Linear Representation Hypothesis

This paper studies the problem of learning disentangled representations of objects and their attributes from raw, unstructured image data. Slot-based methods have shown considerable success in unsupervised learning of object representations from images. Block-slot attention-based methods extend this framework to attribute representations by assuming a uniform factorization of object representations into attributes, which may be suboptimal and consequently limit the quality of the learned representations. We therefore investigate a framework for jointly discovering object and attribute representations. Our key contribution is leveraging the Linear Representation Hypothesis (LRH), which postulates that composable concepts can be represented as linearly additive subspaces in slot representations. Based on this insight, we propose a probabilistic model connecting images, slots (objects), and blocks (attributes). We present an architecture that leverages block attention to connect attribute representations to slots and incorporates LRH in both object and attribute representation spaces. This architecture effectively optimizes the Evidence Lower Bound (ELBO) of the proposed graphical model. Our experiments demonstrate (i) effective discovery of disentangled object and attribute representations, (ii) empirical evidence for LRH in slot space, and (iii) the ability to perform image editing owing to the disentangled and interpretable nature of the learned representations. Our experiments on multiple datasets demonstrate improvements in DCI scores over state-of-the-art methods.

Vision Sparse Autoencoders (SAEs) have become a popular tool in Mechanistic Interpretability due to their presumed ability to disentangle complex features learned by a model into monosemantic concepts. Despite their growing popularity, evaluating their interpretability remains an active topic of research. The bedrock motivating the adoption of SAEs is the Linear Representation Hypothesis (LRH), which claims that polysemantic features can be projected onto a (near) orthogonal basis of sparse, human-understandable representations. Yet, most current frameworks evaluate proxies such as the sparsity of SAE features or the coherence of the inferred dictionary, implicitly assuming that these reflect alignment with human perception. In this paper, we provide empirical evidence that measuring the interpretability of SAE concepts is more difficult than these proxies suggest. To this end, we adapt the Autointerpretability Score (AIS) - previously shown to align with human judgments in Natural Language Processing - to vision tasks and validate our approach in a dedicated user study. We evaluate SAE concept quality using both standard metrics and our adapted AIS. We find that established interpretability metrics for SAEs correlate neither with one another nor with AIS, indicating that no single reference-free metric, whether grounded in the LRH or not, is sufficient for verifying the interpretability of vision SAEs. We argue these findings support recent calls for more verifiable, ground-truth-anchored design and evaluation of explanation methods.

To make claims about representations that generalize beyond a particular trained model, we need to specify when two representations should count as equivalent. The Linear Representation Hypothesis is often discussed without making this equivalence explicit. Different notions of equivalence preserve different structures, so metrics, probes, and interventions that appear to study the same representation may in fact correspond to different hypotheses. We therefore argue that the Linear Representation Hypothesis is not one hypothesis but a family of claims distinguished by representation equivalence. We formalize this idea using group actions, specifying the representation object, the procedure that produces it, and the property ultimately asserted, while accounting for equivalences imposed by the model architecture. This framework clarifies how assumptions can change across metrics, reading points, and analysis stages, and we use it to audit common representation quantities and recent interpretability analyses.

The Linear Representation Hypothesis associates high-level concepts with directions in language models, but it remains unclear how these concept-related linear structures are organized within the model. We propose the Answer-Basin Representation Hypothesis: the probability measure induced over answers by the model's continuation distribution organizes these linear structures, with its statistics represented along linear directions shared across questions. All continuations yielding the same answer form an answer basin, whose mass is their total probability. These basin masses define the pushforward probability measure over answers. We posit that concept-related linear structure emerges from differences in the answer measure rather than being determined by changes in concept labels. Experiments across models and tasks link concept-consistent effects and their reversals in probing and steering to the alignment between concept labels and the answer measure.

In artificial and biological neural networks, concepts are often encoded as consistent linear directions in representation space. In deep learning, this idea is known as the linear representation hypothesis and underpins many interpretability and control methods based on linear probes, from concept detection to activation steering. Yet while prior work has studied whether such directions should exist $\textit{after}$ training, the dynamics of how they emerge $\textit{during}$ training remain poorly understood. Here, we develop a framework to study the alignment of concept directions during training - a process we call "abstraction". In a minimal linear network setting, we obtain exact solutions for the full trajectory of abstraction. These solutions reveal key analytic principles governing abstraction: (i) data and target geometry jointly determine abstraction at the end-of-learning, (ii) abstraction improves with network depth, and (iii) initialization scale controls the maximum abstraction reached during training. Extending our theory to nonlinear networks, we analyze how the choice of nonlinearity affects abstraction dynamics: erf networks approximate the linear theory, while abstraction in ReLU networks depends less on target geometry and more on input geometry. Across both, we prove a striking attenuation law: both nonlinearities weaken abstraction in activations relative to preactivations. We find evidence for this law in open models (DINOv3, Gemma 4) and apply our theory to improve linear probe generalization in LLMs. Together, our results provide a dynamical theory of abstraction with implications for interpretability and control.

Transformer-based architectures have significantly advanced the generation of complex symbolic sequences, yet a significant gap remains in achieving fine-grained, interpretable control over discrete signal attributes. This paper investigates the mechanistic interpretability of the Multitrack Music Transformer (MMT) and proposes a framework for deterministic attribute modulation without retraining to bridge this gap via inference-time activation steering. Utilizing the Difference-in-Means (DiffMean) methodology, we isolate latent directions for signal attributes, specifically Pitch and Duration, within the residual stream. We validate the Linear Representation Hypothesis in this domain, achieving high correlation between steering magnitude and attribute shift. To address the inherent feature entanglement in multi-attribute steering, we introduce a Dual Steering framework utilizing Gram-Schmidt Orthogonalization. Experimental results demonstrate that this geometric decoupling reduces conceptual interference and signal degradation compared to naive vector addition, enabling independent deterministic control even against strong autoregressive conditioning.

Model dimension ($d_{model}$) is a fundamental hyperparameter in transformer language models, yet its role in setting the geometric limits of feature representation remains under-explored. Grounded in the Linear Representation and Superposition Hypotheses - which propose that models encode features as near-orthogonal directions in latent space - we develop a framework for estimating how many such directions a model can support. We first establish the embedding matrix as a measurable proxy for near-orthogonality constraints across the latent space: the boundary between meaningful token relationships and incidental similarity in the pairwise cosine similarity distribution gives a concrete estimate of the model's accepted deviation $\varepsilon$ from perfect orthogonality. Applying this metric across dozens of open-source models reveals two classes: models with high $\varepsilon$ whose embeddings lack near-orthogonal structure, and models with low $\varepsilon$ that maintain it. We then show that the standard Johnson-Lindenstrauss lemma greatly underestimates the packing efficiency of trained representations, and derive an adjusted capacity formula in which the number of near-orthogonal directions depends on the ratio of vectors to dimensions ($k/d$) rather than the raw count - a single modification that cuts prediction error by two orders of magnitude with no extra parameters. Combining these results, we define representational capacity as an upper bound on the number of distinguishable directions available for features and embeddings in a model's latent space. Capacity is exponentially sensitive to $\varepsilon$, and larger models favor tighter orthogonality constraints over maximizing raw capacity - a pattern compatible with several explanations (a stability-capacity trade-off, a ceiling on usable concepts, or confounds with model scale) that we leave to future work.

We investigate the Platonic Representation Hypothesis (PRH) through a tripartite statistical framework of representations: signal, bias, and noise. {1) Signal:} We propose that Platonic alignment arises from the universal relationship between objects and attributes, which is encoded linearly in representations according to the Linear Representation Hypothesis (LRH). We provide evidence that LRH helps explain PRH by extracting linear object-attribute features with sparse autoencoders and showing that these sparse representations often exhibit stronger cross-modal alignment than their dense counterparts. {2) Bias:} Models have different implicit biases due to the diverse architectures and training procedures used. We show that this difference can be partially mitigated. Centering and normalization consistently improve cross-model alignment. {3) Noise:} Finite-sample training leads to noise in representations. We provide evidence that representational noise is driven by data scarcity by revealing a strong and consistent positive correlation between word frequency and alignment in LLMs and text embedding models. Synthesizing signal, bias, and noise, we propose a statistical model that refines the Linear Representation Hypothesis and explains further phenomena related to the alignment of representations emerging from diverse modern AI architectures.

While researchers are finding concepts represented as linear directions in language models, a bag of linear directions fails to capture relational structure. To better understand this dichotomy, we study a model with known linear representations, but trained in a highly structured domain -- the board game Othello. While the model's internal board-state representation is linearly decodable, we find additional structure in the form of tensor product representations (TPRs). We train TPR probes to recover shared structure amongst the linear probes, yielding a factorization into square-embeddings, color-embeddings, and a binding matrix that composes them to construct the model's board-state representation. We find geometric signatures within the weights of our TPR probe that align with the structure of the board, but perhaps more importantly, that the linear probes can be recovered directly from the parameters of our TPR probe. Our findings suggest that directional representations may be projections of more structured underlying representations.

Steering is a widely used technique for controlling large language models, yet its effects are often unstable and hard to predict. Existing theoretical accounts are largely based on the Linear Representation Hypothesis (LRH). While LRH assumes that concepts can be orthogonalized for lossless control, this idealized mapping fails in real representations and cannot account for the observed unpredictability of steering. By relaxing LRH's orthogonality assumption while preserving linear representations, we show that overlapping concept contributions naturally yield a sample-specific axis-orthogonal structure. We formalize this as the Cylindrical Representation Hypothesis (CRH). In CRH, a central axis captures the main difference between concept absence and presence and drives concept generation. A surrounding normal plane controls steering sensitivity by determining how easily the axis can activate the target concept. Within this plane, only specific sensitive sectors strongly facilitate concept activation, while other sectors can suppress or delay it. While the surrounding normal plane can be reliably identified from difference vectors, the sensitive sector cannot, introducing intrinsic uncertainty at the sector level. This uncertainty provides a principled explanation for why steering outcomes often fluctuate even when using well-aligned directions. Our experiments verify the existence of the cylindrical structure and demonstrate that CRH provides a valid and practical way to interpret model steering behavior in real settings: https://github.com/mbzuai-nlp/CRH.

Compositional generalization, the ability to recognize familiar parts in novel contexts, is a defining property of intelligent systems. Although modern models are trained on massive datasets, they still cover only a tiny fraction of the combinatorial space of possible inputs, raising the question of what structure representations must have to support generalization to unseen combinations. We formalize three desiderata for compositional generalization under standard training (divisibility, transferability, stability) and show they impose necessary geometric constraints: representations must decompose linearly into per-concept components, and these components must be orthogonal across concepts. This provides theoretical grounding for the Linear Representation Hypothesis: the linear structure widely observed in neural representations is a necessary consequence of compositional generalization. We further derive dimension bounds linking the number of composable concepts to the embedding geometry. Empirically, we evaluate these predictions across modern vision models (CLIP, SigLIP, DINO) and find that representations exhibit partial linear factorization with low-rank, near-orthogonal per-concept factors, and that the degree of this structure correlates with compositional generalization on unseen combinations. As models continue to scale, these conditions predict the representational geometry they may converge to. Code is available at https://github.com/oshapio/necessary-compositionality.

We consider Representation Misdirection (RM), a class of large language model (LLM) unlearning methods that achieve forgetting by redirecting the latent representations of forget-samples toward a target vector. Despite being important, the roles of the target vector used in RM, however, remain underexplored. Here, we approach and revisit RM through the lens of the Linear Representation Hypothesis. Specifically, if one can identify a one-dimensional representation corresponding to a high-level concept, the Linear Representation Hypothesis enables linear operations on this concept vector within the forget-representation space. Under this view, we hypothesize that, beyond forgetting, machine unlearning via RM elicits controllable side effect behaviors and capabilities corresponding to the high-level concept. Our hypothesis is empirically validated across a wide range of concepts and tasks, including controlling unlearned models' truthfulness, sentiment, stereotypical bias, refusal, language, and in-context learning (ICL) tasks. Our findings reveal that this phenomenon could be either a hidden risk if misused or a mechanism that can be harnessed for developing unlearned models that require stronger capabilities and controllable behaviors.