Shruti Joshi

Arnav Kumar Jain, Shivakanth Sujit, Shruti Joshi et al. · deepmind, mila

Learning world models from their sensory inputs enables agents to plan for actions by imagining their future outcomes. World models have previously been shown to improve sample-efficiency in simulated environments with few objects, but have not yet been applied successfully to environments with many objects. In environments with many objects, often only a small number of them are moving or interacting at the same time. In this paper, we investigate integrating this inductive bias of sparse interactions into the latent dynamics of world models trained from pixels. First, we introduce Variational Sparse Gating (VSG), a latent dynamics model that updates its feature dimensions sparsely through stochastic binary gates. Moreover, we propose a simplified architecture Simple Variational Sparse Gating (SVSG) that removes the deterministic pathway of previous models, resulting in a fully stochastic transition function that leverages the VSG mechanism. We evaluate the two model architectures in the BringBackShapes (BBS) environment that features a large number of moving objects and partial observability, demonstrating clear improvements over prior models.

Vitória Barin Pacela, Shruti Joshi, Isabela Camacho et al.

The linear representation hypothesis states that neural network activations encode high-level concepts as linear mixtures. However, under superposition, this encoding is a projection from a higher-dimensional concept space into a lower-dimensional activation space, and a linear decision boundary in the concept space need not remain linear after projection. In this setting, classical sparse coding methods with per-sample iterative inference leverage compressed sensing guarantees to recover latent factors. Sparse autoencoders (SAEs), on the other hand, amortise sparse inference into a fixed encoder, introducing a systematic gap. We show this amortisation gap persists across training set sizes, latent dimensions, and sparsity levels, causing SAEs to fail under out-of-distribution (OOD) compositional shifts. Through controlled experiments that decompose the failure, we identify dictionary learning -- not the inference procedure -- as the binding constraint: SAE-learned dictionaries point in substantially wrong directions, and replacing the encoder with per-sample FISTA on the same dictionary does not close the gap. An oracle baseline proves the problem is solvable with a good dictionary at all scales tested. Our results reframe the SAE failure as a dictionary learning challenge, not an amortisation problem, and point to scalable dictionary learning as the key open problem for sparse inference under superposition.

Shruti Joshi, Aaron Mueller, David Klindt et al.

Interpretability research on large language models (LLMs) has yielded important insights into model behaviour, yet recurring pitfalls persist: findings that do not generalise, and causal interpretations that outrun the evidence. Our position is that causal inference specifies what constitutes a valid mapping from model activations to invariant high-level structures, the data or assumptions needed to achieve it, and the inferences it can support. Specifically, Pearl's causal hierarchy clarifies what an interpretability study can justify. Observations establish associations between model behaviour and internal components. Interventions (e.g., ablations or activation patching) support claims how these edits affect a behavioural metric (\eg, average change in token probabilities) over a set of prompts. However, counterfactual claims -- i.e., asking what the model output would have been for the same prompt under an unobserved intervention -- remain largely unverifiable without controlled supervision. We show how causal representation learning (CRL) operationalises this hierarchy, specifying which variables are recoverable from activations and under what assumptions. Together, these motivate a diagnostic framework that helps practitioners select methods and evaluations matching claims to evidence such that findings generalise.

Aaron Mueller, Andrew Lee, Shruti Joshi et al.

A central goal of interpretability is to recover representations of causally relevant concepts from the activations of neural networks. The quality of these concept representations is typically evaluated in isolation, and under implicit independence assumptions that may not hold in practice. Thus, it is unclear whether common featurization methods - including sparse autoencoders (SAEs) and sparse probes - recover disentangled representations of these concepts. This study proposes a multi-concept evaluation setting where we control the correlations between textual concepts, such as sentiment, domain, and tense, and analyze performance under increasing correlations between them. We first evaluate the extent to which featurizers can learn disentangled representations of each concept under increasing correlational strengths. We observe a one-to-many relationship from concepts to features: features correspond to no more than one concept, but concepts are distributed across many features. Then, we perform steering experiments, measuring whether each concept is independently manipulable. Even when trained on uniform distributions of concepts, SAE features generally affect many concepts when steered, indicating that they are neither selective nor independent; nonetheless, features affect disjoint subspaces. These results suggest that correlational metrics for measuring disentanglement are generally not sufficient for establishing independence when steering, and that affecting disjoint subspaces is not sufficient for concept selectivity. These results underscore the importance of compositional evaluations in interpretability research.

Manuel Wüthrich, Felix Widmaier, Felix Grimminger et al.

Dexterous object manipulation remains an open problem in robotics, despite the rapid progress in machine learning during the past decade. We argue that a hindrance is the high cost of experimentation on real systems, in terms of both time and money. We address this problem by proposing an open-source robotic platform which can safely operate without human supervision. The hardware is inexpensive (about \SI{5000}[\$]{}) yet highly dynamic, robust, and capable of complex interaction with external objects. The software operates at 1-kilohertz and performs safety checks to prevent the hardware from breaking. The easy-to-use front-end (in C++ and Python) is suitable for real-time control as well as deep reinforcement learning. In addition, the software framework is largely robot-agnostic and can hence be used independently of the hardware proposed herein. Finally, we illustrate the potential of the proposed platform through a number of experiments, including real-time optimal control, deep reinforcement learning from scratch, throwing, and writing.

Shruti Joshi, Andrea Dittadi, Sébastien Lachapelle et al.

Steering methods manipulate the representations of large language models (LLMs) to induce responses that have desired properties, e.g., truthfulness, offering a promising approach for LLM alignment without the need for fine-tuning. Traditionally, steering has relied on supervision, such as from contrastive pairs of prompts that vary in a single target concept, which is costly to obtain and limits the speed of steering research. An appealing alternative is to use unsupervised approaches such as sparse autoencoders (SAEs) to map LLM embeddings to sparse representations that capture human-interpretable concepts. However, without further assumptions, SAEs may not be identifiable: they could learn latent dimensions that entangle multiple concepts, leading to unintentional steering of unrelated properties. We introduce Sparse Shift Autoencoders (SSAEs) that instead map the differences between embeddings to sparse representations. Crucially, we show that SSAEs are identifiable from paired observations that vary in \textit{multiple unknown concepts}, leading to accurate steering of single concepts without the need for supervision. We empirically demonstrate accurate steering across semi-synthetic and real-world language datasets using Llama-3.1 embeddings.

Jean Erik Delanois, Shruti Joshi, Ryan Golden et al.

Artificial Neural Networks (ANNs) are increasingly deployed across diverse real-world settings, where they must operate under data distributions that differ from those seen during training. This challenge is central to Domain Generalization (DG), which trains models to generalize to unseen domains without target data, and Test-Time Adaptation (TTA), which improves robustness by adapting to unlabeled test data at deployment. Existing approaches to address these challenges are often complex, resource-intensive, and difficult to scale. We introduce CONTXT (Contextual augmentatiOn for Neural feaTure X Transforms), a simple and intuitive method for contextual adaptation. CONTXT modulates internal representations using simple additive and multiplicative feature transforms. Within a TTA setting, it yields consistent gains across discriminative tasks (e.g., ANN/CNN classification) and generative models (e.g., LLMs). The method is lightweight, easy to integrate, and incurs minimal overhead, enabling robust performance under domain shift without added complexity. More broadly, CONTXT provides a compact way to steer information flow and neural processing without retraining.

Shruti Joshi, Aiswarya Akumalla, Seth Haney et al.

Mammalian brains handle complex reasoning by integrating information across brain regions specialized for particular sensory modalities. This enables improved robustness and generalization versus deep neural networks, which typically process one modality and are vulnerable to perturbations. While defense methods exist, they do not generalize well across perturbations. We developed a fusion model combining background and foreground features from CNNs trained on Imagenet and Places365. We tested its robustness to human-perceivable perturbations on MS COCO. The fusion model improved robustness, especially for classes with greater context variability. Our proposed solution for integrating multiple modalities provides a new approach to enhance robustness and may be complementary to existing methods.

Nasim Rahaman, Muhammad Waleed Gondal, Shruti Joshi et al.

Modern neural network architectures can leverage large amounts of data to generalize well within the training distribution. However, they are less capable of systematic generalization to data drawn from unseen but related distributions, a feat that is hypothesized to require compositional reasoning and reuse of knowledge. In this work, we present Neural Interpreters, an architecture that factorizes inference in a self-attention network as a system of modules, which we call \emph{functions}. Inputs to the model are routed through a sequence of functions in a way that is end-to-end learned. The proposed architecture can flexibly compose computation along width and depth, and lends itself well to capacity extension after training. To demonstrate the versatility of Neural Interpreters, we evaluate it in two distinct settings: image classification and visual abstract reasoning on Raven Progressive Matrices. In the former, we show that Neural Interpreters perform on par with the vision transformer using fewer parameters, while being transferrable to a new task in a sample efficient manner. In the latter, we find that Neural Interpreters are competitive with respect to the state-of-the-art in terms of systematic generalization

Stefan Bauer, Felix Widmaier, Manuel Wüthrich et al.

Dexterous manipulation remains an open problem in robotics. To coordinate efforts of the research community towards tackling this problem, we propose a shared benchmark. We designed and built robotic platforms that are hosted at MPI for Intelligent Systems and can be accessed remotely. Each platform consists of three robotic fingers that are capable of dexterous object manipulation. Users are able to control the platforms remotely by submitting code that is executed automatically, akin to a computational cluster. Using this setup, i) we host robotics competitions, where teams from anywhere in the world access our platforms to tackle challenging tasks ii) we publish the datasets collected during these competitions (consisting of hundreds of robot hours), and iii) we give researchers access to these platforms for their own projects.

Muhammad Waleed Gondal, Shruti Joshi, Nasim Rahaman et al.

Meta-learning algorithms adapt quickly to new tasks that are drawn from the same task distribution as the training tasks. The mechanism leading to fast adaptation is the conditioning of a downstream predictive model on the inferred representation of the task's underlying data generative process, or \emph{function}. This \emph{meta-representation}, which is computed from a few observed examples of the underlying function, is learned jointly with the predictive model. In this work, we study the implications of this joint training on the transferability of the meta-representations. Our goal is to learn meta-representations that are robust to noise in the data and facilitate solving a wide range of downstream tasks that share the same underlying functions. To this end, we propose a decoupled encoder-decoder approach to supervised meta-learning, where the encoder is trained with a contrastive objective to find a good representation of the underlying function. In particular, our training scheme is driven by the self-supervision signal indicating whether two sets of examples stem from the same function. Our experiments on a number of synthetic and real-world datasets show that the representations we obtain outperform strong baselines in terms of downstream performance and noise robustness, even when these baselines are trained in an end-to-end manner.

Homanga Bharadhwaj, Shruti Joshi

Recommendation systems are an integral part of Artificial Intelligence (AI) and have become increasingly important in the growing age of commercialization in AI. Deep learning (DL) techniques for recommendation systems (RS) provide powerful latent-feature models for effective recommendation but suffer from the major drawback of being non-interpretable. In this paper we describe a framework for explainable temporal recommendations in a DL model. We consider an LSTM based Recurrent Neural Network (RNN) architecture for recommendation and a neighbourhood-based scheme for generating explanations in the model. We demonstrate the effectiveness of our approach through experiments on the Netflix dataset by jointly optimizing for both prediction accuracy and explainability.