ICML 2025 Papers

Large language models (LLMs) exhibit impressive in-context learning (ICL) capability, enabling them to generate relevant responses from a handful of task demonstrations in the prompt. Prior studies have suggested two different explanations for the mechanisms behind ICL: induction heads that find and copy relevant tokens, and function vector (FV) heads whose activations compute a latent encoding of the ICL task. To better understand which of the two distinct mechanisms drives ICL, we study and compare induction heads and FV heads in 12 language models. Through detailed ablations, we find that few-shot ICL is driven primarily by FV heads, especially in larger models. We also find that FV and induction heads are connected: many FV heads start as induction heads during training before transitioning to the FV mechanism. This leads us to speculate that induction facilitates learning the more complex FV mechanism for ICL.

Knowledge Editing (KE) algorithms alter models' weights to perform targeted updates to incorrect, outdated, or otherwise unwanted factual associations. However, recent work has shown that applying KE can adversely affect models' broader factual recall accuracy and diminish their reasoning abilities. Although these studies give insights into the potential harms of KE algorithms, e.g., performance evaluations on benchmarks, little is understood about why such destructive failures occur. Motivated by this, we define a novel synthetic task in which a Transformer is trained from scratch to internalize a "structured" knowledge graph. The structure enforces relationships between entities of the graph, such that editing a factual association has "trickling effects" on other entities (e.g., altering X's parent is Y to Z affects who X's siblings' parent is). Through evaluations of edited models on this task, we show that KE inadvertently affects representations of entities beyond the targeted one, distorting relevant structures that allow a model to infer unseen knowledge about an entity. We call this phenomenon representation shattering and demonstrate that it degrades models' factual recall and reasoning performance. We further corroborate our findings in naturalistic settings with pre-trained Llama and Mamba models as well. Overall, our work yields a precise mechanistic hypothesis to explain why KE has adverse effects on model abilities.

In this position paper, we argue that user studies are key to understanding the value of explainable AI methods, because the end goal of explainable AI is to satisfy societal desiderata. We also argue that the current state of user studies is detrimental to the advancement of the field. We support this argument with a review of general and explainable AI-specific challenges, as well as an analysis of 607 explainable AI papers featuring user studies. We demonstrate how most user studies lack reproducibility, discussion of limitations, comparison with a baseline, or placebo explanations and are of low fidelity to real-world users and application context. This, combined with an overreliance on functional evaluation, results in a lack of understanding of the value explainable AI methods, which hinders the progress of the field. To address this issue, we call for higher methodological standards for user studies, greater appreciation of high-quality user studies in the AI community, and reduced reliance on functional evaluation.

Methods for knowledge editing and unlearning in large language models seek to edit or remove undesirable knowledge or capabilities without compromising general language modeling performance. This work investigates how mechanistic interpretability---which, in part, aims to identify model components (circuits) associated to specific interpretable mechanisms that make up a model capability---can improve the precision and effectiveness of editing and unlearning. We find a stark difference in unlearning and edit robustness when training components localized by different methods. We highlight an important distinction between methods that localize components based primarily on preserving outputs, and those finding high level mechanisms with predictable intermediate states. In particular, localizing edits/unlearning to components associated with the *lookup-table mechanism* for factual recall 1) leads to more robust edits/unlearning across different input/output formats, and 2) resists attempts to relearn the unwanted information, while also reducing unintended side effects compared to baselines, on both a sports facts dataset and the CounterFact dataset across multiple models. We also find that certain localized edits disrupt the latent knowledge in the model more than any other baselines, making unlearning more robust to various attacks.

Understanding the function of individual units in a neural network is an important building block for mechanistic interpretability. This is often done by generating a simple text explanation of the behavior of individual neurons or units. For these explanations to be useful, we must understand how reliable and truthful they are. In this work we unify many existing explanation evaluation methods under one mathematical framework. This allows us to compare and contrast existing evaluation metrics, understand the evaluation pipeline with increased clarity and apply existing statistical concepts on the evaluation. In addition, we propose two simple sanity checks on the evaluation metrics and show that many commonly used metrics fail these tests and do not change their score after massive changes to the concept labels. Based on our experimental and theoretical results, we propose guidelines that future evaluations should follow and identify a set of reliable evaluation metrics.

Sparse Autoencoders (SAEs) are a popular method for decomposing Large Language Model (LLM) activations into interpretable latents, however they have a substantial training cost and SAEs learned on different models are not directly comparable. Motivated by relative representation similarity measures, we introduce Inference-Time Decomposition of Activation models (ITDAs). ITDAs are constructed by greedily sampling activations into a dictionary based on an error threshold on their matching pursuit reconstruction. ITDAs can be trained in 1% of the time of SAEs, allowing us to cheaply train them on Llama-3.1 70B and 405B. ITDA dictionaries also enable cross-model comparisons, and outperform existing methods like CKA, SVCCA, and a relative representation method on a benchmark of representation similarity. Code available at https://github.com/pleask/itda.

Fine-tuning significantly improves the performance of Large Language Models (LLMs), yet its underlying mechanisms remain poorly understood. This paper aims to provide an in-depth interpretation of the fine-tuning process through circuit analysis, a popular tool in *Mechanistic Interpretability (MI)*. Unlike previous studies (Prakash et al. 2024, Chhabra et al. 2024) that focus on tasks where pre-trained models already perform well, we develop a set of mathematical tasks where fine-tuning yields substantial performance gains, bringing the setup closer to real-world scenarios. In our experiments, we identify circuits at various checkpoints during fine-tuning and examine the interplay between circuit analysis, fine-tuning methods, and task complexities. First, we find that while circuits maintain high node similarity before and after fine-tuning, their edges undergo significant changes, contrasting with previous work (Prakash et al. 2024, Chhabra et al. 2024) that reported only small circuit additions after fine-tuning. Based on these observations, we develop a **circuit-aware Low-Rank Adaptation (LoRA)** method that assigns ranks to layers according to edge changes in the circuits. Experimental results demonstrate that our circuit-based LoRA achieves an average improvement of 2.46% over standard LoRA with comparable parameter sizes. Furthermore, we explore how combining circuits from subtasks can enhance fine-tuning in compositional tasks, offering new insights into task design and deepening our understanding of circuit dynamics and fine-tuning mechanisms.

Even when massively overparameterized, deep neural networks show a remarkable ability to generalize. Research on this phenomenon has focused on generalization within distribution, via smooth interpolation. Yet in some settings neural networks also learn to extrapolate to data far beyond the bounds of the original training set, sometimes even allowing for infinite generalization, implying that an algorithm capable of solving the task has been learned. Here we undertake a case study of the learning dynamics of recurrent neural networks trained on the streaming parity task in order to develop an effective theory of algorithm development. The streaming parity task is a simple but nonlinear task defined on sequences up to arbitrary length. We show that, with sufficient finite training experience, RNNs exhibit a phase transition to perfect infinite generalization. Using an effective theory for the representational dynamics, we find an implicit representational merger effect which can be interpreted as the construction of a finite automaton that reproduces the task. Overall, our results disclose one mechanism by which neural networks can generalize infinitely from finite training experience.

We study how multi-head softmax attention models are trained to perform in-context learning on linear data. Through extensive empirical experiments and rigorous theoretical analysis, we demystify the emergence of elegant attention patterns: a diagonal and homogeneous pattern in the key-query weights, and a last-entry-only and zero-sum pattern in the output-value weights. Remarkably, these patterns consistently appear from gradient-based training starting from random initialization. Our analysis reveals that such emergent structures enable multi-head attention to approximately implement a debiased gradient descent predictor --- one that outperforms single-head attention and nearly achieves Bayesian optimality up to proportional factor. We also extend our study to scenarios with anisotropic covariates and multi-task linear regression. Our results reveal that in-context learning ability emerges from the trained transformer as an aggregated effect of its architecture and the underlying data distribution, paving the way for deeper understanding and broader applications of in-context learning.

Uncovering the inner mechanisms of Transformer models offers insights into how they process and represent information. In this work, we analyze the computation performed by Transformers in the layers after the top-1 prediction remains fixed, known as the “saturation event”. We expand this concept to top-k tokens, demonstrating that similar saturation events occur across language, vision, and speech models. We find that these events occur in order of the corresponding tokens’ ranking, i.e., the model first decides on the top ranking token, then the second highest ranking token, and so on. This phenomenon seems intrinsic to the Transformer architecture, occurring across different variants, and even in untrained Transformers. We propose that these events reflect task transitions, where determining each token corresponds to a discrete task. We show that it is possible to predict the current task from hidden layer embedding, and demonstrate that we can cause the model to switch to the next task via intervention. Leveraging our findings, we introduce a token-level early-exit strategy, surpassing existing methods in balancing performance and efficiency and show how to exploit saturation events for better language modeling.

Autoregressive transformers exhibit adaptive learning through in-context learning (ICL), which begs the question of how. Prior works have shown that transformers represent the ICL tasks as vectors in their representations. In this paper, we leverage the encoding-decoding framework to study how transformers form task vectors during pretraining and how their task encoding quality predicts ICL task performance. On synthetic ICL tasks, we analyze the training dynamics of a small transformer and report the coupled emergence of task encoding and decoding. As the model learns to encode different latent tasks (e.g., "Finding the first noun in a sentence.") into distinct, separable representations, it concurrently builds conditional decoding algorithms and improves its ICL performance. We validate this phenomenon across pretrained models of varying scales (Gemma-2 2B/9B/27B, Llama-3.1 8B/70B) and over the course of pretraining in OLMo-7B. Further, we demonstrate that the quality of task encoding inferred from representations predicts ICL performance, and that, surprisingly, finetuning the earlier layers can improve the task encoding and performance more than finetuning the latter layers. Our empirical insights shed light into better understanding the success and failure modes of large language models via their representations.

Language Models (LMs) often encounter knowledge conflicts when parametric memory contradicts contextual knowledge. Previous works attribute this conflict to the interplay between "memory heads" and "context heads", attention heads assumed to promote either memory or context exclusively. In this study, we go beyond this fundamental assumption by uncovering a critical phenomenon we term the *superposition of contextual information and parametric memory*, where highly influential attention heads simultaneously contribute to both memory and context. Building upon this insight, we propose Just Run Twice (JuICE), a test-time attention intervention method that steers LMs toward either parametric beliefs or contextual knowledge without requiring fine-tuning. JuICE identifies a set of reliable attention heads and leverages a dual-run approach to mitigate the superposition effects. Extensive experiments across 11 datasets and 6 model architectures demonstrate that JuICE sets the new state-of-the-art performance and robust generalization, achieving significant and consistent improvement across different domains under various conflict types. Finally, we theoretically analyze knowledge conflict and the superposition of contextual information and parametric memory in attention heads, which further elucidates the effectiveness of JuICE in these settings. Our code is available at https://github.com/GaotangLi/JUICE.