Mattia Jacopo Villani, Nandi Schoots
We constructively prove that every deep ReLU network can be rewritten as a functionally identical three-layer network with weights valued in the extended reals. Based on this proof, we provide an algorithm that, given a deep ReLU network, finds the explicit weights of the corresponding shallow network. The resulting shallow network is transparent and used to generate explanations of the model s behaviour.
Nicholas Pochinkov, Nandi Schoots
Understanding and shaping the behaviour of Large Language Models (LLMs) is increasingly important as applications become more powerful and more frequently adopted. This paper introduces a machine unlearning method specifically designed for LLMs. We introduce a selective pruning method for LLMs that removes neurons based on their relative importance on a targeted capability compared to overall network performance. This approach is a compute- and data-efficient method for identifying and removing neurons that enable specific behaviours. Our findings reveal that both feed-forward and attention neurons in LLMs are specialized; that is, for specific tasks, certain neurons are more crucial than others. Code from all experiments is available at https://github.com/nickypro/selective-pruning
Hugo Fry, Seamus Fallows, Ian Fan et al.
We investigate the optimization target of Contrast-Consistent Search (CCS), which aims to recover the internal representations of truth of a large language model. We present a new loss function that we call the Midpoint-Displacement (MD) loss function. We demonstrate that for a certain hyper-parameter value this MD loss function leads to a prober with very similar weights to CCS. We further show that this hyper-parameter is not optimal and that with a better hyper-parameter the MD loss function attains a higher test accuracy than CCS.
Ali Holmov, Paul Youssef, Nandi Schoots et al.
Knowledge editing methods such as ROME and MEMIT update factual associations in transformer models by modifying MLP weights. While evaluated mainly by output behavior, their internal mechanism remains underexplored. We investigate whether edits rely on a common mechanism, regardless of which fact is modified. Despite fact-specific weight changes, we argue that ROME and MEMIT target the same subset of weights critical for maintaining edits. To isolate this subset, we train a compact binary mask over the edited weights. The mask reverses 80% of edits on the training set and over 70% on the test set, confirming that diverse edits share a common functional structure. Our analysis reveals that the mask reverses edits by eliminating overattention in later layers. Additionally, we show that injecting the mask during editing drops editing success from 98% to 38%, demonstrating that this mechanism is necessary for edits to succeed. Our finding that edits suppress rather than overwrite knowledge explains why ROME and MEMIT fail to propagate changes to related facts. The identified common functional subspace informs detection and defense against unwanted edits.
Satvik Golechha, Dylan Cope, Nandi Schoots
An approach to improve network interpretability is via clusterability, i.e., splitting a model into disjoint clusters that can be studied independently. We find pretrained models to be highly unclusterable and thus train models to be more modular using an ``enmeshment loss'' function that encourages the formation of non-interacting clusters. Using automated interpretability measures, we show that our method finds clusters that learn different, disjoint, and smaller circuits for CIFAR-10 labels. Our approach provides a promising direction for making neural networks easier to interpret.
Lee Sharkey, Bilal Chughtai, Joshua Batson et al. · deepmind
Mechanistic interpretability aims to understand the computational mechanisms underlying neural networks' capabilities in order to accomplish concrete scientific and engineering goals. Progress in this field thus promises to provide greater assurance over AI system behavior and shed light on exciting scientific questions about the nature of intelligence. Despite recent progress toward these goals, there are many open problems in the field that require solutions before many scientific and practical benefits can be realized: Our methods require both conceptual and practical improvements to reveal deeper insights; we must figure out how best to apply our methods in pursuit of specific goals; and the field must grapple with socio-technical challenges that influence and are influenced by our work. This forward-facing review discusses the current frontier of mechanistic interpretability and the open problems that the field may benefit from prioritizing.
Ole Jorgensen, Dylan Cope, Nandi Schoots et al.
Recent work in activation steering has demonstrated the potential to better control the outputs of Large Language Models (LLMs), but it involves finding steering vectors. This is difficult because engineers do not typically know how features are represented in these models. We seek to address this issue by applying the idea of mean-centring to steering vectors. We find that taking the average of activations associated with a target dataset, and then subtracting the mean of all training activations, results in effective steering vectors. We test this method on a variety of models on natural language tasks by steering away from generating toxic text, and steering the completion of a story towards a target genre. We also apply mean-centring to extract function vectors, more effectively triggering the execution of a range of natural language tasks by a significant margin (compared to previous baselines). This suggests that mean-centring can be used to easily improve the effectiveness of activation steering in a wide range of contexts.
Ari Spiesberger, Juan J. Vazquez, Nicky Pochinkov et al.
If LLM training data is polluted with benchmark test data, then benchmark performance gives biased estimates of out-of-distribution (OOD) generalization. Typical decontamination filters use n-gram matching which fail to detect semantic duplicates: sentences with equivalent (or near-equivalent) content that are not close in string space. We study this soft contamination of training data by semantic duplicates. Among other experiments, we embed the Olmo3 training corpus and find that: 1) contamination remains widespread, e.g. we find semantic duplicates for 78% of CodeForces and exact duplicates for 50% of ZebraLogic problems; 2) including semantic duplicates of benchmark data in training does improve benchmark performance; and 3) when finetuning on duplicates of benchmark datapoints, performance also improves on truly-held-out datapoints from the same benchmark. We argue that recent benchmark gains are thus confounded: the prevalence of soft contamination means gains reflect both genuine capability improvements and the accumulation of test data and effective test data in growing training corpora.
Teun van der Weij, Massimo Poesio, Nandi Schoots
Current large language models have dangerous capabilities, which are likely to become more problematic in the future. Activation steering techniques can be used to reduce risks from these capabilities. In this paper, we investigate the efficacy of activation steering for broad skills and multiple behaviours. First, by comparing the effects of reducing performance on general coding ability and Python-specific ability, we find that steering broader skills is competitive to steering narrower skills. Second, we steer models to become more or less myopic and wealth-seeking, among other behaviours. In our experiments, combining steering vectors for multiple different behaviours into one steering vector is largely unsuccessful. On the other hand, injecting individual steering vectors at different places in a model simultaneously is promising.
Sonakshi Chauhan, Maheep Chaudhary, Koby Choy et al.
In this paper we explore where information is collected and how it is propagated throughout layers in large language models (LLMs). We begin by examining the surprising computational importance of punctuation tokens which previous work has identified as attention sinks and memory aids. Using intervention-based techniques, we evaluate the necessity and sufficiency (for preserving model performance) of punctuation tokens across layers in GPT-2, DeepSeek, and Gemma. Our results show stark model-specific differences: for GPT-2, punctuation is both necessary and sufficient in multiple layers, while this holds far less in DeepSeek and not at all in Gemma. Extending beyond punctuation, we ask whether LLMs process different components of input (e.g., subjects, adjectives, punctuation, full sentences) by forming early static summaries reused across the network, or if the model remains sensitive to changes in these components across layers. Extending beyond punctuation, we investigate whether different reasoning rules are processed differently by LLMs. In particular, through interchange intervention and layer-swapping experiments, we find that conditional statements (if, then), and universal quantification (for all) are processed very differently. Our findings offer new insight into the internal mechanisms of punctuation usage and reasoning in LLMs and have implications for interpretability.
Nandi Schoots, Mattia Jacopo Villani, Niels uit de Bos
Kolmogorov-Arnold Networks are a new family of neural network architectures which holds promise for overcoming the curse of dimensionality and has interpretability benefits (arXiv:2404.19756). In this paper, we explore the connection between Kolmogorov Arnold Networks (KANs) with piecewise linear (univariate real) functions and ReLU networks. We provide completely explicit constructions to convert a piecewise linear KAN into a ReLU network and vice versa.
Satvik Golechha, Maheep Chaudhary, Joan Velja et al.
An approach to improve neural network interpretability is via clusterability, i.e., splitting a model into disjoint clusters that can be studied independently. We define a measure for clusterability and show that pre-trained models form highly enmeshed clusters via spectral graph clustering. We thus train models to be more modular using a "clusterability loss" function that encourages the formation of non-interacting clusters. We then investigate the emerging properties of these highly clustered models. We find our trained clustered models do not exhibit more task specialization, but do form smaller circuits. We investigate CNNs trained on MNIST and CIFAR, small transformers trained on modular addition, and GPT-2 and Pythia on the Wiki dataset, and Gemma on a Chemistry dataset. This investigation shows what to expect from clustered models.
Nandi Schoots, Alex Jackson, Ali Kholmovaia et al.
Does the process of training a neural network to solve a task tend to use all of the available weights even when the task could be solved with fewer weights? To address this question we study the effects of pruning fully connected, convolutional and residual models while varying their widths. We find that the proportion of weights that can be pruned without degrading performance is largely invariant to model size. Increasing the width of a model has little effect on the density of the pruned model relative to the increase in absolute size of the pruned network. In particular, we find substantial prunability across a large range of model sizes, where our biggest model is 50 times as wide as our smallest model. We explore three hypotheses that could explain these findings.
Nandi Schoots, Dylan Cope
We study the relationship between the entropy of intermediate representations and a model's robustness to distributional shift. We train models consisting of two feed-forward networks end-to-end separated by a discrete $n$-bit channel on an unsupervised contrastive learning task. Different masking strategies are applied after training that remove a proportion of low-entropy bits, high-entropy bits, or randomly selected bits, and the effects on performance are compared to the baseline accuracy with no mask. We hypothesize that the entropy of a bit serves as a guide to its usefulness out-of-distribution (OOD). Through experiment on three OOD datasets we demonstrate that the removal of low-entropy bits can notably benefit OOD performance. Conversely, we find that top-entropy masking disproportionately harms performance both in-distribution (InD) and OOD.
Adam X. Yang, Maxime Robeyns, Edward Milsom et al.
The successes of modern deep machine learning methods are founded on their ability to transform inputs across multiple layers to build good high-level representations. It is therefore critical to understand this process of representation learning. However, standard theoretical approaches (formally NNGPs) involving infinite width limits eliminate representation learning. We therefore develop a new infinite width limit, the Bayesian representation learning limit, that exhibits representation learning mirroring that in finite-width models, yet at the same time, retains some of the simplicity of standard infinite-width limits. In particular, we show that Deep Gaussian processes (DGPs) in the Bayesian representation learning limit have exactly multivariate Gaussian posteriors, and the posterior covariances can be obtained by optimizing an interpretable objective combining a log-likelihood to improve performance with a series of KL-divergences which keep the posteriors close to the prior. We confirm these results experimentally in wide but finite DGPs. Next, we introduce the possibility of using this limit and objective as a flexible, deep generalisation of kernel methods, that we call deep kernel machines (DKMs). Like most naive kernel methods, DKMs scale cubically in the number of datapoints. We therefore use methods from the Gaussian process inducing point literature to develop a sparse DKM that scales linearly in the number of datapoints. Finally, we extend these approaches to NNs (which have non-Gaussian posteriors) in the Appendices.
Dylan Cope, Nandi Schoots
We introduce two methods for improving the performance of agents meeting for the first time to accomplish a communicative task. The methods are: (1) `message mutation' during the generation of the communication protocol; and (2) random permutations of the communication channel. These proposals are tested using a simple two-player game involving a `teacher' who generates a communication protocol and sends a message, and a `student' who interprets the message. After training multiple agents via self-play we analyse the performance of these agents when they are matched with a stranger, i.e. their zero-shot communication performance. We find that both message mutation and channel permutation positively influence performance, and we discuss their effects.