by leveraging powerful generative models and the appearance and structure of prototype font images as priors. Our approach, ProtoSnap, enforces structural consistency on matches found with deep image features to estimate the diverse configurations of cuneiform characters, snapping a skeleton-based template to photographed cuneiform signs. We provide a new benchmark of expert annotations and evaluate our method on this task. Our evaluation shows that our approach succeeds in aligning prototype skeletons to a wide variety of cuneiform signs. Moreover, we show that conditioning on structures produced by our method allows for generating synthetic data with correct structural configurations, significantly boosting the performance of cuneiform sign recognition beyond existing techniques, in particular over rare signs. Our code, data, and trained models are available at the project page: https://tau-vailab.github.io/ProtoSnap/

Sparse autoencoders (SAEs) are a promising approach to interpreting the internal representations of transformer language models. However, SAEs are usually trained separately on each transformer layer, making it difficult to use them to study how information flows across layers. To solve this problem, we introduce the multi-layer SAE (MLSAE): a single SAE trained on the residual stream activation vectors from every transformer layer. Given that the residual stream is understood to preserve information across layers, we expected MLSAE latents to 'switch on' at a token position and remain active at later layers. Interestingly, we find that individual latents are often active at a single layer for a given token or prompt, but the layer at which an individual latent is active may differ for different tokens or prompts. We quantify these phenomena by defining a distribution over layers and considering its variance. We find that the variance of the distributions of latent activations over layers is about two orders of magnitude greater when aggregating over tokens compared with a single token. For larger underlying models, the degree to which latents are active at multiple layers increases, which is consistent with the fact that the residual stream activation vectors at adjacent layers become more similar. Finally, we relax the assumption that the residual stream basis is the same at every layer by applying pre-trained tuned-lens transformations, but our findings remain qualitatively similar. Our results represent a new approach to understanding how representations change as they flow through transformers. We release our code to train and analyze MLSAEs at https://github.com/tim-lawson/mlsae.

The rising interest in leveraging higher-order interactions present in complex systems has led to a surge in more expressive models exploiting higher-order structures in the data, especially in topological deep learning (TDL), which designs neural networks on higher-order domains such as simplicial complexes. However, progress in this field is hindered by the scarcity of datasets for benchmarking these architectures. To address this gap, we introduce MANTRA, the first large-scale, diverse, and intrinsically higher-order dataset for benchmarking higher-order models, comprising over 43,000 and 250,000 triangulations of surfaces and three-dimensional manifolds, respectively. With MANTRA, we assess several graph- and simplicial complex-based models on three topological classification tasks. We demonstrate that while simplicial complex-based neural networks generally outperform their graph-based counterparts in capturing simple topological invariants, they also struggle, suggesting a rethink of TDL. Thus, MANTRA serves as a benchmark for assessing and advancing topological methods, paving the way towards more effective higher-order models.

Natural Language Explanations (NLEs) interpret the decision-making process of a given model through textual sentences. Current NLEs suffer from a severe limitation; they are unfaithful to the model’s actual reasoning process, as a separate textual decoder is explicitly trained to generate those explanations using annotated datasets for a specific task, leading them to reflect what annotators desire. In this work, we take the first step towards generating faithful NLEs for any visual classification model without any training data. Our approach models the relationship between class embeddings from the classifier of the vision model and their corresponding class names via a simple MLP which trains in seconds. After training, we can map any new text to the classifier space and measure its association with the visual features. We conduct experiments on 38 vision models, including both CNNs and Transformers. In addition to NLEs, our method offers other advantages such as zero-shot image classification and fine-grained concept discovery.