Usha Bhalla

Usha Bhalla, Suraj Srinivas, Himabindu Lakkaraju · harvard

With the increased deployment of machine learning models in various real-world applications, researchers and practitioners alike have emphasized the need for explanations of model behaviour. To this end, two broad strategies have been outlined in prior literature to explain models. Post hoc explanation methods explain the behaviour of complex black-box models by identifying features critical to model predictions; however, prior work has shown that these explanations may not be faithful, in that they incorrectly attribute high importance to features that are unimportant or non-discriminative for the underlying task. Inherently interpretable models, on the other hand, circumvent these issues by explicitly encoding explanations into model architecture, meaning their explanations are naturally faithful, but they often exhibit poor predictive performance due to their limited expressive power. In this work, we identify a key reason for the lack of faithfulness of feature attributions: the lack of robustness of the underlying black-box models, especially to the erasure of unimportant distractor features in the input. To address this issue, we propose Distractor Erasure Tuning (DiET), a method that adapts black-box models to be robust to distractor erasure, thus providing discriminative and faithful attributions. This strategy naturally combines the ease of use of post hoc explanations with the faithfulness of inherently interpretable models. We perform extensive experiments on semi-synthetic and real-world datasets and show that DiET produces models that (1) closely approximate the original black-box models they are intended to explain, and (2) yield explanations that match approximate ground truths available by construction. Our code is made public at https://github.com/AI4LIFE-GROUP/DiET.

Sheridan Feucht, Tal Haklay, Usha Bhalla et al.

Does structure in representations imply structure in computation? We study how Llama-3.1-8B reasons over cyclic concepts (e.g., "what month is six months after August?"). Even though Llama-3.1-8B's representations for these concepts are circularly structured, we find that instead of directly computing modular addition in the period of the cyclic concept (e.g., 12 for months), the model re-uses a generic addition mechanism across tasks that operates independently of concept-specific geometry. First, it computes the sum of its two inputs using base-10 addition (six + August=14). Then, it maps this sum back to cyclic concept space (14->February). We show that Llama-3.1-8B uses task-agnostic Fourier features to compute these sums--in fact, these features have periods that respect standard base-10 addition, e.g., 2, 5, and 10, rather than the cyclic concept period (e.g., 12 for months). Furthermore, we identify a sparse set of 28 MLP neurons re-used across all tasks (approximately 0.2% of the MLP at layer 18) that can be partitioned into disjoint clusters, each computing the sum for a Fourier feature with a different period. Our work highlights how an interplay between causal abstraction and feature geometry can deepen our mechanistic understanding of LMs.

Charumathi Badrinath, Usha Bhalla, Alex Oesterling et al. · harvard

Do different generative image models secretly learn similar underlying representations? We investigate this by measuring the latent space similarity of four different models: VAEs, GANs, Normalizing Flows (NFs), and Diffusion Models (DMs). Our methodology involves training linear maps between frozen latent spaces to "stitch" arbitrary pairs of encoders and decoders and measuring output-based and probe-based metrics on the resulting "stitched'' models. Our main findings are that linear maps between latent spaces of performant models preserve most visual information even when latent sizes differ; for CelebA models, gender is the most similarly represented probe-able attribute. Finally we show on an NF that latent space representations converge early in training.

Roy Rinberg, Usha Bhalla, Igor Shilov et al.

Targeted interventions on language models, such as unlearning, debiasing, or model editing, are a central method for refining model behavior and keeping knowledge up to date. While these interventions aim to modify specific information within models (e.g., removing virology content), their effects often propagate to related but unintended areas (e.g., allergies); these side-effects are commonly referred to as the ripple effect. In this work, we present RippleBench-Maker, an automatic tool for generating Q&A datasets that allow for the measurement of ripple effects in any model-editing task. RippleBench-Maker builds on a Wikipedia-based RAG pipeline (WikiRAG) to generate multiple-choice questions at varying semantic distances from the target concept (e.g., the knowledge being unlearned). Using this framework, we construct RippleBench-Bio, a benchmark derived from the WMDP (Weapons of Mass Destruction Paper) dataset, a common unlearning benchmark. We evaluate eight state-of-the-art unlearning methods and find that all exhibit non-trivial accuracy drops on topics increasingly distant from the unlearned knowledge, each with distinct propagation profiles. To support ongoing research, we release our codebase for on-the-fly ripple evaluation, along with the benchmark, RippleBench-Bio.

Usha Bhalla, Alex Oesterling, Claudio Mayrink Verdun et al.

Translating the internal representations and computations of models into concepts that humans can understand is a key goal of interpretability. While recent dictionary learning methods such as Sparse Autoencoders (SAEs) provide a promising route to discover human-interpretable features, they suffer from a variety of problems, including a systematic failure to capture the rich conceptual information that drives linguistic understanding. Instead, they exhibit a bias towards shallow, token-specific, or noisy features, such as "the phrase 'The' at the start of sentences". In this work, we propose that this is due to a fundamental issue with how dictionary learning methods for LLMs are trained. Language itself has a rich, well-studied structure spanning syntax, semantics, and pragmatics; however, current unsupervised methods largely ignore this linguistic knowledge, leading to poor feature discovery that favors superficial patterns over meaningful concepts. We focus on a simple but important aspect of language: semantic content has long-range dependencies and tends to be smooth over a sequence, whereas syntactic information is much more local. Building on this insight, we introduce Temporal Sparse Autoencoders (T-SAEs), which incorporate a novel contrastive loss encouraging consistent activations of high-level features over adjacent tokens. This simple yet powerful modification enables SAEs to disentangle semantic from syntactic features in a self-supervised manner. Across multiple datasets and models, T-SAEs recover smoother, more coherent semantic concepts without sacrificing reconstruction quality. Strikingly, they exhibit clear semantic structure despite being trained without explicit semantic signal, offering a new pathway for unsupervised interpretability in language models.

Alex Oesterling, Usha Bhalla, Suresh Venkatasubramanian et al.

As Artificial Intelligence (AI) tools are increasingly employed in diverse real-world applications, there has been significant interest in regulating these tools. To this end, several regulatory frameworks have been introduced by different countries worldwide. For example, the European Union recently passed the AI Act, the White House issued an Executive Order on safe, secure, and trustworthy AI, and the White House Office of Science and Technology Policy issued the Blueprint for an AI Bill of Rights (AI BoR). Many of these frameworks emphasize the need for auditing and improving the trustworthiness of AI tools, underscoring the importance of safety, privacy, explainability, fairness, and human fallback options. Although these regulatory frameworks highlight the necessity of enforcement, practitioners often lack detailed guidance on implementing them. Furthermore, the extensive research on operationalizing each of these aspects is frequently buried in technical papers that are difficult for practitioners to parse. In this write-up, we address this shortcoming by providing an accessible overview of existing literature related to operationalizing regulatory principles. We provide easy-to-understand summaries of state-of-the-art literature and highlight various gaps that exist between regulatory guidelines and existing AI research, including the trade-offs that emerge during operationalization. We hope that this work not only serves as a starting point for practitioners interested in learning more about operationalizing the regulatory guidelines outlined in the Blueprint for an AI BoR but also provides researchers with a list of critical open problems and gaps between regulations and state-of-the-art AI research. Finally, we note that this is a working paper and we invite feedback in line with the purpose of this document as described in the introduction.

Daniel Wurgaft, Can Rager, Matthew Kowal et al.

Neural representations carry rich geometric structure; but does that structure causally shape behavior? To address this question, we intervene along paths through activation space defined by different geometries, and measure the behavioral trajectories they induce. In particular, we test whether interventions that respect the geometry of activation space will yield behaviors close to those the model exhibits naturally. Concretely, we first fit an activation manifold $M_h$ to representations and a behavior manifold $M_y$ to output probability distributions. We then test the link $M_h \leftrightarrow M_y$ via interventions: we find that steering along $M_h$, which we term manifold steering, yields behavioral trajectories that follow $M_y$, while linear steering -- which assumes a Euclidean geometry -- cuts through off-manifold regions and hence produces unnatural outputs. Moreover, optimizing interventions in activation space to produce paths along $M_y$ recovers activation trajectories that trace the curvature of $M_h$. We demonstrate this bidirectional relationship between the geometry of representation and behavior across tasks and modalities. In language models, we use reasoning tasks with cyclic and sequential geometries as well as in-context learning tasks with more complex graph geometries. In a video world model, we use a task with geometry corresponding to physical dynamics. Overall, our work shows that geometry in neural representation is not merely incidental, but is in fact the proper object for enabling principled control via intervention on internals. This recasts the core problem of steering from finding the right direction to finding the right geometry.

Usha Bhalla, Alex Oesterling, Suraj Srinivas et al. · harvard

CLIP embeddings have demonstrated remarkable performance across a wide range of multimodal applications. However, these high-dimensional, dense vector representations are not easily interpretable, limiting our understanding of the rich structure of CLIP and its use in downstream applications that require transparency. In this work, we show that the semantic structure of CLIP's latent space can be leveraged to provide interpretability, allowing for the decomposition of representations into semantic concepts. We formulate this problem as one of sparse recovery and propose a novel method, Sparse Linear Concept Embeddings, for transforming CLIP representations into sparse linear combinations of human-interpretable concepts. Distinct from previous work, SpLiCE is task-agnostic and can be used, without training, to explain and even replace traditional dense CLIP representations, maintaining high downstream performance while significantly improving their interpretability. We also demonstrate significant use cases of SpLiCE representations including detecting spurious correlations and model editing.

Usha Bhalla, Thomas Fel, Can Rager et al.

Sparse autoencoders (SAEs) are widely used to extract interpretable features from neural network representations, often under the implicit assumption that concepts correspond to independent linear directions. However, a growing body of evidence suggests that many concepts are instead organized along low-dimensional manifolds encoding continuous geometric relationships. This raises three basic questions: what does it mean for an SAE to capture a manifold, when do existing SAE architectures do so, and how? We develop a theoretical framework that answers these questions and show that SAEs can capture manifolds in two fundamentally different ways: globally, by allocating a compact group of atoms whose linear span contains the entire manifold, or locally, by distributing it across features that each selectively tile a restricted region of the underlying geometry. Empirically, we find that SAEs suboptimally recover continuous structures, mixing the global subspace and local tiling solutions in a fragmented regime we call dilution. This explains why manifold structure is rarely visible at the level of individual concepts and motivates post-hoc unsupervised discovery methods that search for coherent groups of atoms rather than isolated directions. More broadly, our results suggest that future representation learning methods should treat geometric objects, not just individual directions, as the basic units of interpretability.

Usha Bhalla, Suraj Srinivas, Asma Ghandeharioun et al. · harvard

With the growing complexity and capability of large language models, a need to understand model reasoning has emerged, often motivated by an underlying goal of controlling and aligning models. While numerous interpretability and steering methods have been proposed as solutions, they are typically designed either for understanding or for control, seldom addressing both. Additionally, the lack of standardized applications, motivations, and evaluation metrics makes it difficult to assess methods' practical utility and efficacy. To address the aforementioned issues, we argue that intervention is a fundamental goal of interpretability and introduce success criteria to evaluate how well methods can control model behavior through interventions. To evaluate existing methods for this ability, we unify and extend four popular interpretability methods-sparse autoencoders, logit lens, tuned lens, and probing-into an abstract encoder-decoder framework, enabling interventions on interpretable features that can be mapped back to latent representations to control model outputs. We introduce two new evaluation metrics: intervention success rate and coherence-intervention tradeoff, designed to measure the accuracy of explanations and their utility in controlling model behavior. Our findings reveal that (1) while current methods allow for intervention, their effectiveness is inconsistent across features and models, (2) lens-based methods outperform SAEs and probes in achieving simple, concrete interventions, and (3) mechanistic interventions often compromise model coherence, underperforming simpler alternatives, such as prompting, and highlighting a critical shortcoming of current interpretability approaches in applications requiring control.

Aaron J. Li, Suraj Srinivas, Usha Bhalla et al. · harvard

Sparse autoencoders (SAEs) are commonly used to interpret the internal activations of large language models (LLMs) by mapping them to human-interpretable concept representations. While existing evaluations of SAEs focus on metrics such as the reconstruction-sparsity tradeoff, human (auto-)interpretability, and feature disentanglement, they overlook a critical aspect: the robustness of concept representations to input perturbations. We argue that robustness must be a fundamental consideration for concept representations, reflecting the fidelity of concept labeling. To this end, we formulate robustness quantification as input-space optimization problems and develop a comprehensive evaluation framework featuring realistic scenarios in which adversarial perturbations are crafted to manipulate SAE representations. Empirically, we find that tiny adversarial input perturbations can effectively manipulate concept-based interpretations in most scenarios without notably affecting the outputs of the base LLMs themselves. Overall, our results suggest that SAE concept representations are fragile and may be ill-suited for applications in model monitoring and oversight.

Shichang Zhang, Tessa Han, Usha Bhalla et al.

The increasing complexity of AI systems has made understanding their behavior critical. Numerous interpretability methods have been developed to attribute model behavior to three key aspects: input features, training data, and internal model components, which emerged from explainable AI, data-centric AI, and mechanistic interpretability, respectively. However, these attribution methods are studied and applied rather independently, resulting in a fragmented landscape of methods and terminology. This position paper argues that feature, data, and component attribution methods share fundamental similarities, and a unified view of them benefits both interpretability and broader AI research. To this end, we first analyze popular methods for these three types of attributions and present a unified view demonstrating that these seemingly distinct methods employ similar techniques (such as perturbations, gradients, and linear approximations) over different aspects and thus differ primarily in their perspectives rather than techniques. Then, we demonstrate how this unified view enhances understanding of existing attribution methods, highlights shared concepts and evaluation criteria among these methods, and leads to new research directions both in interpretability research, by addressing common challenges and facilitating cross-attribution innovation, and in AI more broadly, with applications in model editing, steering, and regulation.

Tian Yun, Usha Bhalla, Ellie Pavlick et al.

Vision-language (VL) pretrained models have achieved impressive performance on multimodal reasoning and zero-shot recognition tasks. Many of these VL models are pretrained on unlabeled image and caption pairs from the internet. In this paper, we study whether representations of primitive concepts--such as colors, shapes, or the attributes of object parts--emerge automatically within these pretrained VL models. We propose a two-step framework, Compositional Concept Mapping (CompMap), to investigate this. CompMap first asks a VL model to generate primitive concept activations with text prompts, and then learns to construct a composition model that maps the primitive concept activations (e.g. the likelihood of black tail or red wing) to composite concepts (e.g. a red-winged blackbird). We show that a composition model can be reliably learn from ground truth primitive concepts. We thus hypothesize that if primitive concepts indeed emerge in a VL pretrained model, its primitive concept activations can be used to learn a composition model similar to the one designed by experts. We propose a quantitative metric to measure the degree of similarity, and refer to the metric as the interpretability metric. We also measure the classification accuracy when using the primitive concept activations and the learned composition model to predict the composite concepts, and refer to it as the usefulness metric. Our study reveals that state-of-the-art VL pretrained models learn primitive concepts that are highly useful for fine-grained visual recognition on the CUB dataset, and compositional generalization tasks on the MIT-States dataset. However, we observe that the learned composition models have low interpretability in our qualitative analyses. Our results reveal the limitations of existing VL models, and the necessity of pretraining objectives that encourage the acquisition of primitive concepts.