Subspace

Deep reinforcement learning (RL) agents commonly rely on high-dimensional neural representations, despite growing evidence that task-relevant value and policy structure may be intrinsically low-dimensional. In this work, we present a simple yet effective representation-level prior that inserts a fixed orthonormal projection to constrain encoder features to a low-dimensional subspace, requiring no auxiliary objectives, pretraining, or changes to the underlying RL algorithm. Under a linear realizability assumption, we prove that when the bottleneck dimension exceeds the intrinsic rank of the optimal value function in feature space, the bottleneck preserves expressivity and leaves the induced gradient dynamics unchanged up to an equivalent low-dimensional parameterization. Empirically, we find that across both single and multi-task benchmarks, baseline performance is either matched or improved once the bottleneck dimension exceeds a small task-dependent threshold; in many cases, value representations can be compressed to extremely low dimensions without loss, and the minimal sufficient dimension depends far more on environment complexity than encoder width. In addition, we analyze representation geometry and find that orthogonal bottlenecks stabilize feature norms and are associated with higher effective rank. Together, these results support a representation-space interpretation of the manifold hypothesis in reinforcement learning and position orthogonal bottlenecks as a lightweight, architecture-agnostic mechanism for shaping RL representations.

Reliable quality control (QC) of ultrasound images is essential for both real-time acquisition guidance and retrospective clinical audit, yet existing approaches rely heavily on per-plane annotations, or employ pseudo-labeling prone to systematic bias under spatial deformations inherent in clinical acquisition. We present STRIQ, a registration-driven framework that recasts annotation-free US plane quality control as a subspace-guided consistency measurement problem. Specifically, STRIQ introduces a Latent Registration Aligner (LRA) to establish hierarchical feature space correspondences between query images and variance-driven anchors, which are autonomously distilled from unlabeled data via a variance spectrum criterion to serve as structurally stable prototypes. To further disambiguate anatomical planes and mitigate negative knowledge transfer, we propose an Orthogonal Knowledge Subspace (OKS) module. The OKS decomposes plane-specific representations into mutually orthogonal subspaces, enabling fine-grained expert collaboration while preventing inter-plane interference, ensuring that the quality metric is grounded in principled subspace proximity. Extensive experiments on the in-house US4QA and public CAMUS datasets demonstrate that STRIQ achieves state-of-the-art correlation with clinical quality scores, establishing a new paradigm for annotation-free, real-time reliable ultrasound quality control. Our code is available at https://github.com/zhcz328/STRIQ.

Reservoir computing typically relies on large, randomly generated reservoirs, enabling simple, often linear readouts. Over the past two decades, most constructions have exploited the freedom to select the reservoir, constrained primarily by stability conditions based on state contraction or memory capacity. However, these designs are largely independent of the input data and learning objective, resulting in a trial-and-error methodology driven by randomness. In high dimensions, the reservoir acts as a random embedding of the input history, implicitly relying on Johnson--Lindenstrauss--type concentration phenomena to preserve information. In contrast, we develop reservoir design principles from a geometric perspective for inputs generated by deterministic dynamical systems. Rather than relying on random embeddings, we require reservoir state increments to align within a cone around an input-determined vector subspace, and prove that such a cone concentration reduces ridge-regression training error. When the cone angle is small, the variance of reservoir states concentrates in the input-determined subspace, improving conditioning of the empirical second-moment matrix and strengthening alignment between dominant covariance directions and the state-target cross-covariance. For echo state networks, we provide a constructive approach to reservoir design. The reservoir matrix is chosen so that associated Krylov-chain directions remain nearly closed within an input-determined subspace while permitting controlled mixing in its orthogonal complement. We also provide a spectral diagnostic for ridge regression training that identifies when reservoir geometry concentrates predictive information into a few dominant covariance modes and when ``spectral pollution'' inhibits forecasting. Numerical experiments demonstrate consistent performance gains over arbitrary reservoir constructions.

Jailbreak prompts can trigger harmful completions on aligned LLMs, In accordance, safety steering has been proposed: test-time activation interventions that steer jailbreak activations to trigger refusal while preserving benign utility. However, existing steering methods are fundamentally supervised and tied to a static, limited training set, whereas real jailbreaks evolve and are often out-of-distributed from the training set, leading to failures on unseen attacks. In this paper, we tackle the failure on unseen jailbreaks problem, base on unsupervised latent direction discovery. We propose a bi-level adversarial training framework for zero-shot jailbreak defense. In the inner step, we simulate diverse jail-broken activations by extrapolating from refusal-state harmful-request activations via unsupervised latent direction discovery, which expands the coverage of real jailbreak activation subspaces. In the outer step, we train a potential-induced steering field to push these adversarial jailbroken states into refusal regions while keeping benign unchanged. Across three LLMs and six classical jailbreak families, our method achieves strong defense with attack success rates mostly below 5%, and rising subspace coverage throughout training helps explain the improved generalization.

Multimodal large language models (MLLMs) need efficient mechanisms to update knowledge without degrading existing capabilities. While intrinsic multimodal knowledge editing achieves strong reliability and locality, it often exhibits limited generality, failing to propagate edits across semantically equivalent visual and linguistic variations. This issue arises from the lack of explicit semantic supervision, rigid editing scopes, and biased anchoring to individual samples in high-dimensional multimodal spaces. We address robust intrinsic multimodal knowledge editing by explicitly targeting generalization. We formalize robustness through knowledge units that group semantically equivalent multimodal inputs and define generality as consistent predictions within each unit. To expose fragile semantic regions, we introduce Latent Adversarial Robustification (LAR), which generates adversarial yet semantically coherent variants in the joint latent space. We further propose Rank-Constrained Subspace Learning (RCSL), enforcing low-rank alignment of adversarial representations at the edit layer via a singular value-based objective. Extensive analysis demonstrates the effectiveness of ASAM empirically.

Bayesian optimization is widely employed for optimizing complex black-box functions but struggles with the curse of dimensionality. Random embedding, as a dimension reduction strategy, simplifies tasks that possess the effective dimension by optimizing within a low-dimensional subspace. However, determining the effective dimension of a task in advance remains a significant challenge, which influences the selection of the subspace dimensionality and the optimization performance. Traditional methods use fixed subspace dimensions provided by experts or rely on trial and error to estimate subspace dimensions with resources consumed. To this end, this paper proposes an automated random embedding for high-dimensional Bayesian optimization with unknown effective dimension, called Dynamic Shared Embedding Bayesian Optimization (DSEBO). DSEBO starts with a low dimension and switches to a higher subspace if the solutions in the current subspace show preliminary convergence. DSEBO dynamically determines the dimension of the next subspace based on the quality of the solutions in different subspaces and shares the queried solutions with the new subspace for a better initialization. Theoretically, we derive a regret bound for DSEBO and demonstrate that DSEBO can better balance approximation and optimization errors. Extensive experiments on functions with dimensionality of varying magnitudes and real-world tasks with unknown effective dimensions reveal that, compared with state-of-the-art methods, alternating optimization across different subspaces results in significant improvements in high-dimensional optimization, both in terms of optimization regret and time.

Self-distillation bootstraps large language models (LLMs) by training on their own generations. However, existing methods either rely on external signals to curate self-generated outputs (e.g., correctness filtering, execution feedback, and reward search), which are costly and unavailable for the best-performing frontier models, or skip curation entirely and train on all raw outputs, an approach that is often domain-specific and hard to generalize. Both also share a deeper weakness that self-generated outputs entangle task-relevant capability with others, such as stylistic patterns, formatting artifacts, and model-specific errors, diluting the signal for the specific capability one aims to improve. In this paper, we propose Self-Policy Distillation (SPD), which achieves generalizable, capability selective without any external signal. Specifically, SPD extracts a low-rank capability subspace from the model's own gradients on correctness-defining tokens, projects key-value (KV) activations into this subspace during self-generation, and fine-tunes on the resulting raw outputs with standard next-token prediction loss. Through extensive experiments across code generation, mathematical reasoning, and multiple-choice QA, we show that SPD achieves up to 13% improvement over state-of-the-art self-distillation methods without external signals and up to 16% improvement over pre-trained baselines. Notably, SPD demonstrates superior generalizability, achieving 15% better performance under out-of-domain generalization settings.

Two of the most widely used methods for analysing graph data, Adjacency Spectral Embedding and Laplacian Spectral Embedding, often produce different results when applied to the same graph. Yet the structural reasons behind this disagreement remain incompletely understood. This paper provides an end-to-end account of ASE-LSE latent subspace disagreement. We first prove that the two methods produce identical latent subspaces for every embedding dimension whenever the Laplacian is a scalar multiple of the adjacency matrix, and show that this scalar relationship holds if and only if the graph is either regular or bipartite biregular. This anchor result identifies a sufficient condition for perfect agreement that pins down the floor of the disagreement spectrum and supplies the baseline for the perturbation analysis. We then prove that no maximal-disagreement graph or family of graphs exists: the disagreement is always strictly below its theoretical ceiling, and we exhibit a witness family demonstrating that no finite maximum is attainable, so the disagreement landscape has no maximiser. With both endpoints established, we derive a Regularity Departure Bound whose two terms isolate degree heterogeneity and eigengap as the primary structural factors influencing disagreement in the middle regime. Empirical validation across thousands of simulated graphs confirms the mechanisms predicted by the bound: heterogeneity pushes disagreement up, eigengap suppresses it, and their joint ratio emerges as a unified predictor of ASE-LSE disagreement, suggesting when the two embeddings can be treated as interchangeable and when they cannot.

We show that singular value decomposition of the lm_head} weight matrix of a transformer-based large language model -- requiring only five lines of PyTorch and no model inference -- reveals interpretable semantic subspaces directly from the model weights. Each left singular vector identifies the vocabulary tokens most readily selected when the hidden state aligns with the corresponding singular direction; inspecting these clusters exposes the model's training data composition and curation philosophy. Analysing GPT-OSS-120B, Gemma-2-2B, and Qwen2.5-1.5B, we find that singular value spectra and vocabulary cluster structures differ systematically across models: GPT exhibits a graduated hierarchy of functionally differentiated subspaces; Gemma is dominated by pre-nineteenth-century English orthography, forming a stepwise clustering structure that may contribute to high output controllability; and Qwen exhibits broad multilingual coverage alongside subspaces whose vocabulary the authors have determined to be ethically inappropriate for direct publication. Base-instruct comparison reveals that ethically concerning subspaces originate in pretraining and are not removed by post-training alignment. We introduce the Vocabulary Cluster Score (VCS) to quantify subspace coherence, and the Weighted Projection Score (WPS) as a static glitch token detector; applying WPS to GPT-OSS-120B recovers shokubutsu-hyakka-tsu (ID 137606), a well-known glitch token widely reported in the CJK language community, without any model inference. We propose a taxonomy of root causes for problematic vocabulary content and call for lm_head} SVD analysis to be adopted as a standard pre-release safety auditing step. Our findings further suggest directions toward SVD-guided tokenizer optimisation and more controllable LLM design.

Mechanistic interpretability aims to explain a model's behavior by identifying causally responsible internal structures. Dictionary-based explainers such as sparse autoencoders and transcoders are a primary tool, but their faithfulness under out-of-distribution (OOD) shift has received little systematic attention. We show that distribution shift rotates the subspace that the model actively uses, misaligning the explainer's dictionary trained on in-distribution (ID) activations. We formalize this misalignment as the faithfulness gap, a geometric distance between the ID dictionary and the OOD-active subspace, and show that it controls OOD faithfulness degradation. To reduce this gap, we propose the Geometry-Adaptive Explainer (GAE), which realigns the explainer's dictionary with the OOD-active subspace while preserving the original feature structure. This requires only unlabeled OOD activations and no gradient updates. We prove that GAE improves over the unadapted ID explainer, with excess loss bounded quadratically by the second-moment shift. Empirically, GAE even matches or surpasses all training-based baselines in causal faithfulness across multiple models and OOD settings.

Reinforcement learning with verifiable rewards (RLVR) has become a dominant paradigm for improving reasoning in large language models (LLMs), yet the underlying geometry of the resulting parameter trajectories remains underexplored. In this work, we demonstrate that RLVR weight trajectories are extremely low-rank and highly predictable. Specifically, we find that the majority of downstream performance gains are captured by a rank-1 approximation of the parameter deltas, where the magnitude of this projection evolves near-linearly with training steps. Motivated by this, we propose a simple and compute-efficient method RELEX (REinforcement Learning EXtrapolation), which estimates the rank-1 subspace from a short observation window and extrapolates future checkpoints via linear regression, with no learned model required. Across three models (i.e., Qwen2.5-Math-1.5B, Qwen3-4B-Base, and Qwen3-8B-Base), RELEX produces checkpoints that match or exceed RLVR performance on both in-domain and out-of-domain benchmarks, requiring as few as 15% steps of full RLVR training. Remarkably, RELEX is able to extrapolate far beyond the observation window at no training cost, predicting checkpoints up to 10-20$\times$ beyond the observed prefix with continued improvement (e.g., observe only the first 50 steps and extrapolate to 1000 steps). Our ablation analysis confirms the minimalist sufficiency of RELEX: neither increasing the subspace rank nor employing non-linear modeling yields further gains in extrapolation. Finally, we show that RELEX's success stems from a "denoising" effect: by projecting updates onto the rank-1 subspace, the model discards stochastic optimization noise that would otherwise degrade performance during extrapolation. Our code is available at https://github.com/weizhepei/RELEX.

Distributed principal component analysis (PCA) produces node-level estimates of both a mean vector and a principal subspace. Robustly aggregating these heterogeneous objects requires a relative scale between mean error and subspace error. We study a scale-calibrated median-of-means estimator for this problem using the product geometry of Euclidean space and the Grassmann manifold. A node-level PCA expansion shows that the mean component has the usual linear influence, whereas the subspace component is an eigengap-weighted covariance perturbation. We prove a local reduction showing that the proposed product-manifold median-of-means estimator is asymptotically equivalent to a scaled spatial median of node influence errors. This yields fixed-node non-Gaussian limits, growing-node Gaussian limits with finite-block bias, and an explicit scale-dependent covariance formula. We propose robust block-scale and inference-optimal calibration rules, establish high-probability median-of-means bounds, characterize factorwise bad-node influence, and prove node-bootstrap validity. Simulations and large-scale single-cell RNA-seq data show that scale calibration adapts to eigengap-driven subspace uncertainty and provides a robust distributed PCA summary.

Modern matrix completion problems often involve heterogeneous data whose rows simultaneously belong to many meta-categories, such as demographic and age groups in recommendation systems, or region and recording session labels in neural electrophysiological experiments. Standard low-rank estimators impose a single global latent geometry, which can recover average structure but may smooth away subgroup-specific variation, especially when observations are unevenly distributed across groups. We introduce Group-Aware Matrix Estimation (GAME), a convex estimator for overlapping subgroup-wise low-rank matrix estimation. GAME regularizes category-specific submatrices through overlapping nuclear-norm penalties, allowing related groups to borrow information while preserving local latent structure in a shared coordinate system. We provide finite-sample guarantees for both reconstruction error and subgroup-specific subspace recovery, showing how performance depends on sampling density, subgroup rank, and overlap structure. Experiments on synthetic, recommendation, ecological, and neuroscience datasets show that GAME is most beneficial in structured missingness regimes, where subgroup-aware regularization improves both reconstruction accuracy and latent subspace fidelity. Across these benchmarks, GAME is competitive or best among global low-rank, side-information, and modern imputation baselines, with the largest gains when subgroups exhibit distinct low-rank structure.

While deep neural networks have achieved remarkable success across a wide range of domains, their underlying mechanisms remain poorly understood, and they are often regarded as black boxes. This gap between empirical performance and theoretical understanding poses a challenge analogous to the pre-axiomatic stage of classical geometry. In this work, we introduce the Pursuit of Subspaces (PoS) hypothesis, an axiomatic framework that formulates neural network behavior through a set of geometric postulates. These axioms, together with their derived consequences, provide a unified perspective on representation, computation, and generalization in both shallow and deep architectures. We show that this framework yields geometric explanations for fundamental questions in deep learning, including representation structure, architectural mechanisms, and generalization behavior, offering a principled step toward a coherent theoretical foundation.

Vision-Language Models require efficient adaptation to continually emerging downstream tasks. While Parameter-Efficient Fine-Tuning mitigates catastrophic forgetting, assigning isolated modules per task leads to parameter explosion. Conversely, recent similarity-driven sharing mechanisms falsely equate superficial visual similarity with underlying alignment consistency. This fundamental mismatch triggers severe negative transfer between visually similar but logically distinct tasks and fails to exploit alignment reuse across visually diverse ones. We argue thatalignment sharing is fundamentally a geometric problem of overlapping optimization trajectories within shared low-rank subspaces. Grounded in this insight, we propose iGSP, a novel framework that achieves efficient adaptation via implicit gradient subspace projection. Leveraging the early convergence of MoE routers to establish the subspace basis, iGSP bifurcates the adaptation process into two phases. First, the Subspace Identification phase introduces candidate experts via basis pre-expansion, applies a novel subspace-constrained regularization to implicitly project new task gradients onto the historical subspace, and precisely prunes redundant dimensions by treating routing probabilities as gradient flow indicators, ultimately to maximize knowledge reuse. Second, the Orthogonal Subspace Fine-Tuning phase fixes this structural basis and removes the regularization to rapidly fit the task-specific residual loss. Extensive experiments on the MTIL benchmark demonstrate that iGSP achieves state-of-the-art accuracy while significantly improving training efficiency, reducing the average trainable parameters by 42.7% compared to current SOTA methods, and decreasing the final total parameters by 86.9% relative to counterparts. The source code is available at https://github.com/GeoX-Lab/iGSP.

Online model editing for multimodal large language models (MLLMs) requires assimilating a stream of corrections under tight compute and memory budgets. Yet editors developed for text-only LLMs often degrade on MLLMs: visually dominant activations skew the statistics that shape updates, causing cross-modal conflict, while sequential writes become entangled in a shared edit space and amplify long-horizon interference, causing inter-edit interference. To address these, we propose M-ORE, a modality-decoupled online recursive editor for lifelong MLLM adaptation. M-ORE is derived from a unified proximal-projection formulation and admits a closed-form update with a Sherman-Morrison recursion, yielding constant per-edit overhead. It maintains module-wise locality statistics for the text stack and the visual projector to avoid visually dominated update shaping and performs continual updates in a fixed orthogonal low-rank edit subspace via a Sherman-Morrison recursion to mitigate long-horizon interference. Experiments on multiple MLLM backbones and online editing benchmarks show that our M-ORE method consistently improves reliability, generality, and locality over strong baselines, while achieving favorable quality-efficiency scaling. Our code is publicly available at https://github.com/lab-klc/M-ORE.

Many bandit deployments (recommendation, clinical dosing, ad targeting) share two facts prior work handles only in isolation: rewards live on a low-dimensional latent subspace, and that subspace drifts. Stationary low-rank bandits exploit rank but break under subspace change; non-stationary linear bandits adapt to drift but pay ambient rate $\widetilde{O}(d\sqrt{T})$. We study piecewise-stationary low-rank linear contextual bandits with scalar feedback: $θ_t = B_k^\star w_t$ with rank-$r$ factor $B_k^\star\in\mathbb{R}^{d\times r}$ constant within each of $K$ unknown segments and able to shift at boundaries. Our results are tight along three axes. (i) Identification boundary. With single-play scalar rewards, the moving subspace is recoverable through quadratic functionals of rewards iff three probe-side conditions hold: known noise variance, bounded state-noise coupling, and full-dimensional probe support. Each is necessary in the unrestricted-second-moment problem, and jointly they are sufficient, characterizing the boundary of the solvable region. (ii) Algorithm and dynamic regret. SPSC interleaves isotropic probes with windowed projected ridge-UCB exploitation inside the learned $r$-dimensional subspace; a CUSUM-style variant discovers segment boundaries online. The costed dynamic regret is $\widetilde{O}(r\sqrt{T})+\widetilde{O}(T^{2/3})+O(W\,V_{\mathrm{in}})$, replacing the ambient $d\sqrt{T}$ rate with the intrinsic rank. (iii) Empirics. On eleven benchmarks spanning synthetic, UCI/MovieLens, semi-synthetic clinical, and ZOZOTOWN production-log data, SPSC outperforms non-stationary and low-rank baselines whenever $d-r\gtrsim T^{1/6}$, matching the analytical crossover. To our knowledge, this is the first work to characterize the identification boundary and attain the intrinsic-rank dynamic-regret rate in this setting.

We study exact constant collapse in variational autoencoders: the deterministic encoder mean becomes independent of the input. The prior remains the standard Gaussian. Before VAE training, we select a fixed teacher posterior from a GMM-based view of the data and attach a fixed latent-only simplex witness to the encoder mean. This construction yields two linked objects. The first is a certificate: if the witness prediction improves on the best constant predictor of the teacher, the encoder mean cannot be input-independent constant. The second is a local escape direction: on the collapsed manifold, the teacher residual gives a sample-dependent descent direction for the alignment loss. For any full-support teacher posterior, the same geometry also gives a closed-form latent code with zero teacher-witness alignment error. Its scaled versions trace a margin-energy path from the constant predictor to the exact teacher code, which quantifies non-collapse inside the protected witness subspace. We instantiate the method on MNIST, CIFAR-10, and CIFAR-100. With searched unsupervised PCA-GMM teachers, vanilla VAEs fail the teacher-witness certificate in all five seeds on CIFAR-10 and CIFAR-100, while RST variants pass in all five seeds. Under collapse-stress settings with $β_{\mathrm{KL}}\in\{2,4,8\}$, vanilla VAE again fails in all seeds, whereas RST-alpha-prefit remains certificate-positive. Escape trajectories on both natural-image datasets increase the witness margin from a low-margin initialization and exhibit nonzero teacher-induced gradient norms. The analysis is confined to exact constant collapse of the encoder mean; generation quality, decoder use, and other collapse modes remain separate questions.

Inductive matrix completion (IMC) is a variant of low-rank matrix completion that incorporates row and column side-information. In principle, it can reduce the effective dimension of the recovery problem from the ambient matrix size to the dimension of the side-information features. Existing theory, however, does not fully realize this advantage in the noisy setting: sample-efficient guarantees only apply to noiseless recovery, while noisy guarantees require sample sizes comparable to ordinary matrix completion. This paper closes this gap for noisy IMC. We analyze a nonconvex projected gradient descent algorithm with spectral initialization and prove that, under exact side-information, it achieves linear convergence and stable recovery at a sample complexity governed by the effective side-information dimension rather than the ambient matrix dimension. The key technical ingredient is a local regularity condition for the IMC loss that holds at this reduced sample size, despite the mismatch between the observation pattern and the side-information subspaces. We further extend the analysis to inexact side-information, showing that the same reduced sample complexity is preserved and that the estimation error degrades optimally with the level of subspace misspecification. Motivated by this trade-off, we also propose a penalized interpolation between IMC and ordinary matrix completion that balances sample efficiency against robustness to imperfect side-information. Simulations and experiments on the MovieLens dataset support the theoretical findings and illustrate the practical benefits of exploiting side-information in low-sample regimes.

We develop a geometric theory of projection heads in self-supervised learning by modeling the head as a trainable Riemannian metric on the backbone representation manifold. We show that linear heads perform implicit subspace whitening, while nonlinear heads adapt local metrics to satisfy the specific topological constraints of the loss, with head depth empirically dictating this capacity. Analyzing dimensional collapse, we prove that smooth nonlinear heads natively induce negative eigenvalues in the Hessian at collapsed equilibria, making them unstable. We empirically validate this by continuously tracking the optimization geometry during training, which reveals that smooth activations like Swish can generate explicit negative curvature to escape collapse, whereas linear and ReLU heads under continuous-time gradient flow cannot, relying instead on discrete-time optimization dynamics and BatchNorm. Finally, we geometrically characterize how metric degeneracy governs the information-invariance trade-off, explaining why the head must be discarded. Evaluated across contrastive and decorrelation-based objectives on foundation models, our results demonstrate that the projection head acts as a universal geometric buffer, decoupling the semantic backbone from the rigid, destructive constraints of the pretraining objective.

Vision-Language-Action (VLA) models leverage powerful perceptual priors from web-scale Vision-Language Model (VLM) pre-training, yet they remain surprisingly brittle in practice, frequently failing at simple robotic tasks. To mitigate this, we propose Contrastive Conceptor Activation Steering (COAST). COAST builds on the notion of a "conceptor", a linear operator that soft-projects data into the principal components of a target distribution. COAST uses conceptors to identify success-critical subspaces for a target robotic task from a few examples of success and failure rollouts. At inference time, it steers VLA latents into these identified success subspaces to improve task outcomes. Across three architecturally distinct neural policies (flow-matching VLA, autoregressive VLA, and Diffusion Policy), COAST improves absolute mean simulation and real-robot task success rate by over 20 and 40% respectively. The activation subspace geometry reveals that failure modes share substantial structure across tasks while success representations remain largely task-specific. When tasks share similar failure modes, this structure enables previously fitted conceptors to improve performance on new tasks without refitting. Ultimately, our results suggest that current VLAs retain substantial task-relevant knowledge in their latent representations, and that the action expert's decoding bottleneck could be mitigated by steering its residual stream toward task-relevant subspaces. COAST provides a lightweight, training-free path to unlocking these latent capabilities by steering the model towards its own "success" distributions.

In language models, what a representation encodes is determined by the geometry of its representation space: distances, not activations, carry meaning. Existing tools characterize the shape of this geometry but do not ask what that shape is organized for. We introduce Subspace PGA, a metric that tests whether a layer's distance structure aligns with the readout subspace of the unembedding matrix $W_U$ more than with random subspaces of equal size. Across seven Pythia models (70M--6.9B) and three cross-family models, intermediate geometry is significantly organized for prediction (peak $z = 9$--$24$), but the degree is scale-dependent: small models ($d \leq 1024$) progressively lose it at late layers during training -- even as loss keeps improving -- while large models ($d \geq 2048$) preserve it throughout. We trace this to a capacity trade-off: a few dominant directions migrate away from $W_U$'s readout, masking rather than destroying the predictive structure beneath, and removing them restores alignment. Neither spectral metrics nor loss curves capture this distinction. Scale thus determines not only how well a model predicts, but how its representation geometry is organized to do so.

While Local Differential Privacy (LDP) serves as a foundational primitive for distributed data collection, its stringent randomization requirements often lead to severe degradation in data representation utility. This degradation stems from the task-agnostic nature of conventional LDP mechanisms, which perturb all dimensions without accounting for their relative importance to the downstream objective. To address this issue, we propose a novel approach that mitigates noise in task-relevant subspaces of the data representation. Our method identifies task-critical subspaces via the Jacobian of a public downstream model, selectively attenuates noise along these directions, and reshapes the isotropic noise of standard LDP mechanisms into an anisotropic distribution. The resulting mechanism preserves the privacy guarantee of the underlying LDP randomizer while heterogeneously modulating the impact of noise across task directions, thereby substantially enhancing data utility. The approach is applicable to both linear and nonlinear models and can be seamlessly integrated with existing LDP mechanisms. Extensive experiments on CIFAR-10-C under brightness corruption at the highest severity level demonstrate that integrating our approach improves classification accuracy by approximately 8 percentage points for Laplace and 20 percentage points for PrivUnit variants at $ε=7.5$. The source code is available at https://github.com/ymha/jacobian-anr-ldp.

We developed a tool for detecting domain shifts, namely subtle differences in the probability distributions of datasets. We identify these shifts using an algorithm designed to detect localised density anomalies in high-dimensional feature spaces. If an anomaly is present, we then identify the feature subspace in which the anomaly is most pronounced. This allows us to trace the domain shift to a small set of features, making the shift interpretable. Moreover, we provide a protocol for compensating domain shifts by extracting, from two unlabelled datasets, subsets of samples with no detectable residual distributional difference. We validate the framework on controlled 20-dimensional benchmarks with known ground truth, recovering both broad and localized shifts together with their supporting feature subspaces. We then apply it to healthy electrocardiogram (ECG) recordings represented by 782 features. In age- and sex-matched cohort comparisons differing in measurement-device composition, the method detects device-induced shifts, extracts representative subsets enriched in the imbalanced device components, and identifies ECG features associated with the acquisition contrast. These results suggest that density-shift detection and subspace attribution provide a practical framework for uncovering hidden cohort biases before downstream modelling.

We study a prototypical situation when a learned predictor can discover useful low-dimensional structure in data, while using fewer samples than are needed for accurate prediction. Specifically, we consider the problem of recovering a multi-index polynomial $f^*(x)=h(Ux)$, with $U\in\mathbb{R}^{r\times d}$ and $r\ll d$, from finitely many data/label pairs. Importantly, the target function depends on input $x$ only through the projection onto an unknown $r$-dimensional central subspace. The algorithm we analyze is appealingly simple: fit kernel ridge regression (KRR) to the data and compute the Average Gradient Outer Product (AGOP) from the fitted predictor. Our main results show that under reasonable assumptions the top $r$-dimensional eigenspace of AGOP provably recovers the central subspace, even in regimes when the prediction error remains large. Specifically, if the target function $f^*$ has degree $p^*$, it is known that $n\asymp d^{p^*}$ samples are necessary for KRR to achieve accurate prediction. In contrast, we show that if a low degree $p$ component of $f^*$ already carries all relevant directions for prediction, subspace recovery occurs in the much lower sample regime $n\asymp d^{p+δ}$ for any $δ\in(0,1)$. Our results thus demonstrate a separation between prediction and representation, and provide an explanation for why iterative kernel methods such as Recursive Feature Machines (RFM) can be sample-efficient in practice.

Electroencephalogram (EEG) signals are highly susceptible to artifacts, resulting in a low signal-to-noise ratio which makes extraction of meaningful neural information challenging. Artifact Subspace Reconstruction (ASR) is one of the most widely used artifact filtering techniques in EEG-based BCI applications, owing to its real-time applicability. ASR reconstructs artifact-free signals by operating in Principal Component (PC) space within sliding windows. However, ASR performance is critically sensitive to its threshold parameter - an incorrect threshold risks removing task-relevant neural features alongside artifacts. Furthermore, since PCs are linear combinations of all channels, subspace reconstruction in PC space may alter the underlying data structure, potentially discarding essential neural information. To address these limitations, we propose nASR, a novel end-to-end trainable Keras layer that jointly optimizes artifact rejection and downstream decoding. nASR introduces two trainable threshold parameters: K, which governs artifact detection in PC variance space, and L, which quantifies eigen-spread to pinpoint the primary artifact--contributing channels, enabling selective channel-level reconstruction that preserves clean channel information. An ablation study comprising five model variants (m01 - m05), evaluated across two subjects from the BCI Competition IV Dataset 1, confirms that nASR variants consistently outperform traditional ASR on test classification metrics, while achieving a 6-8x reduction in inference time, making nASR a strong candidate for real-time BCI applications demanding both low latency and high decoding performance.

Despite the popularity of the actor-critic method and the practical needs of collaborative policy training, existing works typically either overlook environmental heterogeneity or give up personalization altogether by training a single shared policy across all agents. We consider a federated actor-critic framework in which agents share a common linear subspace representation while maintaining personalized local policy components, and agents iteratively estimate the common subspace, local critic heads, and local policies (i.e., actors). Under canonical single-timescale updates with Markovian sampling, we establish finite-time convergence via a novel joint linear approximation framework. Specifically, we show that the critic error converges to zero at the rate of $\tilde{\mathcal{O}}(1/((1-γ)^4\sqrt{TK}))$, and the policy gradient norm converges to zero at the rate of $\tilde{\mathcal{O}}(1/((1-γ)^6\sqrt{TK}))$, where $T$ is the number of rounds, $K$ is the number of agents, and $γ\in (0,1)$ is the discount factor. These results demonstrate linear speedup with respect to the number of agents $K$, despite heterogeneous Markovian trajectories under distinct transition kernels and coupled learning dynamics. To address these challenges, we develop a new perturbation analysis for the projected subspace updates and QR decomposition steps, together with conditional mixing arguments for heterogeneous Markovian noise. Furthermore, to handle the additional complications induced by policy updates and temporal dependence, we establish fine-grained characterizations of the discrepancies between function evaluations under Markovian sampling and under temporally frozen policies. Experiments instantiate the framework within PPO on federated \texttt{Hopper-v5} action-map heterogeneity, showing gains over Single PPO and FedAvg PPO and downstream transfer from the learned shared trunk.

Complex reinforcement learning environments frequently employ multi-task and mixed-reward formulations. In these settings, heterogeneous reward distributions and correlated reward dimensions often destabilize the construction of scalar advantages. To address these challenges, we propose Reward-Decorrelated Policy Optimization (RDPO), a reward-processing method designed to explicitly target both failure modes. RDPO first utilizes Magnitude-Aware Quantile normalization to stabilize prompt-level advantage allocation across binary, fractional, and continuous rewards. It then applies Mahalanobis whitening within each active reward subspace to mitigate correlation redundancy prior to aggregation. When applied during the post-training of LongCat-Flash, RDPO enhances instruction following, writing quality, and robustness to hard prompts while remaining broadly competitive on reasoning and coding evaluations.

Diffusion models are often trained in low-dimensional latent spaces, which are then reused for related but shifted datasets. In this work, we study when such latent reuse remains reliable under distribution shift. We consider a source-target setting in which both datasets are approximately low-dimensional but may lie near different subspaces. We show that freezing and reusing a source latent space induces a target-domain score error governed by two quantities: the principal-angle misalignment between the source and target subspaces, and the target ambient noise amplified by the diffusion time scale. Motivated by these limits, we further study mixed source-target training and characterize how the required shared latent dimension depends on the relative geometry of the two distributions. Our results provide theoretical guidance on when latent reuse is reliable and when learning a shared representation may be necessary.

Regression and Bayesian accounts of in-context learning (ICL) explain how demonstrations can induce predictors, while mechanistic analyses often identify compact activation directions that steer prompted behavior. However, it remains unclear whether structured demonstrations induce low-dimensional concept inference. We study this question through a concept-subspace view of ICL, in which tasks vary only along intrinsic concept coordinates, although inputs are observed in a high-dimensional ambient space. For ridge and least-squares ICL proxies, prediction decomposes exactly into concept-coordinate regression and off-subspace leakage. Under block-diagonal or near-block-diagonal covariance assumptions, the leading estimation and nuisance-sensitivity terms scale with the dimension of the concept subspace, while residual effects are controlled by cross-subspace coupling. This separation gives a mechanistic prediction: recoverable task information should concentrate in a low-dimensional, task-aligned activation subspace. On CounterFact-derived multi-relation prompts with Llama-3-8B, a 68--73-dimensional subspace of the 4096-dimensional residual stream restores 78.8% of the clean--corrupted accuracy gap, whereas patching the complementary subspace restores 0%. Concept swaps redirect predictions toward injected relations, while random and cross-task matched-rank controls are largely ineffective. Additional experiments on Qwen2.5-7B and a controlled cross-lingual rule task show the same qualitative pattern. These results support concept subspaces as compact, task-aligned mediators of recoverable ICL behavior in structured task families, without implying full-circuit recovery.

Stochastic bilevel optimization (SBO) has become a standard framework for hyperparameter learning, data reweighting, representation learning, and data-mixture optimization in deep learning. Existing exact single-loop SBO methods and memory-efficient surrogate SBO methods either create severe memory pressure for large lower-level neural networks or lack competitive convergence guarantees under standard assumptions. In this paper, we propose BROS, a memory-efficient single-loop SBO method with the same convergence rate order as exact single-loop SBO methods. BROS performs lower and auxiliary updates in randomized subspaces with a Rademacher bi-probe correction that recovers an unbiased Hessian-action estimator. We prove that BROS preserves the $\mathcal O(\varepsilon^{-2})$ sample complexity of MA-SOBA for finding an $\varepsilon$-stationary point under only standard assumptions. Experiments on hyper-data cleaning, data-mixture learning, hyper-representation learning, and ViT sample reweighting show that BROS reduces peak memory by up to 44.9% while closely matching full-space baseline performance.

Alignment, the tendency of adjacent weight matrices in deep networks to develop compatible subspace orientations, underlies gradient flow, Neural Collapse, and representation similarity across architectures. Despite extensive empirical documentation, these phenomena have resisted unified theoretical treatment: existing explanations are post-hoc, each fitted to a specific observation with whatever mathematics is at hand. We reverse this direction by deriving the mathematical structure that layerwise alignment inherently demands. Using geometric invariant theory, we prove that alignment geometry has a canonical closed, polystable stratum given by a flag variety, and that subspace intersection dimension is its unique reparameterization-invariant observable, establishing that subspace metrics are not empirical conventions but mathematical necessities. This unified framework yields two dynamical consequences: ridge regularization drives subspace alignment at an exponential rate set by weight decay, whereas nonlinear activations induce a commutator obstruction to exact basis alignment, generically present in nonlinear networks and absent in linear ones. Together these give a geometric explanation of the Level-2/3 hierarchy in Neural Collapse from first principles rather than post-hoc analysis. The commutator magnitude and head subspace overlap further serve as weight-space windows into internal alignment structure, requiring no forward passes. Experiments on multilayer perceptrons, residual networks, and pretrained language models support the proposed diagnostics and delineate their scope.

Large language models represent the same reasoning in vastly different surface forms -- English prose, Python code, mathematical notation -- yet whether they share a common internal substrate across these symbolic systems remains unknown. We introduce the TriForm Benchmark (18 concepts x 6 forms x 3 instances = 324 stimuli) and study five LLMs (1.6B-8B) across three architecture families. Using permutation-corrected RSA, cross-form probing, and activation patching, we find converging evidence for a Format-Agnostic Reasoning Subspace (FARS) in middle layers. We make FARS concrete: concept-centroid PCA extracts a 10-dimensional subspace that amplifies concept structure 3x while suppressing form information to near zero. Replacing only these 10 dimensions during cross-form patching preserves 90-96% of model output -- far exceeding both full activation replacement (44-56%) and variance-maximizing PCA (60-74%) -- while ablating them causes targeted disruption. FARS generalizes to held-out concepts and converges across architectures (CCA > 0.79 for all model pairs), providing within-modality evidence for the Platonic Representation Hypothesis. We further discover a declarative-procedural asymmetry: representations are far more compatible between prose and mathematics than between either and code, suggesting that the critical axis of divergence is not linguistic vs. formal but declarative vs. procedural.

The advancement of Multimodal Large Language Models (MLLMs) has expanded Role-Playing Agents (RPAs) into visually grounded environments. However, human vision is inherently subjective and identity-driven, whereas existing MLLMs extract objective, character-agnostic features for general tasks. In RPAs, this generic visual noise overpowers fragile character traits, causing Modality-Role Interference (MRI), where agents struggle to integrate visual grounding and character consistency. To address this, we introduce the training-free Character-Aware Visual Intervention (CAVI) framework, enabling agents to perceive the world through the lens of character. CAVI systematically targets MRI: macroscopically, Character-Guided Token Pruning (CTP) restricts the visual receptive field to role-relevant entities; microscopically, Orthogonal Feature Modulation (OFM) projects tokens onto a character-context subspace to extract aligned facts; and during decoding, Modality-Adaptive Role Steering (MARS) dynamically optimizes steering intensity based on visual reliance. Extensive experiments show CAVI effectively alleviates MRI, significantly enhancing character-consistent multimodal interactions.

This paper investigates robust representation learning in offline goal-conditioned reinforcement learning (GCRL). Particularly in sparse reward scenarios, learning representations that align state and goal latents is a challenge that frequently culminates in representation divergence where the encoder drifts toward a low-dimensional, goal-agnostic subspace that destabilizes policy learning. We address this issue by showing that an agent must acquire a fundamental understanding of its environment across multiple scales, from local physical dynamics to long-horizon goal-directed structure. Building on this insight, we propose Ms.PR, a framework that leverages multi-scale predictive supervision to enforce goal-directed alignment within the latent space. We demonstrate that Ms.PR leads to improved representation quality and strong performance on both vision and state-based tasks. Furthermore, we show that our approach is exceptionally resilient under realistic, challenging data regimes, maintaining state-of-the-art performance across a wide variety of tasks, trajectory stitching scenarios, and extreme noise conditions.

We investigate the geometry of predictive information across the layers of large language models (LLMs). We repurpose representation lenses-learned affine maps trained to predict the next token from intermediate residual streams-as geometric diagnostic tools. Rather than asking what the model predicts at each layer, we ask where predictive information resides and how it evolves across depth. We define at each layer a predictive readout subspace as the dominant k-dimensional singular subspace of such a map on the d-dimensional residual stream (where k is a resolution parameter), and track its trajectory on the Grassmann manifold as a similarity profile across layers. The profile is well described by unimodal distributions exhibiting a rise, near-plateau, and descent; varying k from 1% to 50% of d traces a Pareto frontier between visibility and energy retention, yet the same structure emerges at all scales. Across eight models from two families (Qwen2.5 and OLMo2, 1B-32B), we identify three geometric phases. Updates are approximately orthogonal to the residual stream throughout; what distinguishes the phases is their effect on the effective rank, which expands, stabilizes, and concentrates. In the first, Seeding Multiplexing, feed-forward memories and attention layers seed a candidate set in superposition in family-specific proportions, with the final token rising as leading candidate from 20% to 35% of positions across this phase. In the second, Hoisting Overriding, updates override existing subspaces to concentrate the candidate distribution without expanding the rank. In the third, Focal Convergence, high-energy low-rank updates write the winner into a form aligned with the unembedding direction. Phases 1 and 3 grow slowly with model depth, while Phase 2 expands linearly. The additional capacity of deeper LLMs is largely absorbed by candidate disambiguation.

Supervised fine-tuning (SFT) improves in-domain performance but can degrade out-of-domain (OOD) generalization. Prior work suggests that this degradation is related to changes in dominant singular subspaces of pretrained weight matrices. However, directly identifying loss-sensitive directions with Hessian or Fisher information is computationally expensive at LLM scale. In this work, we propose preserving projected rotations in pretrained singular subspaces as an efficient proxy for Fisher-sensitive directions, which we call Rotation-Preserving Supervised Fine-Tuning (RPSFT). RPSFT penalizes changes in the projected top-$k$ singular-vector block of each pretrained weight matrix, limiting unnecessary rotation while preserving task adaptation. Across model families and sizes trained on math reasoning data, RPSFT improves the in-domain/OOD trade-off over standard SFT and strong SFT baselines, better preserves pretrained representations, and provides stronger initializations for downstream RL fine-tuning. Code is available at \href{https://github.com/jinhangzhan/RPSFT.git}{https://github.com/jinhangzhan/RPSFT}.

Hopfield networks and their generalizations have established deep connections among biological associative memories, statistical physics, and transformers. Yet most models treat retrieval as a fixed query-to-memory mapping, ignoring the role of external context in recall. In this work, we propose a two-stage associative memory architecture, wherein a context-gate subcircuit reshapes the retrieval energy landscape before and during recall. We show theoretically that context gating increases inter-memory separation while inducing sparsity, translating into exponential improvements in retrieval. Crucially, we prove that the system admits a unique self-consistent fixed point, revealing that the resulting retrieval state is driven by both a direct contextual bias and a second-order retrieval-gate feedback loop. We then bridge this theory to transformers; specifically, we evaluate a first-order approximation on Llama-3, confirming that in-context learning acts as context-gated retrieval. Native dynamics mirror our theory: context localizes a memory subspace, enabling the zero-shot query to cleanly discriminate. Ultimately, this framework provides a mechanistic link between associative memory theory and LLM phenomenology.

Learning hierarchical features in Sparse Autoencoders (SAEs) is essential for capturing the structured nature of real-world data and mitigating issues like feature absorption or splitting. Existing works attempt to identify hierarchical relationships within independent feature sets by relying on activation coverage, the assumption that child feature should only activate when its parent feature activates. However, we demonstrate that this condition alone is insufficient; that is, it often produces false positives where parent and child concepts are semantically unrelated. To address this, we introduce a novel reconstruction condition that enforces a deeper functional link between hierarchical levels. By combining both activation and reconstruction constraints, we propose the Tree SAE, a model designed to learn hierarchical structures directly from within the feature set. Our results demonstrate that Tree SAEs significantly surpass the existing SAEs at learning hierarchical pairs while maintaining competitive performance to the state-of-the-art on several key benchmarks. Finally, we demonstrate the practical utility of our Tree SAE in mapping the geometry of child feature subspaces and uncovering the complex hierarchical concept structures encoded within large language models.

The convergence of Krylov-based linear iterative solvers applied to parametric partial differential equations (PDEs) is often highly sensitive to the domain, its discretization, the location/values of the applied Dirichlet/Neumann boundary conditions, body forces and material properties, among others. We have previously introduced hybridization of classical linear iterative solvers with neural operators for specific geometries, but they tend to not perform well on geometries not previously seen during training. We partially addressed this challenge by introducing the deep operator network Geo-DeepONet and hybridizing it with Krylov-based iterative linear solvers, which, despite learning effectively across arbitrary unstructured meshes without requiring retraining, led to only modest reductions in iterations compared to state-of-the-art preconditioners. In this study we introduce Neural Subspace Proper Orthogonal Decomposition (NSPOD), a multigrid-like deep operator network-based preconditioner which can dramatically reduce the number of iterations needed for convergence in Krylov-based linear iterative solvers, even when compared to state-of-the-art methods such as algebraic multigrid preconditioners. We demonstrate its efficiency via numerical experiments on a linearized version of solid mechanics PDEs applied to unstructured domains obtained from complex CAD geometries. We expect that the findings in this study lead to more efficient hybrid preconditioners that can match, or possibly even surpass, the convergence properties of the current gold standard preconditioning methods for solid mechanics PDEs.

Low-Rank Adaptation (LoRA) enables parameter-efficient fine-tuning of large models by decomposing weight updates into low-rank matrices, significantly reducing storage and computational overhead. While effective, standard LoRA lacks mechanisms for uncertainty quantification, leading to overconfident and poorly calibrated models. Bayesian variants of LoRA address this limitation, but at the cost of a significantly increased number of trainable parameters, partially offsetting the original efficiency gains. Additionally, these models are harder to train and may suffer from unstable convergence. In this work, we propose a novel framework for parameter-efficient Bayesian fine-tuning, demonstrating that effective uncertainty quantification can be achieved in very low-dimensional parameter spaces. The proposed method achieves strong performance with improved calibration and generalization while maintaining computational efficiency. Our empirical findings show that, with the appropriate projection of the weight space uncertainty can be effectively modeled in a low-dimensional space, and weight covariances exhibit low ranks.

Class-incremental learning aims to continuously acquire new knowledge while preserving previously learned information, thereby mitigating catastrophic forgetting. Existing methods primarily restrict parameter updates but often overlook their structural properties in high-dimensional spaces. From a subspace perspective, updates induced by different tasks tend to lie in multiple overlapping low-rank subspaces, leading to cross-task subspace interference and severe forgetting. To address this issue, we propose HDSD, a Hierarchical Dual-Subspace Decoupling framework for continual learning in vision-language models. Specifically, we introduce a lightweight Feature Modulation Module (FMM) that explicitly decomposes the parameter space into general and task-specific subspaces. Building on this design, we develop two complementary components. First, a General Fusion Module (GFM) evaluates relative parameter changes across tasks and uses an adaptive threshold to capture stable and transferable knowledge. Second, a Hierarchical Learning Module (HLM) performs structured parameter decomposition via Singular Value Decomposition (SVD) and uses a scaling mechanism to constrain updates within distinct subspace scales. Together, these designs reduce subspace interference and parameter drift. Extensive experiments on conventional benchmarks show that HDSD achieves state-of-the-art results.

Generative personalization often suffers from the semantic collapsing problem (SCP), where a learned personalized concept overpowers the rest of the text prompt, causing the model to ignore important contextual details. To address this, we first analyze the underlying cause, revealing that the semantic drift responsible for SCP is not random but is concentrated within a specific low-dimensional subspace. We also discover that the personalization process perturbs the embedding of the original base concept, making it an unstable reference point. Based on these insights, we introduce Test-time Embedding Adjustment with Adaptive Subspace Projection (AdaptSP), a training-free method that uses the stable, pre-trained embedding as an anchor. AdaptSP isolates the semantic drift and projects it onto the identified subspace, performing a precise adjustment that mitigates SCP while maintaining the subject identity. Our experiments show that this targeted approach significantly improves prompt fidelity and contextual alignment.

Transformer architecture search (TAS) discovers optimal vision transformer (ViT) architectures automatically, reducing human effort to manually design ViTs. However, existing TAS methods suffer from the feature collapse problem, where subnets within a supernet fail to learn subnet-specific features, mainly due to the shared weights in a supernet, limiting the performance of individual subnets. To address this, we propose TAS-LoRA, a novel method that introduces parameter-efficient low-rank adaptation (LoRA) to enable subnet-specific feature learning, while maintaining computational efficiency. TAS-LoRA incorporates a Mixture-of-LoRAExperts (MoLE) strategy, where a lightweight router dynamically assigns LoRA experts based on subnet architectures, and introduces a group-wise router initialization technique to encourage diverse feature learning across experts early in training. Extensive experiments on ImageNet and several transfer learning benchmarks, including CIFAR-10/100, Flowers, CARS, and INAT-19, demonstrate that TAS-LoRA mitigates feature collapse effectively, improving performance over state-of-the-art TAS methods significantly.

Data reconstruction attacks on trained neural networks aim to recover the data on which the network has been trained and pose a significant threat to privacy, especially if the training dataset contains sensitive information. Here, we propose a unified optimization formulation of the data reconstruction problem based on initial and trained parameter values, incorporating state-of-the-art proposals. We show that in the random feature model, this formulation provably leads to training data reconstruction with high probability, provided the network width is sufficiently large; this unprecedented finite-width result uses PAC-style bounds. Furthermore, when the data lies in a low-dimensional subspace, we show that the network width requirement for successful reconstruction can be relaxed, with bounds depending on the subspace dimension rather than the ambient dimension. For general neural network models and unknown data orientations, we propose an efficient reconstruction algorithm that approximates the low-dimensional data subspace through the change in the first-layer weights during training and uses only the last-layer weights for reconstruction, thus reducing the search space dimension and the required network width for high-quality reconstructions. Our numerical experiments on synthetic datasets and CIFAR-10 confirm that our subspace-aware reconstruction approach outperforms standard full-space techniques.

Heterogeneous federated learning (HtFL) aims to enable collaboration among clients that differ in both data distributions and model architectures. Prototype-based methods, which communicate class-level feature centers (prototypes) instead of full model parameters, have recently shown strong potential for HtFL. Existing prototype-based HtFL methods typically reuse the MSE-based or cosine-based alignment mechanism developed for homogeneous FL when aligning client-specific representations with global prototypes. These approaches are essentially coordinate alignment, where representations of clients are forced to match the global prototypes in the embedding space in an element-wise manner. Such alignment implicitly assumes that all clients should map their representations into the feature subspace defined by the global prototypes. This assumption is reasonable in homogeneous FL, where all clients share the same feature extractor. However, it becomes problematic in HtFL, since heterogeneous feature extractors naturally induce client-specific feature subspaces, and forcing all clients to optimize within a single global subspace unnecessarily suppresses their learning capacity. We observe that coordinate alignment implicitly couples two distinct objectives: aligning inter-class semantic structure, which is directly beneficial for classification, and enforcing a shared feature basis, which is unnecessary and even harmful under model heterogeneity. Building on this insight, we design FedSAF, which shifts the alignment objective from absolute coordinates to inter-class relational structure. We demonstrate that structural alignment consistently outperforms coordinate alignment in heterogeneous settings. Experiments on multiple benchmarks show that our structural alignment outperforms state-of-the-art prototype-based HtFL methods by up to 3.52%.

Human Motion Segmentation (HMS), which aims to partition a video into non-overlapping segments corresponding to different human motions, has recently attracted increasing research attention. Existing HMS approaches are predominantly based on subspace clustering, which are grounded on the assumption that the distribution of high-dimensional temporal features well aligns with a Union-of-Subspaces (UoS). For videos in the real world, however, the raw frame-level features often violate the UoS assumption and yield unsatisfactory segmentation performance. To address this issue, we propose an efficient and effective approach for HMS, named Temporal Deep Self-expressive subspace Clustering (TDSC), which jointly learns temporally consistent structured representations and stabilized affinity for accurate and robust HMS. Specifically, in TDSC, we alternately learn structured representations of the input frame features and self-expressive coefficients via a properly regularized self-expressive model, in which a coding-rate maximization regularizer is incorporated to avoid representation collapse and conform the learned representations to span a desired UoS distribution, and meanwhile, temporal constraints are incorporated to promote temporally adjacent frames to be partitioned into the same groups. Moreover, we develop a temporal momentum averaging mechanism to stabilize affinity evolution and design a reparameterization strategy to enable efficient optimization. We conduct extensive experiments on five benchmark HMS datasets using both conventional (HoG) and up-to-date deep features (i.e., CLIP, DINOv2) to validate the effectiveness of our approach.

Domain Generalization (DG) aims to learn representations that remain robust under out-of-distribution (OOD) shifts and generalize effectively to unseen target domains. While recent invariant learning strategies and architectural advances have achieved strong performance, explicitly discovering a structured domain-invariant subspace through second-order statistics remains underexplored. In this work, we propose CPCANet, a novel framework grounded in Common Principal Component Analysis (CPCA), which unrolls the iterative Flury-Gautschi (FG) algorithm into fully differentiable neural layers. This approach integrates the statistical properties of CPCA into an end-to-end trainable framework, enforcing the discovery of a shared subspace across diverse domains while preserving interpretability. Experiments on four standard DG benchmarks demonstrate that CPCANet achieves state-of-the-art (SOTA) performance in zero-shot transfer. Moreover, CPCANet is architecture-agnostic and requires no dataset-specific tuning, providing a simple and efficient approach to learning robust representations under distribution shift. Code is available at https://github.com/wish44165/CPCANet.

Activation-based steering provides control of LLM behavior at inference time, but the dominant paradigm reduces each concept to a single direction whose geometry is left largely unexamined. Rather than selecting a single steering direction, we use conceptors: soft projection matrices estimated from activations pooled across both poles of a bipolar concept, which preserve the concept's full multidimensional subspace. A geometric analysis shows the bipolar subspace strictly subsumes the single-vector baseline. We further show that the conceptor quota provides a parameter-free layer-selection diagnostic, predicting concept separability with Pearson correlations up to r=0.96 across three instruction-tuned models and three semantic dimensions. Beyond selection, conceptors admit a closed-form Boolean algebra (AND, OR, NOT): we evaluate conceptor compositionality on thematically related sub-concepts. Across a systematic five-axis design-space evaluation, conceptors match or outperform additive baselines at layers where concept subspaces are multi-dimensional while producing substantially fewer degenerate outputs. Conceptor steering is a geometrically principled, compositional, and practically safer alternative to single-direction steering from a limited number of contrastive pairs.

To mitigate the memory constraints associated with fine-tuning large pre-trained models, existing parameter-efficient fine-tuning (PEFT) methods, such as LoRA, rely on low-rank updates. However, such updates fail to fully capture the rank characteristics of the weight modifications observed in full-parameter fine-tuning, resulting in a performance gap. Furthermore, LoRA and other existing PEFT methods still require substantial memory to store the full set of frozen weights, limiting their efficiency in resource-constrained settings. To addres these limitations, we introduce Cumulative Energy-Retaining Subspace Adaptation (CERSA), a novel fine-tuning paradigm that leverages singular value decomposition (SVD) to retain only the principal components responsible for 90% to 95% of the spectral energy. By fine-tuning low-rank representations derived from this principal subspace, CERSA significantly reduces memory consumption. We conduct extensive evaluations of CERSA across models of varying scales and domains, including image recognition, text-to-image generation, and natural language understanding. Empirical results demonstrate that CERSA consistently outperforms state-of-the-art PEFT methods while achieving substantially lower memory requirements. The code will be publicly released.

We provide a unified theoretical analysis of Linear Discriminant Analysis with simultaneous multilabel scatter matrix formulations and Stiefel orthogonality constraints. Our contributions span both algebraic structure and statistical guarantees. On the algebraic side, we characterize the rank of the multilabel between-class scatter matrix, showing that the effective discriminant dimensionality can strictly exceed the classical single-label bound of $C-1$; we establish a multilabel partition of variance and prove that all four Fisher objectives are equivalent under the $W^\top S_t^{ML} W = I_r$ constraint while characterizing their divergence under the Stiefel constraint; and we prove a two-sided label-distance preservation bound relating projected distances to Hamming distances in label space. On the statistical side, we establish a finite-sample $O(k_{\max}\sqrt{d\log d/n}/gap_r)$ bound on the subspace estimation error under sub-Gaussian noise with a matching $Ω(σ^2 d/(n\,gap_r))$ minimax lower bound, establishing a near-minimax-optimal rate (matching up to logarithmic and $k_{\max}$ factors) for multilabel discriminant subspace estimation. We further provide high-probability distance concentration, robustness guarantees under label interactions, and a regularization analysis preserving the spectral structure when $d \gg n$. All results are verified numerically on synthetic data generated from the linear label-effect model, covering both the algebraic identities and the multilabel-specific quantities ($k_{\max}$, $κ(S_t^{ML})$, $\|Γ/n\|_2$, $Δ_r$) that govern the statistical bounds. The numerical experiments are designed as a sanity check for the theorems rather than as an empirical benchmark; evaluation on real multilabel datasets is left to future work targeting application-oriented venues.

Activation-alignment measures such as Representational Similarity Analysis (RSA), Canonical Correlation Analysis (CCA), and Centered Kernel Alignment (CKA) are widely used to compare biological and artificial neural representations. Recent theoretical work interprets many of these methods as assessing agreement between optimal linear readouts over broad families of global tasks. However, agreement at the level of global readouts does not determine how a system uses local stimulus evidence. Specifically, representations may align in activation space yet differ in their sensitivity to small perturbations. To address this challenge, we introduce a complementary framework based on local decodable information, which focuses on a representation's ability, under noise, to discriminate small perturbations within a specified stimulus-coordinate subspace. Building on Fisher information and local representation geometry, we summarize each representation using the expected projected pullback/Fisher metric over that subspace. This formulation induces a second-moment family of local discrimination tasks, for which the resulting operator provides a minimal, complete dataset-level summary of expected discriminability. We compare these regularized signatures using a log-spectral distance on the manifold of symmetric positive definite (SPD) matrices, yielding the Spectral Riemannian Alignment Score (S-RAS) and a uniform multiplicative certificate over the corresponding family of lifted task values. Empirically, this framework enables the recovery of corresponding layers across independently trained artificial neural networks, supports transferable class-conditional probes, reveals controlled dissociations between standard and robust training, and uncovers stimulus-coordinate family effects across mouse visual cortex using the Allen Brain Observatory static gratings dataset.

A method is presented for accelerating inference in transformer language models by exploiting the low effective rank of the token activation manifold at each layer. The method decomposes each activation vector into a subspace component and a residual, computes the linear-layer output on the subspace component via a cached low-rank weight image at reduced memory bandwidth, and applies a per-token gate that determines whether the residual correction is computed or skipped. The gate ensures that the output distribution is preserved to within a controllable tolerance. Validation on three model families (GPT-2 124M, GPT-J 6B, OPT 6.7B) on AMD MI300X demonstrates effective speedups of 3.0x to 10.5x on linear-layer weight reads with perplexity ratios below 1.00 and top-1 token agreement above 98%. The method requires no retraining, no architectural modification, and no approximation of the attention mechanism. At the operating point (k = 256, ε = 0.05) on GPT-J 14 6B, the accelerated model produces character-for-character identical output to the baseline.

Adapting large pretrained models to diverse tasks is now routine, yet the two dominant strategies of parameter-efficient fine-tuning (PEFT) and low-rank compression are typically composed in sequence. This decoupled practice first compresses and then fine-tunes adapters, potentially misaligning the compressed subspace with downstream objectives and squandering a global parameter budget. To overcome this limitation, we introduce JACTUS (Joint Adaptation and Compression with a Task-aware Union of Subspaces), a single framework that unifies compression and adaptation. From a small calibration set, JACTUS estimates input and pre-activation gradient covariances, forms their orthogonal union with the pretrained weight subspace, performs a projected low-rank approximation inside this union, allocates rank globally by marginal gain per parameter, and trains only a compact core matrix. This explicitly mitigates the potential misalignment between the compressed subspace and downstream objectives by coupling the directions preserved for compression with those required for adaptation, yielding a deployable low-rank model that avoids retaining full frozen weights while enabling fast and robust tuning. On vision, JACTUS attains an average 89.2% accuracy on ViT-Base across eight datasets at 80% retained parameters, surpassing strong 100% PEFT baselines (e.g., DoRA 87.9%). On language, JACTUS achieves an 80.9% average on Llama2-7B commonsense QA at the same 80% retained-parameter budget, outperforming 100% PEFT (e.g., DoRA 79.7%) and exceeding prior compress-then-finetune pipelines under the same ratained-parameter budget. We will release code.

We present a method for diagnosing interpretation in neural networks by identifying an input subspace where a proposed interpretation is highly faithful. Our method is particularly useful for causal-abstraction-style interpretability, where a high-level causal hypothesis is evaluated by interchange interventions. Rather than treating interchange intervention accuracy as a single global summary, we refine this framework by partitioning the input space into well-interpreted and under-interpreted regions according to pairwise interchange-intervention behavior. This turns causal abstraction from a purely global evaluation into a more diagnostic tool: it not only measures whether an interpretation works, but also reveals where it works, where it fails, and what distinguishes the two cases. This diagnostic view also provides practical heuristics for improving interpretations. By analyzing the structure of the well-interpreted and under-interpreted regions, we can identify missing distinctions in a high-level hypothesis, discover previously unmodeled intermediate variables, and combine complementary partial interpretations into a stronger one. We instantiate this idea as a simple four-step recipe and show that it yields informative error analyses across multiple causal abstraction settings. In a toy logic task, recursively applying the recipe recovers a high-level hypothesis from scratch. More broadly, our results suggest that partitioning the input space is a useful step toward more precise, constructive, and scalable mechanistic interpretability.

Understanding how neural models represent human-interpretable concepts is challenging. Prior work has explored linear concept subspaces from diverse perspectives, such as probing and concept erasure. We introduce a unified framework to study these subspaces along two axes: \textit{containment}, which tests if a concept is fully represented in a subspace but not outside it, and \textit{disentanglement}, which tests for isolation from other concepts. In experiments on both text and speech models, we first highlight that concept subspaces may not be uniquely determined, and discuss the implications for concept subspace analysis. Then, we compare properties of concept subspaces estimated using five estimators, proposed in different communities. We find that (1) the choice of estimator impacts the containment and disentanglement properties; (2) the state-of-the-art concept erasure method, LEACE, performs well on both testing axes, but still struggles to generalize to unseen data; and (3) in HuBERT speech representations, phone information is both contained and disentangled from speaker information, while speaker information is hard to contain in a compact subspace, despite being disentangled from phones.

LoRA adapts large language models (LLMs) by restricting updates to low-rank subspaces of pre-trained weights. While this substantially reduces training cost, the effectiveness of adaptation critically depends on which subspace is chosen at initialization: a poor initialization that allocates capacity to task-irrelevant directions can severely hinder downstream performance. Existing initialization strategies primarily rely on the intrinsic properties of pre-trained weights, implicitly assuming that weight geometry alone reflects task relevance. However, such criteria overlook how the model interacts with the downstream data distribution. In this work, we formulate LoRA initialization as identifying the degree of impact of directions in parameter space under the target data distribution. We argue that data-aware sensitivity, rather than weight-only magnitude, should govern the choice of adaptation subspaces. Building on this perspective, we propose a Fisher-guided framework that leverages curvature information induced by downstream data to characterize how parameter perturbations influence model predictions. This perspective yields a principled, task-dependent criterion for selecting LoRA directions that better align adaptation with the target objective. Empirical results across diverse tasks and modalities demonstrate that data-aware initialization consistently and significantly improves downstream performance over existing approaches.

While Graph Foundation Models (GFMs) have achieved remarkable success in homogeneous graphs, extending them to multi-domain heterogeneous graphs (MDHGs) remains a formidable challenge due to cross-type feature shifts and intra-domain relation gaps. Existing global feature alignment methods (PCA or SVD) enforce a shared feature space blindly, which distorts type-specific semantics and disrupts original topologies, inevitably leading to "Type Collapse" and "Relation Confusion". To address these fundamental limitations, we propose Decoupled relation Subspace Alignment (DRSA), a novel, plug-and-play relation-driven alignment framework. DRSA fundamentally shifts the paradigm by decoupling feature semantics from relation structures. Specifically, it introduces a dual-relation subspace projection mechanism to coordinate cross-type interactions within a shared low-rank relation subspace explicitly. Furthermore, a feature-structure decoupled representation is designed to decompose aligned features into a semantic projection component and a structural residual term, adaptively absorbing intra-domain variations. Optimized via a stable alternating minimization strategy based on Block Coordinate Descent, DRSA constructs a well-calibrated, structure-aware latent space. Extensive experiments on multiple real-world benchmark datasets demonstrate that DRSA can be seamlessly integrated as a universal preprocessing module, significantly and consistently enhancing the cross-domain and few-shot knowledge transfer capabilities of state-of-the-art GFMs. The code is available at: https://github.com/zhengziyu77/DSRA.

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.

Parameter-efficient fine-tuning (PEFT) methods face a tradeoff between adapter size and expressivity: ultra-low-parameter adapters are confined to fixed low-rank subspaces, capping performance even with extended training. We propose BoostLoRA, a gradient-boosting framework that overcomes this limit by iteratively training and merging minimal adapters on the examples the current model gets wrong. A ROTATE SVD basis strategy assigns each round to an orthogonal subspace, so cumulative effective rank grows linearly with the number of rounds while each adapter remains ultra-low-rank. After merging, adapters are discarded, leaving zero inference overhead. On Qwen2.5-3B, BoostLoRA reaches 89.1% on GSM8K and 68.8% on MATH-500, surpassing both the best single-shot ultra-low parameter adapter (TinyLoRA) and full fine-tuning; on code generation it reaches 57.2% on MBPP and 80.4% on HumanEval while full fine-tuning drops below the zero-shot baseline. We also demonstrate cross-architecture transfer on protein binding classification with ESM2-650M and cross-entropy training. BoostLoRA is, to our knowledge, the first PEFT method whose effective rank grows with training, separating per-round parameter cost from total representational capacity.