Representation-Level Certification

Multimodal retrieval integrates video, audio, subtitles, and text; however, recent geometric aggregators, such as Gramian volumes, hyperbolic volumes, and spectral objectives, treat all modalities symmetrically. Under a unified evaluation protocol, their joint scores frequently lag behind the strongest single-modality pathway by 1.9 to 27.6 R@1. Controlled analyses attribute this outcome to uniform modality influence. This work introduces Spherical Centroid Aggregation with Learned Adaptive Relevance (SCALAR), a query-conditioned aggregator that assigns relevance-based weights to each available modality before computing a spherical centroid. SCALAR accommodates arbitrary modality subsets and is trained on masked, reduced-arity views using rank-8 LoRA adapters. Across five benchmarks, SCALAR achieves positive aggregation gain on four, reaching +4.0 R@1, while none of the evaluated prior aggregators is positive on more than one. A uniform-weight ablation reproduces the degradation observed with symmetric aggregation. With only 4.8 million trainable parameters, SCALAR attains the highest text-to-video R@1 on three and performs within seed variation of the best result on a fourth. Under test-time modality dropout, SCALAR's representation-stage score surpasses the released GRAM checkpoint at every evaluated masking rate and benchmark by 3.2 to 10.9 R@1. Finally, as modalities are removed, rerankers trained exclusively on complete modality sets increasingly converge toward their video-only pathways, diminishing these representation-level gains and underscoring a limitation of standard two-stage retrieval pipelines.

A deployed language model may refuse a harmful request in English yet comply with its faithful translation, revealing a cross-lingual safety failure that cannot be characterized reliably by output behavior alone. We formalize this phenomenon through an audited equivalence relation and show that, for a declared quotient, representation, metric, feature dictionary, scoring head, threshold, and contrast model, the resulting safety drift admits an exact linear-algebraic characterization. Specifically, the drift is a cross-Gram functional of the within-fiber contrast; its worst admissible value is a support function, while margin invariance is characterized by an annihilator condition. We introduce an intrinsic calibrated exposure measure, governed by the leverage duality $χ^2=1/\ell-1$, which separates observed drift into three diagnostically distinct regimes: a reader fault removable by recalibration, an exact correction that is too ill-conditioned to be reliable, and a representation-level collision that no readout-only intervention can remove. Thus, identical observed exposure can lead to fundamentally different remediation verdicts. The framework also extends to cone-valued safety heads. An untied order-swap identity provides a diagnostic for the linear control interface; its calibration-state residual predicts a distinct three-control composition error on unseen states and targets, achieving median Spearman correlation $0.964$, compared with $0.269$ for a static cross-Gram baseline. etc.....

Pretrained generative robot policies learn expressive action priors from demonstrations. However, existing reinforcement learning methods only steer the noisy space but fail to modulate intermediate action representations during the generation process, resulting in performance degradation and inefficiency. To address this limitation, we propose a novel Dual-Latent Space Reinforcement Learning (DLSRL) framework, which complements initial-noise steering with representation-level control inside the frozen generator. Specifically, our actor network predicts two distinct latent variables: an initial-noise latent variable that steers behavior generation, and an action-representation latent variable for intermediate feature modulation. Moreover, this representation latent variable is mapped to adapter features and ingeniously injected into the hidden states of intermediate action tokens via residual connections. Our dual-control design enables direct adjustment of action representations without updating the base policy. Experiments across generative policy architectures and robotic manipulation tasks show that DLSRL effectively accelerates online robot policy adaptation and achieves competitive performance. Our code is available at \href{https://github.com/xianchaoxiu/DLSRL}{https://github.com/xianchaoxiu/DLSRL}.

Activation Oracles (AOs) are language models trained to answer natural-language questions about another model's internal activations. They offer a flexible interface for reading hidden information from model states, especially when relevant information is internally represented but absent or incomplete in visible behavior. However, AOs are themselves learned systems: their answers are shaped by training data, objectives, and learned reporting behavior, rather than being neutral readouts of represented information. We study this in a controlled Taboo Word Guessing setting, where subject models are fine-tuned to internally use a hidden concept while avoiding direct disclosure. Contrary to the expectation that an AO trained on such a subject becomes a specialist reader, we find that fine-tuned AOs can become concept-specific anti-readers: they selectively fail to recover the concept persistently present during their own training. This failure is not simply explained by absence of the concept from the subject or oracle representations: the target remains decodable inside the oracle, while LogitLens and layer-ablation analyses indicate that the failure arises in the AO readout pathway. Our results show that behavioral leakage, representation-level decodability, and AO-verbalizability can come apart, raising a reliability concern for learned interpretability interfaces.

Video world models are increasingly used in autonomous driving to forecast future scene evolution and provide future-aware spatio-temporal representations for downstream action prediction. In perception-to-action pipelines, these representations can directly influence ego-vehicle waypoint planning, making the learned future dynamics a critical security-sensitive component. Despite their promise, the training-time security risks of autonomous-driving video world models remain largely unexplored. We present BadDreamer, a transferable spatio-temporal backdoor attack that targets the perception side of this pipeline. Unlike conventional backdoors that manipulate image labels, prompt outputs, or action supervision, BadDreamer poisons the learned transition dynamics of a video world model. It constructs trigger-erasure sequences in which an oncoming yellow delivery rider is visible in the observed context frames but erased from the future frames. After fine-tuning on a small fraction of such sequences, the compromised world model learns a hidden conditional association: when the physical trigger appears, it hallucinates a future where the rider disappears and the road appears clear. We further show that this corrupted future-aware representation can transfer to the downstream action module without directly modifying ego-trajectory labels, inducing unsafe non-evasive waypoint predictions. Our experiments instantiate this attack on a representative open-source perception-to-action pipeline, revealing a representation-level safety risk in autonomous-driving video world models and highlighting the need for backdoor-aware validation beyond clean generation quality.

Large Language Model (LLM) safety has often been evaluated at the behavior level, which provides limited evidence of internal robustness, as these evaluations target outputs rather than representation-level vulnerability under intervention. We formalize this discrepancy as the audit gap: the difference between behavioral safety and robustness under intervention. To study this gap, we construct dissociated models that preserve safe outward behavior while remaining vulnerable in the latent space. We introduce an intervention-based evaluation framework to test model robustness through soft interventions in parameter and latent spaces, including harmful fine-tuning and layer-wise latent perturbations. To formalize the evaluation, we propose the Latent Vulnerability Score (LVS) to measure how easily harmful behavior can be elicited by bounded latent perturbations. Using this evaluation framework, we show that behavioral safety metrics are insufficient measures of representation-level robustness across multiple safely and unsafely aligned state-of-the-art models. Notably, dissociated models show substantially elevated LVSs despite comparable refusal behavior under harmful intervention, with intermediate representations being the most sensitive to intervention. Our results suggest that behavioral safety evaluation alone provides an incomplete picture of model robustness, motivating representation-aware audits of latent vulnerability and observable behavior.

Causal video generators must predict from the past, but they need not learn only from it. In streaming autoregressive video diffusion, each emitted segment becomes a commitment that future segments must preserve. Standard training, however, only asks each causal state to explain the present. This creates what we call a representation-level planning gap: states that fit the current segment may discard identity, layout, and motion information needed for a consistent future. We introduce Video-Mirai, a training-only method that closes this gap without changing causal inference: the generator rolls out causally, a frozen foresight encoder reads the completed rollout non-causally, and a lightweight predictor distills the resulting stopped-gradient targets into causal states. Future frames supervise representations, never generator inputs. At inference, the encoder and predictor are discarded, leaving the original architecture, per-step FLOPs, and KV-cache behavior unchanged. Video-Mirai improves a strong Causal-Forcing baseline on 5-second VBench from 83.8 to 84.6 in terms of Total Score. On 30-second rollouts beyond the training horizon, subject consistency improves from 84.9 to 88.5 and background consistency from 90.2 to 91.9. Ablations identify future-conditioned targets as the key ingredient, and probes show that future frames become more decodable from current features. Causality should constrain inference, not representation supervision. Our study highlights that visual autoregressive models need foresight. Project page: https://y0uroy.github.io/Video-Mirai.

Reinforcement learning (RL) post-training has shown to improve reasoning in large language models (LLMs). However, there has been little exploration on the problem of data contamination in RL post-training, potentially undermining generalization and evaluation reliability of the training process itself. Existing detection methods primarily rely on output-level signals such as likelihood or entropy, which become unreliable for RL-trained models since RL shapes behavior through trajectory-level rewards rather than token likelihoods. We propose LaRA, a layer-wise representation analysis framework for detecting contamination in RL post-trained LLMs. LaRA introduces three complementary metrics, measuring perturbation sensitivity, directional collapse, and local representation rigidity under controlled perturbations. We find that contamination produces progressive geometric deviations across layers, including amplified perturbation sensitivity, stronger directional collapse, and enhanced local rigidity. Based on our findings, we also develop a contamination detection protocol that aggregates representation-level deviations across layers and metrics. Experiments on RL-trained reasoning models show that our protocol outperforms existing output-level baselines for contamination detection.

Machine unlearning aims to remove the influence of specific training records from a deployed model without retraining from scratch. Current protocols verify this at the output level through membership inference, retain accuracy, and forget-set accuracy, but a model can satisfy all three whilst still encoding forgotten records in its intermediate representations. We introduce RULER, a set of representation-level verification metrics. The oracle-comparative metric M2 measures whether forget-set records occupy the same representational position as in a model retrained without them. The oracle-free metric M4 detects residuals from the unlearned model's internal similarity structure alone, without retraining. Four approximate unlearning methods all pass output-level evaluation, yet under a linear mixed-effects model M2 detects significant residuals in 10 of 12 conditions (p<0.05), with effect sizes growing as the forget fraction increases. A fifth method, Bad Teacher, shows the same residuals despite a different forgetting mechanism. M4 acts as a pre-unlearning diagnostic across tabular, image, clinical text, and face-identity settings: it detects identity-level memorisation in face recognition models where no tested method fully erases the signal.

Federated learning enables collaborative training across distributed clients with heterogeneous data, but such heterogeneity often leads to unstable updates and degraded global performance. Moreover, in practical deployments, client updates may deviate from the expected behavior not only due to benign not i.i.d. distributions, but also due to distributional shifts or anomalous inputs, raising concerns about the reliability of the aggregation process. In this work, we propose a lightweight geometric signal to quantify the functional deviation of a client with respect to the global model. Instead of comparing model parameters or gradients, our approach measures how the local training of each client alters the activation-induced partition of the input space, evaluated on a shared probe set. This yields a permutation-invariant, interpretable metric of client--global divergence that captures differences in how data is processed by the model. We show that this signal effectively identifies clients that induce atypical functional changes, distinguishing stable yet heterogeneous clients from those whose updates significantly diverge from the global regime. As a result, the proposed metric provides a simple tool for monitoring client behavior and enabling risk-aware aggregation strategies in federated learning systems.

Machine unlearning in Vertical Federated Learning (VFL) has attracted growing interest, yet existing methods certify forgetting solely using output-level metrics. We challenge these works by introducing Mirage, a representation-level auditing framework that comprises four complementary diagnostics: Linear probe recovery (LPR), centered kernel alignment (CKA), feature separability scoring, and layer-wise recovery analysis. Extensive experiments across seven datasets and seven baseline methods following recent VFL unlearning protocols reveal three key findings: (1) Forgetting gap: methods that pass output-level certification still retain substantial class structure in their representations, with LPR exceeding the retrained baseline by up to 15.4 points; CKA shows that these models remain structurally closer to the original than to the retrained reference, while separability scores indicate persistent geometric discrimination. (2) Unlearning trilemma: no existing method simultaneously achieves high utility, output-level forgetting, and representation-level forgetting. (3) Class-sample asymmetry: class-level forgetting leaves strong representational traces (LPR exceeding 96 percent on several datasets), whereas sample-level forgetting is indistinguishable from chance (LPR is approximately 50 percent); layer-wise analysis further shows that residual class information persists across network depths. These findings call for representation-aware evaluation standards in federated unlearning research. Code is publicly available at https://github.com/YuZhenyuLindy/Mirage.

For reinforcement learning in data-scarce domains like real-world robotics, intensive data reuse enhances efficiency but induces overfitting. While prior works focus on critic bias, representation-level instability in Self-Predictive Learning (SPL) under high Update-to-Data (UTD) regimes remains underexplored. To bridge this gap, we propose Robust Representation via Redundancy Reduction (R2R2), a regularization method within SPL. We theoretically identify that standard zero-centering conflicts with SPL's spectral properties and design a non-centered objective accordingly. We verify R2R2 on SPL-native algorithms like TD7. Furthermore, to demonstrate its orthogonality to prior advancements, we extend the state-of-the-art SimbaV2, which originally lacks SPL, by integrating a tailored SPL module, termed SimbaV2-SPL. Experiments across 11 continuous control tasks confirm that R2R2 effectively mitigates overfitting; specifically, at a UTD ratio of 20, it improves TD7 by ~22% and provides additional gains on top of SimbaV2-SPL, which itself establishes a new state-of-the-art. The code can be found at: https://github.com/songsang7/R2R2

Large language models from different families use different hidden dimensions, tokenizers, and training procedures, making behavioral directions difficult to compare or transfer across models. We introduce an anchor-projection framework that maps hidden representations from each model into a shared anchor coordinate space (ACS). Behavioral directions extracted from source models are projected into ACS and averaged into a canonical direction. For a new model, the canonical direction is reconstructed into its native hidden space using only anchor activations, without fine-tuning or target-specific direction extraction. We evaluate five instruction-tuned model families and ten behavioral axes. We find that same-axis directions align tightly across the Llama-Qwen-Mistral-Phi (LQMP) cluster in ACS. This shared structure transfers to downstream tasks. For the aligned LQMP cluster, held-out targets achieve (0.83) ten-way detection accuracy and (0.95) mean binary AUROC, while canonical steering induces refusal-rate shifts of up to +0.46% under distribution shift. Sensitivity analyses show that two source models and small anchor pools already suffice to approximate transferable directions. Overall, ACS provides a novel perspective on cross-family interpretability, revealing that representation-level transfer remains robust across model families.

Fine-tuning safety-aligned language models for downstream tasks often leads to substantial degradation of refusal behavior, making models vulnerable to adversarial misuse. While prior work has shown that safety-relevant features are encoded in structured representations within the model's activation space, how these representations change during fine-tuning and why alignment degrades remains poorly understood. In this work, we investigate the representation-level mechanisms underlying alignment degradation. Our analysis shows that standard fine-tuning induces systematic drift in safety-relevant representations, distorts their geometric structure, and introduces interference between task optimization and safety features. These effects collectively lead to increased harmful compliance. Motivated by these findings, we introduce REFUSALGUARD, a representation-level fine-tuning framework that preserves safety-relevant structure during model adaptation. Our approach constrains updates in hidden representation space, ensuring that safety-mediating components remain stable while allowing task-specific learning in complementary directions. We evaluate REFUSALGUARD across multiple model families, including LLaMA, Gemma, and Qwen, on adversarial safety benchmarks such as AdvBench, DirectHarm4, and JailbreakBench, as well as downstream utility tasks. Our approach achieves attack success rates comparable to base safety-aligned models while maintaining competitive task performance, significantly outperforming baselines.

Unified Multimodal Models (uMMs) aim to support both visual understanding and visual generation within a shared representation. However, existing evaluation protocols assess these two capabilities independently and do not examine whether they are semantically aligned. As a result, it remains unclear whether current uMMs learn coherent unified representations that remain consistent across tasks given a visual concept. We introduce XTC-Bench, a scene-graph-grounded evaluation framework that measures cross-task visual semantic consistency. By deriving both generation prompts and understanding queries from a structured scene graph, our framework enables fact-level alignment analysis across objects, attributes, and relations. We propose Continuous Cross-Task Agreement (CCTA), a fine-grained metric that quantifies semantic agreement between generation and understanding over matched atomic facts, isolating internal consistency from standalone task accuracy. Extensive experiments on eight open-source and one commercial unified models reveal that high generation or understanding performance does not imply strong cross-task alignment, and architectural analysis shows consistency is governed by how tightly learning objectives are coupled across modalities, not by architectural unification alone. XTC-Bench provides a reproducible and model-agnostic framework for diagnosing representation-level misalignment, offering a concrete direction for advancing unified multimodal modeling beyond isolated task performance.

Reliable deployment of language models requires two capabilities that appear distinct but share a common geometric foundation: predicting whether a model will accept targeted behavioral control, and detecting when its internal structure degrades. We show that geometric stability, the consistency of a representation's pairwise distance structure, addresses both. Supervised Shesha variants that measure task-aligned geometric stability predict linear steerability with near-perfect accuracy ($ρ= 0.89$-$0.97$) across 35-69 embedding models and three NLP tasks, capturing unique variance beyond class separability (partial $ρ= 0.62$-$0.76$). A critical dissociation emerges: unsupervised stability fails entirely for steering on real-world tasks ($ρ\approx 0.10$), revealing that task alignment is essential for controllability prediction. However, unsupervised stability excels at drift detection, measuring nearly $2\times$ greater geometric change than CKA during post-training alignment (up to $5.23\times$ in Llama) while providing earlier warning in 73% of models and maintaining a $6\times$ lower false alarm rate than Procrustes. Together, supervised and unsupervised stability form complementary diagnostics for the LLM deployment lifecycle: one for pre-deployment controllability assessment, the other for post-deployment monitoring.

Aligned language models that are trained to refuse harmful requests also exhibit over-refusal: they decline safe instructions that seemingly resemble harmful instructions. A natural approach is to ablate the global refusal direction, steering the hidden-state vectors away or towards the harmful-refusal examples, but this corrects over-refusal only incidentally while disrupting the broader refusal mechanism. In this work, we analyse the representational geometry of both refusal types to understand why this happens. We show that harmful-refusal directions are task-agnostic and can be captured by a single global vector, whereas over-refusal directions are task-dependent: they reside within the benign task-representation clusters, vary across tasks, and span a higher-dimensional subspace. Linear probing suggests that the two refusal types are representationally distinct from the early transformer layers. These findings provide a mechanistic explanation of why global direction ablation alone cannot address over-refusal, and establish that task-specific geometric interventions are necessary.

The limited availability of dysarthric speech data makes cross-lingual detection an important but challenging problem. A key difficulty is that speech representations often encode language-dependent structure that can confound dysarthria detection. We propose a representation-level language shift (LS) that aligns source-language self-supervised speech representations with the target-language distribution using centroid-based vector adaptation estimated from healthy-control speech. We evaluate the approach on oral DDK recordings from Parkinson's disease speech datasets in Czech, German, and Spanish under both cross-lingual and multilingual settings. LS substantially improves sensitivity and F1 in cross-lingual settings, while yielding smaller but consistent gains in multilingual settings. Representation analysis further shows that LS reduces language identity in the embedding space, supporting the interpretation that LS removes language-dependent structure.

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.