ICLR 2025 Papers

In this study, we propose a novel rule-based interpretable choice model, {\bf Logic-Logit}, designed to effectively learn and explain human choices. Choice models have been widely applied across various domains—such as commercial demand forecasting, recommendation systems, and consumer behavior analysis—typically categorized as parametric, nonparametric, or deep network-based. While recent innovations have favored neural network approaches for their computational power, these flexible models often involve large parameter sets and lack interpretability, limiting their effectiveness in contexts where transparency is essential. Previous empirical evidence shows that individuals usually use {\it heuristic decision rules} to form their consideration sets, from which they then choose. These rules are often represented as {\it disjunctions of conjunctions} (i.e., OR-of-ANDs). These rules-driven, {\it consider-then-choose} decision processes enable people to quickly screen numerous alternatives while reducing cognitive and search costs. Motivated by this insight, our approach leverages logic rules to elucidate human choices, providing a fresh perspective on preference modeling. We introduce a unique combination of column generation techniques and the Frank-Wolfe algorithm to facilitate efficient rule extraction for preference modeling—a process recognized as NP-hard. Our empirical evaluation, conducted on both synthetic datasets and real-world data from commercial and healthcare domains, demonstrates that Logic-Logit significantly outperforms baseline models in terms of interpretability and accuracy.

Traditional data influence estimation methods, like influence function, assume that learning algorithms are permutation-invariant with respect to training data. However, modern training paradigms—especially for foundation models using stochastic algorithms and non-convergent, multi-stage curricula—are sensitive to data ordering, thus violating this assumption. This mismatch renders influence functions inadequate for answering some critical questions in current machine learning: How can we differentiate the influence of the same data contributing at different stages of training? More generally, how can we capture the dependence of data influence on the optimization trajectory during training? To address this gap, we formalize the concept of \emph{trajectory-specific leave-one-out (LOO) influence}, which quantifies the impact of removing a data point from a specific iteration during training, accounting for the exact sequence of data encountered and the model's optimization trajectory. However, exactly evaluating the trajectory-specific LOO presents a significant computational challenge. To address this, we propose \emph{data value embedding}, a novel technique enabling efficient approximation of trajectory-specific LOO. Specifically, we compute a training data embedding that encapsulates the cumulative interactions between data and the evolving model parameters. The LOO can then be efficiently approximated through a simple dot-product between the data value embedding and the gradient of the given test data. As data value embedding captures training data ordering, it offers valuable insights into model training dynamics. In particular, we uncover distinct phases of data influence, revealing that data points in the early and late stages of training exert a greater impact on the final model. These insights translate into actionable strategies for managing the computational overhead of data selection by strategically timing the selection process, potentially opening new avenues in data curation research.

Large language models (LLMs) have been shown to face hallucination issues due to the data they trained on often containing human bias; whether this is reflected in the decision-making process of LLM agents remains under-explored. As LLM Agents are increasingly employed in intricate social environments, a pressing and natural question emerges: Can we utilize LLM Agents' systematic hallucinations to mirror human cognitive biases, thus exhibiting irrational social intelligence? In this paper, we probe the irrational behavior among contemporary LLM agents by melding practical social science experiments with theoretical insights. Specifically, we propose CogMir, an open-ended Multi-LLM Agents framework that utilizes hallucination properties to assess and enhance LLM Agents’ social intelligence through cognitive biases. Experimental results on CogMir subsets show that LLM Agents and humans exhibit high consistency in irrational and prosocial decision-making under uncertain conditions, underscoring the prosociality of LLM Agents as social entities and highlighting the significance of hallucination properties. Additionally, CogMir framework demonstrates its potential as a valuable platform for encouraging more research into the social intelligence of LLM Agents.

Learning a transition model via Maximum Likelihood Estimation (MLE) followed by planning inside the learned model is perhaps the most standard and simplest Model-based Reinforcement Learning (RL) framework. In this work, we show that such a simple Model-based RL scheme, when equipped with optimistic and pessimistic planning procedures, achieves strong regret and sample complexity bounds in online and offline RL settings. Particularly, we demonstrate that under the conditions where the trajectory-wise reward is normalized between zero and one and the transition is time-homogenous, it achieves nearly horizon-free and second-order bounds.

Large language models (LLMs) store extensive factual knowledge, but the mechanisms behind how they store and express this knowledge remain unclear. The Knowledge Neuron (KN) thesis is a prominent theory for explaining these mechanisms. This theory is based on the **Knowledge Localization (KL)** assumption, which suggests that a fact can be localized to a few knowledge storage units, namely knowledge neurons. However, this assumption has two limitations: first, it may be too rigid regarding knowledge storage, and second, it neglects the role of the attention module in knowledge expression. In this paper, we first re-examine the KL assumption and demonstrate that its limitations do indeed exist. To address these, we then present two new findings, each targeting one of the limitations: one focusing on knowledge storage and the other on knowledge expression. We summarize these findings as **Query Localization** assumption and argue that the KL assumption can be viewed as a simplification of the QL assumption. Based on QL assumption, we further propose the Consistency-Aware KN modification method, which improves the performance of knowledge modification, further validating our new assumption. We conduct 39 sets of experiments, along with additional visualization experiments, to rigorously confirm our conclusions. Code will be made public soon.

Sparse autoencoders provide a promising unsupervised approach for extracting interpretable features from a language model by reconstructing activations from a sparse bottleneck layer. Since language models learn many concepts, autoencoders need to be very large to recover all relevant features. However, studying the properties of autoencoder scaling is difficult due to the need to balance reconstruction and sparsity objectives and the presence of dead latents. We propose using k-sparse autoencoders [Makhzani and Frey, 2013] to directly control sparsity, simplifying tuning and improving the reconstruction-sparsity frontier. Additionally, we find modifications that result in few dead latents, even at the largest scales we tried. Using these techniques, we find clean scaling laws with respect to autoencoder size and sparsity. We also introduce several new metrics for evaluating feature quality based on the recovery of hypothesized features, the explainability of activation patterns, and the sparsity of downstream effects. These metrics all generally improve with autoencoder size. To demonstrate the scalability of our approach, we train a 16 million latent autoencoder on GPT-4 activations for 40 billion tokens. We release training code and autoencoders for open-source models, as well as a visualizer.

Language Model (LM) agents for cybersecurity that are capable of autonomously identifying vulnerabilities and executing exploits have potential to cause real-world impact. Policymakers, model providers, and researchers in the AI and cybersecurity communities are interested in quantifying the capabilities of such agents to help mitigate cyberrisk and investigate opportunities for penetration testing. Toward that end, we introduce Cybench, a framework for specifying cybersecurity tasks and evaluating agents on those tasks. We include 40 professional-level Capture the Flag (CTF) tasks from 4 distinct CTF competitions, chosen to be recent, meaningful, and spanning a wide range of difficulties. Each task includes its own description, starter files, and is initialized in an environment where an agent can execute commands and observe outputs. Since many tasks are beyond the capabilities of existing LM agents, we introduce subtasks for each task, which break down a task into intermediary steps for a more detailed evaluation. To evaluate agent capabilities, we construct a cybersecurity agent and evaluate 8 models: GPT-4o, OpenAI o1-preview, Claude 3 Opus, Claude 3.5 Sonnet, Mixtral 8x22b Instruct, Gemini 1.5 Pro, Llama 3 70B Chat, and Llama 3.1 405B Instruct. For the top performing models (GPT-4o and Claude 3.5 Sonnet), we further investigate performance across 4 agent scaffolds (structured bash, action-only, pseudoterminal, and web search). Without subtask guidance, agents leveraging Claude 3.5 Sonnet, GPT-4o, OpenAI o1-preview, and Claude 3 Opus successfully solved complete tasks that took human teams up to 11 minutes to solve. In comparison, the most difficult task took human teams 24 hours and 54 minutes to solve. Anonymized code and data are available at https://drive.google.com/file/d/1kp3H0pw1WMAH-Qyyn9WA0ZKmEa7Cr4D4 and https://drive.google.com/file/d/1BcTQ02BBR0m5LYTiK-tQmIK17_TxijIy.

Large language models (LLMs) are vulnerable to adversarial attacks that can elicit harmful responses. Defending against such attacks remains challenging due to the opacity of jailbreaking mechanisms and the high computational cost of training LLMs robustly. We demonstrate that adversarial attacks share a universal mechanism for circumventing LLM safeguards that works by ablating a dimension in the residual stream embedding space called the refusal feature. We further show that the operation of refusal feature ablation (RFA) approximates the worst-case perturbation of offsetting model safety. Based on these findings, we propose Refusal Feature Adversarial Training (ReFAT), a novel algorithm that efficiently performs LLM adversarial training by simulating the effect of input-level attacks via RFA. Experiment results show that ReFAT significantly improves the robustness of three popular LLMs against a wide range of adversarial attacks, with considerably less computational overhead compared to existing adversarial training methods.

Prior work has shown that text-conditioned diffusion models can learn to identify and manipulate primitive concepts underlying a compositional data-generating process, enabling generalization to entirely novel, out-of-distribution compositions. Beyond performance evaluations, these studies develop a rich empirical phenomenology of learning dynamics, showing that models generalize sequentially, respecting the compositional hierarchy of the data-generating process. Moreover, concept-centric structures within the data significantly influence a model's speed of learning the ability to manipulate a concept. In this paper, we aim to better characterize these empirical results from a theoretical standpoint. Specifically, we propose an abstraction of prior work's compositional generalization problem by introducing a structured identity mapping (SIM) task, where a model is trained to learn the identity mapping on a Gaussian mixture with structurally organized centroids. We mathematically analyze the learning dynamics of neural networks trained on this SIM task and show that, despite its simplicity, SIM's learning dynamics capture and help explain key empirical observations on compositional generalization with diffusion models identified in prior work. Our theory also offers several new insights---e.g., we find a novel mechanism for non-monotonic learning dynamics of test loss in early phases of training. We validate our new predictions by training a text-conditioned diffusion model, bridging our simplified framework and complex generative models. Overall, this work establishes the SIM task as a meaningful theoretical abstraction of concept learning dynamics in modern generative models.

Large language models (LLMs) have demonstrated the ability to generate text that realistically reflects a range of different subjective human perspectives. This paper studies how LLMs are seemingly able to reflect more liberal versus more conservative viewpoints among other political perspectives in American politics. We show that LLMs possess linear representations of political perspectives within activation space, wherein more similar perspectives are represented closer together. To do so, we probe the attention heads across the layers of three open transformer-based LLMs (Llama-2-7b-chat, Mistral-7b-instruct, Vicuna-7b). We first prompt models to generate text from the perspectives of different U.S. lawmakers. We then identify sets of attention heads whose activations linearly predict those lawmakers' DW-NOMINATE scores, a widely-used and validated measure of political ideology. We find that highly predictive heads are primarily located in the middle layers, often speculated to encode high-level concepts and tasks. Using probes only trained to predict lawmakers' ideology, we then show that the same probes can predict measures of news outlets' slant from the activations of models prompted to simulate text from those news outlets. These linear probes allow us to visualize, interpret, and monitor ideological stances implicitly adopted by an LLM as it generates open-ended responses. Finally, we demonstrate that by applying linear interventions to these attention heads, we can steer the model outputs toward a more liberal or conservative stance. Overall, our research suggests that LLMs possess a high-level linear representation of American political ideology and that by leveraging recent advances in mechanistic interpretability, we can identify, monitor, and steer the subjective perspective underlying generated text.

In this paper, I introduce the retrieval problem, a simple yet common reasoning task that can be solved only by transformers with a minimum number of layers, which grows logarithmically with the input size. I empirically show that large language models can solve the task under different prompting formulations without any fine-tuning. To understand how transformers solve the retrieval problem, I train several transformers on a minimal formulation. Successful learning occurs only under the presence of an implicit curriculum. I uncover the learned mechanisms by studying the attention maps in the trained transformers. I also study the training process, uncovering that attention heads always emerge in a specific sequence guided by the implicit curriculum.

This paper introduces PhyloLM, a method adapting phylogenetic algorithms to Large Language Models (LLMs) to explore whether and how they relate to each other and to predict their performance characteristics. Our method calculates a phylogenetic distance metric based on the similarity of LLMs' output. The resulting metric is then used to construct dendrograms, which satisfactorily capture known relationships across a set of 111 open-source and 45 closed models. Furthermore, our phylogenetic distance predicts performance in standard benchmarks, thus demonstrating its functional validity and paving the way for a time and cost-effective estimation of LLM capabilities. To sum up, by translating population genetic concepts to machine learning, we propose and validate a tool to evaluate LLM development, relationships and capabilities, even in the absence of transparent training information.

Outliers have been widely observed in Large Language Models (LLMs), significantly impacting model performance and posing challenges for model compression. Understanding the functionality and formation mechanisms of these outliers is critically important. Existing works, however, largely focus on reducing the impact of outliers from an algorithmic perspective, lacking an in-depth investigation into their causes and roles. In this work, we provide a detailed analysis of the formation process, underlying causes, and functions of outliers in LLMs. We define and categorize three types of outliers—activation outliers, weight outliers, and attention outliers—and analyze their distributions across different dimensions, uncovering inherent connections between their occurrences and their ultimate influence on the attention mechanism. Based on these observations, we hypothesize and explore the mechanisms by which these outliers arise and function, demonstrating through theoretical derivations and experiments that they emerge due to the self-attention mechanism's softmax operation. These outliers act as implicit context-aware scaling factors within the attention mechanism. As these outliers stem from systematic influences, we term them systematic outliers. Our study not only enhances the understanding of Transformer-based LLMs but also shows that structurally eliminating outliers can accelerate convergence and improve model compression. The code is avilable at \url{https://github.com/an-yongqi/systematic-outliers}.

The need to explain decisions made by AI systems is driven by both recent regulation and user demand. The decisions are often explainable only post hoc. In counterfactual explanations, one may ask what constitutes the best counterfactual explanation. Clearly, multiple criteria must be taken into account, although "distance from the sample" is a key criterion. Recent methods that consider the plausibility of a counterfactual seem to sacrifice this original objective. Here, we present a system that provides high-likelihood explanations that are, at the same time, close and sparse. We show that the search for the most likely explanations satisfying many common desiderata for counterfactual explanations can be modeled using Mixed-Integer Optimization (MIO). We use a Sum-Product Network (SPN) to estimate the likelihood of a counterfactual. To achieve that, we propose an MIO formulation of an SPN, which can be of independent interest. The source code with examples is available at https://github.com/Epanemu/LiCE.

Kolmogorov–Arnold Network (KAN) is a network structure recently proposed in Liu et al. (2024) that offers improved interpretability and a more parsimonious design in many science-oriented tasks compared to multi-layer perceptrons. This work provides a rigorous theoretical analysis of KAN by establishing generalization bounds for KAN equipped with activation functions that are either represented by linear combinations of basis functions or lying in a low-rank Reproducing Kernel Hilbert Space (RKHS). In the first case, the generalization bound accommodates various choices of basis functions in forming the activation functions in each layer of KAN and is adapted to different operator norms at each layer. For a particular choice of operator norms, the bound scales with the $l_1$ norm of the coefficient matrices and the Lipschitz constants for the activation functions, and it has no dependence on combinatorial parameters (e.g., number of nodes) outside of logarithmic factors. Moreover, our result does not require the boundedness assumption on the loss function and, hence, is applicable to a general class of regression-type loss functions. In the low-rank case, the generalization bound scales polynomially with the underlying ranks as well as the Lipschitz constants of the activation functions in each layer. These bounds are empirically investigated for KANs trained with stochastic gradient descent on simulated and real data sets. The numerical results demonstrate the practical relevance of these bounds.

Graphs are ubiquitous in social networks, chemical molecules, and financial data, where Graph Neural Networks (GNNs) achieve superior predictive accuracy. Graphs can be evolving, while understanding how GNN predictions respond to the evolution provides significant insight and trust. We explore the problem of explaining evolving GNN predictions due to continuously changing edge weights. We introduce a layer edge-based explanation to balance explanation fidelity and interpretability. We propose a novel framework to address the challenges of axiomatic attribution and the entanglement of multiple computational graph paths due to continuous change of edge weights. We first design an axiomatic attribution of the evolution of the model prediction to message flows, then develop Shapley value to fairly map message flow contributions to layer edges. We formulate a novel optimization problem to find the critical layer edges based on KL-divergence minimization. Extensive experiments on eight datasets for node classification, link prediction, and graph classification tasks with evolving graphs demonstrate the better fidelity and interpretability of the proposed method over the baseline methods. The code is available at https://github.com/yazhengliu/Axiomatic-Layer-Edges/tree/main.

As LLMs increasingly impact society, their ability to represent diverse perspectives is critical. However, recent studies reveal that alignment algorithms such as RLHF and DPO significantly reduce the diversity of LLM outputs. Not only do aligned LLMs generate text with repetitive structure and word choice, they also approach problems in more uniform ways, and their responses reflect a narrower range of societal perspectives. We attribute this problem to the KL divergence regularizer employed in preference learning algorithms. This causes the model to overweight majority opinions and sacrifice diversity in exchange for optimal reward. To address this, we propose Soft Preference Learning, which decouples the entropy and cross-entropy terms in the KL penalty — allowing for fine-grained control over LLM generation diversity. From a capabilities perspective, LLMs trained using Soft Preference Learning attain higher accuracy on difficult repeated sampling tasks and produce outputs with greater semantic and lexical diversity. From an alignment perspective, they are capable of representing a wider range of societal viewpoints and display improved logit calibration. Notably, Soft Preference Learning resembles, but is a Pareto improvement over, standard temperature scaling.

In this paper, we study the loss landscape of one-hidden-layer neural networks with ReLU-like activation functions trained with the empirical squared loss using gradient descent (GD). We identify the stationary points of such networks, which significantly slow down loss decrease during training. To capture such points while accounting for the non-differentiability of the loss, the stationary points that we study are directional stationary points, rather than other notions like Clarke stationary points. We show that, if a stationary point does not contain "escape neurons", which are defined with first-order conditions, it must be a local minimum. Moreover, for the scalar-output case, the presence of an escape neuron guarantees that the stationary point is not a local minimum. Our results refine the description of the *saddle-to-saddle* training process starting from infinitesimally small (vanishing) initialization for shallow ReLU-like networks: By precluding the saddle escape types that previous works did not rule out, we advance one step closer to a complete picture of the entire dynamics. Moreover, we are also able to fully discuss how network embedding, which is to instantiate a narrower network with a wider network, reshapes the stationary points.

Pre-trained large language models (LLMs) have been reliably integrated with visual input for multimodal tasks. The widespread adoption of instruction-tuned image-to-text vision-language assistants (VLAs) like LLaVA and InternVL necessitates evaluating gender biases. We study gender bias in 22 popular open-source VLAs with respect to personality traits, skills, and occupations. Our results show that VLAs replicate human biases likely present in the data, such as real-world occupational imbalances. Similarly, they tend to attribute more skills and positive personality traits to women than to men, and we see a consistent tendency to associate negative personality traits with men. To eliminate the gender bias in these models, we find that fine-tuning-based debiasing methods achieve the best trade-off between debiasing and retaining performance on downstream tasks. We argue for pre-deploying gender bias assessment in VLAs and motivate further development of debiasing strategies to ensure equitable societal outcomes. Code is available at https://github.com/ExplainableML/vla-gender-bias.