Papers

Concept erasure aims to remove a target concept, such as a copyrighted style, a recognizable character, or unsafe content, from a pretrained text-to-image diffusion model while preserving its ability to generate other content. Existing activation steering methods build an erasure direction mainly from the target concept and adjust model activations along it at inference time. However, target and retained concepts often overlap in the model's representation space, so this direction also contains shared components that retained concepts rely on. Steering directly along this direction can therefore suppress retained concepts and harm the generation of non-target content. To address this issue, we propose Retain-aware Activation Steering (RASteer), a training-free method. RASteer first builds a retain subspace from the concepts to preserve. Retain-Orthogonal Steering (ROS) then removes components aligned with this subspace from the erasure direction, making steering more specific to the target. Since fully removing the shared components can weaken erasure, we further introduce Overlap-Adaptive Calibration (OAC). At each layer and denoising step, OAC uses the overlap between the erasure direction and the retain subspace to control how much of each shared component is removed, balancing target erasure and concept preservation. Experiments on unsafe-content, instance, and artistic-style erasure across multiple backbones and benchmarks show that RASteer matches or outperforms the activation steering and weight editing baselines we evaluate, achieving a better balance between erasure and preservation.

In this paper, we study the following Gierer–Meinhardt system \begin{equation*} \begin{cases} u_t=d_1\Delta u-a_1u+u^p/v^q+\delta_1(x),\quad x\in \Omega, t>0,\\ v_t=d_2\Delta v-a_2v+u^r/v^s+\delta_2(x), \quad x\in \Omega, t>0,\\ \partial_\nu u=\partial_\nu v=0,\qquad\qquad \qquad\qquad\quad x\in \partial \Omega, t>0,\\ u(x,0)=u_0(x),\quad v(x,0)=v_0(x),\quad x\in \Omega, \end{cases} \end{equation*} where $\Omega\subset \mathbb{R}^N$ is a bounded domain, $\nu$ is the outer norm vector with respect to the smooth boundary $\partial \Omega$, $d_1,d_2, p,q,r>0$, $s>-1$, $p-1<r$, $qr=(p-1)(s+1)$, $\delta_1(x)$, $\delta_2(x)$, $u_0(x)$ and $v_0(x)$ are continuous functions. For every pair $d_1,d_2>0$, we respectively establish the global existence, uniform boundedness and asymptotic behavior for Gierer–Meinhardt system with critical exponent $qr=(p-1)(s+1)$ under certain conditions.

As large language models (LLMs) keep growing in size and complexity, their training frameworks evolve at a rapid pace as well. Therefore, continuous integration (CI) is critical for maintaining the quality and stability of these frameworks. However, unlike traditional software, CI for LLM training frameworks relies on GPU-intensive tests, which usually involve complete model training or evaluation. This leads CI itself to become a new bottleneck for fast-paced development. In this paper, we introduce FastCI, a framework that improves the efficiency of CI for LLM training frameworks. FastCI leverages runtime evidence to select affected tests and prune tests that execute changed code in equivalent contexts. Then FastCI prioritizes high-risk tests to expose potential failures earlier, and optimizes test workloads along dimensions outside the intended validation scope of each test. Evaluated on the CI workload of our LLM training framework, FastCI reduces the CI latency by 77.5% and the GPU resource usage by 63.9%, while improving the modified code coverage retention by 3.2%, compared with the currently deployed CI pipelines. FastCI has now been integrated into the CI pipelines of our LLM training framework at ByteDance.

In this paper we prove a generalized version of Mel'nikov's Criterion, which has also been proved by Haran, and use it in order to connect the realization as absolute Galois groups of free pronite and free pro-p products of a sheaf of pro-p groups.

We present an optimal method for studying new physics effects in high-mass dilepton production at the LHC. This method relies on a reweighting procedure that is precise to the next-to-leading order in quantum chromodynamics. The only computationally expensive input required is a standard model $\mathrm{pp\rightarrow V^{(*)}\rightarrow\ell^+\ell^-}$ sample, with $\mathrm{V=Z,\gamma^*}$, generated at the same level of precision and with sufficient statistical coverage of the phase space at high invariant masses. The proposed method is suitable for any $\mathrm{q\overline{q}\ell^+\ell^-}$ contact-interaction scenario, regardless of the assumed underlying structure and couplings. It is also ideal for searching for medium-to-large-width new neutral resonances produced in the s-channel, as changes to the input values for mass, coupling, or width can be easily implemented using the same initial generated samples.

Numerical relativity is a term used to describe the solution of the full non-linear Einstein Equations as an initial value problem using numerical methods. Individual simulations are computationally expensive and in many applications the high complexity of the parameter space and non-linear sensitivity of the solutions makes locating particular outcomes of interest challenging. In this work, we demonstrate how auto-differentiation can efficiently find dynamical solutions of interest, using the classic problem of scalar critical collapse as a proof of principle. We locate a line of critical solutions in a 2D parameter space using 35 runs, compared to the 270 runs required to locate 10 critical points using bisection, and directly calculate the unstable mode, confirming its universality along the line. The method is straightforwardly extendable to higher-dimensional parameter spaces, offering a promising method to efficiently target more complex solutions in strong dynamical gravity regimes.

Emergency department (ED) triage is a high-stakes prioritization task in which demographic, socioeconomic, and system-context information may improperly influence acuity assignment. Although open-source large language models (LLMs) are increasingly considered for local and privacy-preserving clinical decision support, it remains unclear how counterfactual bias varies across model families, sizes, medical-domain models, and domain-adapted models. We present a comparative counterfactual audit of ten open-source LLMs for pediatric Emergency Severity Index (ESI) prediction. Starting from real and handbook-style clinical vignettes, we construct paired counterfactual variants that change only one injected demographic, socioeconomic, healthcare-access, behavioral, social, or system-context variable while holding the clinical presentation fixed. Models include Qwen2.5-7B, Qwen2.5-14B-Instruct, a QLoRA fine-tuned Qwen2.5-7B, MedGemma variants, MedLLaMA2-7B, GPT-OSS-20B, and GPT-OSS-120B. We measure any counterfactual shift, undertriage, overtriage, shifts greater than one ESI level, mean shift, and mean absolute shift. Counterfactual sensitivity varied substantially and did not consistently decrease with larger model size or medical-domain pretraining. The fine-tuned Qwen2.5-7B showed the lowest overall sensitivity, with a 5.27% any-shift rate and mean absolute shift of 0.0534, versus 16.02% and 0.1706 for the base model. Several larger or medical-domain models showed more significant shifts. Stratified and correlation analyses further revealed clinically important directionality and shared failure patterns hidden by aggregate rates. These findings support counterfactual auditing as a lightweight, clinically interpretable framework for comparing fairness risks in open-source LLMs before clinical deployment.

The ability of vision-language models (VLMs) to associate visual identities with biographical information creates a need for selective unlearning of personally identifiable information (PII) while preserving permitted knowledge about the same individual. This setting is challenging because both sensitive and retained information can share the same visual inputs and intermediate representations. We introduce SIEVE, a simple and effective framework for selective VLM unlearning. SIEVE directly regularizes attention-value representations while also controlling model outputs. SIEVE suppresses attention values for forget examples toward a constant zero, while preserving retain-example representations by matching them to a frozen reference model. These objectives are combined with sequence-level forget and retain supervision, enabling targeted forgetting without largely affecting retained knowledge. Extensive experiments show that SIEVE achieves state-of-the-art performance on unlearning with multiple model-modality settings, while maintaining competitive retained utility. Ablation studies further show that value suppression and negative cross-entropy contribute complementary forgetting signals, while reference-based value matching substantially reduces utility degradation. These results demonstrate that attention values provide an effective intervention point for selective multimodal unlearning when sensitive and retained knowledge are closely related.

Accent conversion (AC) requires changing accent while preserving speaker identity and linguistic content, yet parallel recordings are scarce. Speech synthesis provides an alternative source of supervision, but generated targets vary in accent realization and source preservation. We propose a multi-sample synthetic supervision framework that constructs conversion targets by jointly assessing these properties across candidate waveforms for each source–accent condition. Selected candidates provide associated discrete speech codes, representing linguistic content, prosody, and speaking style, as targets for accent-conditioned autoregressive adaptation. Compared with using one generated target per training example, our method improves target-accent classification accuracy by 2.49 percentage points, while speaker similarity remains nearly unchanged, and word error rate increases by 0.29 percentage points. Across six target accents, our method achieves 7.81% WER and the highest mean listening ratings among evaluated systems.

Recent work on embedded post-quantum cryptography has focused primarily on instruction-level optimization, including arithmetic-kernel improvements, assembly tuning, register allocation, and instruction scheduling. Using the Module-Lattice-Based Key-Encapsulation Mechanism (ML-KEM) on an Arm Cortex-M7 as a case study, we examine the additional gains available from memory-hierarchy utilization, tightly coupled memory placement, peripheral integration, clock configuration, and deterministic public-data reuse. The evaluation starts from a state-of-the-art SLOTHY-optimized implementation and covers all three ML-KEM parameter sets. Without modifying the cryptographic algorithm or standardized wire formats, the evaluated profiles without auxiliary public state reduce cycles by up to 2.5%. A selected public-data-reuse profile reduces encapsulation and decapsulation cycles by up to 74.6% and 58.8%, respectively. These results demonstrate that substantial deployment gains remain after arithmetic-kernel optimization and motivate a two-stage methodology that also examines the surrounding execution system.

Path finding and multi-robot motion planning require trajectories that are smooth, goal-directed, and collision-free in environments with complex geometric constraints. Recent diffusion-based planners have shown that trajectory generation can be cast as iterative denoising which has opened the doors to learning-based approaches that can handle multi-modal trajectory distributions and refine entire trajectories. However, a key limitation is that diffusion planners require training on large collections of feasible trajectories, rendering them map-specific, and difficult to deploy when high-quality demonstrations are unavailable. This paper introduces a training-free diffusion-based motion planner that replaces learned global trajectory scores with analytical local scores derived from obstacle, smoothness, velocity, and inter-agent feasibility terms. The proposed idea relies on a key observation: the score of a trajectory can be reconstructed by considering only local interactions between neighboring waypoints and nearby constraints. This structure exploitation yields a decomposed denoising procedure that retains the optimization structure of classical trajectory methods while inheriting the iterative refinement behavior of diffusion models. Experiments on a large collection of complex environments and large multi-agent planning tasks show that the proposed analytical score produces smooth and feasible trajectories within limited computational costs, for example in generating feasible paths for 300+ agents in environments containing 100+ obstacles in under 6 seconds on a GPU, outperforming strong learning-based and optimization baselines, while avoiding the data requirements of learned diffusion planners.

We develop a measure-theoretic framework for small-order limits of nonlocal energies with metric-valued maps. For globally $L^p$ maps, we consider interaction measures whose marginals have uniformly bounded $L^\infty$ densities $a_\alpha$ and $b_\alpha$. If these densities converge weakly-* to $a_\infty$ and $b_\infty$, the normalized energies converge to the $L^p$ energy weighted by $a_\infty+b_\infty$ if and only if the interaction measures escape every bounded rectangle. For maps that are only locally $L^p$, we introduce target laws describing the distribution of the values sampled at infinity. Convergence of their $p$-distance profiles yields the limiting interaction energy for every $1\leq p<\infty$ and arbitrary Polish targets, with integrated $p$-Wasserstein convergence as a sufficient criterion. The arguments include the endpoint $p=1$ and require no linear structure on the target space. Applications include directional and anisotropic Maz'ya–Shaposhnikova formulas, a Bernoulli-law interpretation of fractional perimeters, and an Abelian principle for heat-semigroup energies.

In this paper, we present an algorithm for constructing all Calabi-Yau orbifolds and their mirror pairs associated with an arbitrary Berglund-Hubsch-type model, and prove its correctness. In particular, the proposed method yields explicit generators for the groups defining the corresponding orbifolds. The method for classifying maximal admissible groups is proposed. We illustrate the application of the classification approach to an arbitrary Fermat-type Calabi-Yau manifold. In addition, as an illustrative example, we classify all maximal admissible groups of Fermat-type Calabi-Yau threefolds.

Complex surgical procedures around critical anatomy, such as the endoscopic endonasal skull base surgery, requires significant practice and training on the part of the surgeon before they are allowed to perform the operation on a live patient. This training in typically done in cadaveric specimens, due to them containing the same critical structures as a living human. However, cadavers are not a perfect 1-to-1 substitute for a living patient. The dead and preserved tissues of a cadaver are colored completely differently than a living human, and – without complex and expensive pumping systems – do not bleed in the same way. As a result, identifying the critical pieces of anatomy that make this procedure so complex can be quite different in a live case than in a surgeon's cadaveric practice. This paper presents EndoLive, a framework for real-time style transfer between cadaveric endoscopic video and living human endoscopic video. Our method combines the ConStructS GAN model for realistic style transfer for surgical applications, with the HyPER-GAN model that can learn complex translations and perform them in real-time. We train EndoLive on unpaired cadaveric and live images taken from an endoscope, and test the trained model with cadaveric video, on a variety of devices. Experimental results demonstrate that EndoLive can perform cadaveric-to-live translation at speeds well above the minimum necessary for real-time, while maintaining semantic consistency of critical anatomical structures. Our source code is available at https://github.com/griffhurt/endolive.

Outcome-based reinforcement learning can train language models to forecast real-world events, but prior forecasting work either freezes research context before training or deploys agentic research only at test time, so the skill of gathering evidence is never shaped by the reward. We introduce an agentic forecasting environment, dataset, and harness built from 2,100+ resolved Polymarket questions; the agent acquires its own context at rollout time (web search, page reading, and financial time series, all restricted by layered leak filtering to information published before each question's cutoff), and we train Qwen3.5-35B-A3B (3B active parameters) on it with single-epoch GRPO under a Brier-score reward. Training changes how the agent interacts with information: calibration improves 30-40%, and search attempts fall from 3.8 to 2.25 per rollout as evidence discipline is learned. Evaluated in an identical harness against four frontier models, the trained policy also finishes ahead of every frontier model tested at evidence-based forecasting, including Claude Opus 4.5 (soft-Brier 0.254 vs. 0.256, n=265), at about 5% of the inference cost, and its margin is widest on the hardest questions, the ones the crowd itself had not decided. We release the environment, dataset, and per-rollout records as a reusable harness for temporal forecasting agents.

Compact ground-state preparation circuits are essential for early fault-tolerant quantum chemistry. We introduce Variational Imaginary-time Majorana Evolution (VIME), a classical pre-training algorithm that extends operator-projected variational quantum imaginary-time evolution (OVQITE) to molecular electronic structure using a single compiled Majorana-propagation (MP) surrogate graph. We pair VIME with a compressed tiled Unitary Product State (c-tUPS) ansatz that we develop to reduce both classical simulation and quantum state-preparation costs. In the strongly correlated ruthenium complex TLD-1411, active-space calculations requiring up to 52 qubits achieve chemical precision against DMRG reference energies, with energy errors approximately $400\times$ smaller than those from ADAPT-VQE-based variational Majorana-propagation (ADAPT-VMPE). Across the strongly multireference acene series up to heptacene (60 qubits), VIME matches or improves on the MP-reported energy accuracy of the ADAPT-VMPE comparison points with up to $26\times$ fewer CNOTs and $167\times$ lower two-qubit depth. These results establish imaginary-time classical pre-training of compressed variational ansätze as a promising route to resource-efficient molecular ground-state preparation.

Mass-loss in red supergiants (RSGs) is poorly understood and stellar evolution models rely on empirical prescriptions. In recent years, new superwind mass-loss ideas, including an $L/M$ dependence have been proposed, unlike traditional RSG wind recipes which are formulated on a $L$ dependence alone. We investigate how contrasting RSG mass-loss prescriptions affect the late-time evolution of massive stars from their RSG phase onwards. We used MESA to create a series of stellar evolution models with masses ranging from $10\,{\rm M}_\odot - 40\,{\rm M}_\odot$, using the RSG \cite{VS23} $L/M$ recipe and traditional mass-loss prescriptions. Independent of the treatment of the stellar envelope, we find for a 25$\,{\rm M}_\odot$ star that the $L/M$ recipe reaches an order of magnitude higher mass-loss rate than traditional recipes due to its inverse mass dependence. As the RSG luminosity is set by the core mass and becomes largely insensitive to the current stellar mass, this creates a positive feedback loop driving runaway envelope stripping. Consequently, stars with an initial mass $M_{\rm init} \sim \,20\,{\rm M}_\odot$ are able to strip their hydrogen envelope and form classical Wolf-Rayet (WR) stars. We draw this conclusion not solely on the basis of their HRD location, but also on their predicted emission-line spectrum that we compute with the PoWR stellar atmosphere code. Stars with $M_{\rm init}<20\,{\rm M}_\odot$ do not fully strip, meaning they expand after core helium burning. In Appendix A we present a generic analytical model showing that the tendency for runaway stripping increases rapidly with the strength of the inverse mass-loss scaling with mass. Our models demonstrate a viable single-star pathway into the yellow supergiant and WR regimes. The amounts of full and partial stripping have a direct effect on supernova (SN) progenitors, including partially stripped type IIb SNe.

In one-shot training-free voice conversion (VC), the source and reference utterances may contain different linguistic content, so reliable frame-level correspondence between them cannot be assumed. We propose StateVC, which jointly defines a common set of local regions from pooled frame-level WavLM representations of the source and reference utterances; we refer to these regions as states. These shared states are obtained by fitting a pair-specific Gaussian mixture model to the pooled representations, without explicit source–reference frame matching. Within each state, StateVC estimates a source-to-reference mean shift in the original WavLM space. Source-frame posterior probabilities then combine the state-specific shifts so that different frames can receive different local updates. For the LibriSpeech one-shot protocol, StateVC achieves the lowest word error rate (WER) and character error rate (CER) among the evaluated systems, at 8.01% and 3.22%, respectively, with a speaker similarity (SIM) of 0.9512. It also achieves the highest mean perceived speaker similarity among the evaluated systems and the highest mean naturalness among the evaluated training-free systems.

Top-two algorithms are simple and effective for fixed-confidence best-arm identification, but their sharp non-asymptotic behavior is still not well understood. We study this question for Bernoulli bandits through $\beta$-EB-TCI, the empirical-best top-two rule of Jourdan et al., whose challenger minimizes a Bernoulli transportation cost plus a logarithmic penalty in its pull count. We first analyze the stabilized regime, in which the empirical leader has become the true best arm and its sampling fraction stays close to $\beta$. In this regime, every suboptimal arm is sampled linearly often, and, with high probability, the stopping time is at most $T_{\beta}^{\star}(\mu)\log(1/\delta)$ plus stabilization and lower-order concentration terms, with the exact fixed-$\beta$ leading coefficient. The main remaining difficulty is therefore to control when the regime is reached. For the original rule, we obtain a non-asymptotic high-probability stopping bound on every Bernoulli instance with a unique best arm. Under a finite-mean sufficient-exploration condition, the bound yields the sharp expected sample complexity; combined with the exploration result of Jourdan et al., this covers pairwise-distinct means under their tie-breaking convention. A mild forced-exploration guard, which uses only $O(\sqrt{Kt})$ pulls up to time $t$, gives a self-contained sharp expected bound for any number of arms with a unique best arm. We also identify a limitation of proof strategies that try to handle equal suboptimal means through a single index-comparison argument.

Sparse expert activation reduces MoE models' computation, yet expert weights can exceed limited device memory. Offloading makes inference feasible on a compact AI appliance but exposes host-to-device transfers to the inference path. We present MoE-CORE, a system that coordinates expert offloading and residency for memory-constrained MoE inference. It stages complete expert layers in alternating buffers during prefill. During decode, it combines nonuniform layer-wise cache capacity, domain-informed initialization, routing-history-aware replacement, and cross-layer prefetching. The main configuration executes router-selected experts exactly; an optional score-based substitution path handles eligible low-score misses. The main comparison uses 1K- and 128-token output caps for MoE-CORE and vLLM Prefetch, respectively. Across five workloads per model, MoE-CORE records a mean time per output token (TPOT) of 38.0-44.8 ms versus 1268.9-1269.1 ms for the evaluated vLLM Prefetch configuration on DeepSeek-V4-Flash-W4A8; the corresponding values on GLM-5.2-W4A8C8 are 206.6-220.5 and 5941.5-5941.8 ms. Under an 84-GB NPU-memory cap, the best measured DeepSeek GSM8K configuration achieves a TPOT of 21.5 ms with approximate expert substitution and multi-token prediction (MTP) at depth 2. These results support coordinated expert residency and transfer scheduling under a device-memory constraint. The code is here.

Ransomware has emerged as a major cybersecurity threat, with incidents increasing in frequency and impact across critical sectors. These attacks are typically launched through phishing emails, malicious downloads, or exploitation of software vulnerabilities to gain system access. Once inside, the malware encrypts files and demands a ransom, often in cryptocurrency, for the decryption key. Conventional detection methods often struggle with novel or scarce samples, leaving systems vulnerable. To address these challenges, this paper proposes a hybrid deep learning framework that combines an Autoencoder Feature Extractor (AFE) with a Model Agnostic Meta Learning (MAML) classifier for few shot malware detection. The AFE generates compact latent features that reduce noise and dimensionality, while the MAML classifier rapidly adapts to new threats using limited labeled data. Experiments conducted on the Ransomware Dataset 2024 demonstrate the effectiveness of the framework in binary classification tasks. Across one to fifty shot settings, the proposed model consistently achieves high accuracy, F1 score, and Matthews Correlation Coefficient values, maintaining reliable classification even under extreme scarcity. These results highlight the model's robustness and effectiveness in adapting to limited data scenarios, demonstrating the potential of combining feature extraction with meta learning to enhance resilience against malware, particularly in sectors such as healthcare, manufacturing, and public infrastructure, where cyberattacks can cause significant operational and financial disruption.

In this article, we study the invertibility of matrices of the form $A+EDF$, where $A$ and $D$ are square matrices over a unital, not necessarily commutative, ring and $A$ is singular. Under natural hypotheses on kernels and images, we prove that $A+EDF$ is invertible if and only if $D$ is invertible, and obtain an explicit formula for the inverse of $A+EDF$ by purely algebraic methods. When the coefficient ring is stably finite, or more specifically a field, we obtain an invertibility criterion under weaker hypotheses. Our results extend and sharpen a theorem of Eriksson and Nordqvist for complex matrices.

Large language models offer a promising foundation for chemical reasoning, bringing together chemical knowledge and multistep problem solving. Chemical intuition can provide an initial sense of plausible outcomes before the details of a solution are fully worked out. Inspired by how such expectations complement explicit analysis, we study how continuous latent thoughts can be trained to anticipate informative aspects of future solutions without verbalizing every intermediate step. We introduce Latent JEPA, a framework that combines autoregressive learning with joint-embedding prediction of one or more future views. For chemical reasoning, we develop textual and molecular prediction objectives that connect latent thoughts to both subsequent reasoning and molecular outcomes. Experiments on ChemCoTBench show gains in molecular optimization and on several editing and reaction metrics. Representation analyses show that future prediction makes latent thoughts more informative about molecular outcomes and strengthens their correspondence with chemical structure. These findings support abstract future prediction as a learning principle for connecting continuous latent reasoning with scientific outcomes.

The linear list Hadwiger conjecture asserts that every $K_t$-minor-free graph is $O(t)$-choosable. Following an approach outlined by Delcourt and Postle, we recently established $O(t\log\log t)$-choosability. Soon afterward, OpenAI provided a proof of the conjecture. A key new ingredient of their proof is a linear bound on the choice number of $K_t$-minor-free graphs with at most $t^{11/10}$ vertices. In this paper, we give a relatively simple proof of this small-graph bound for graphs with at most $t(\ln t)^5$ vertices.Combining this with the argument from our earlier $O(t\log\log t)$ bound, we obtain a shorter and somewhat more accessible proof of the conjecture.

We prove a generalization of Grauert's higher coherence theorem for a class of curved differential graded (non-commutative) Fréchet algebras. This allows us to extend the Fourier-Mukai calculus to derived categories arising in many new contexts. We then apply it and prove equivalence of derived categories of dual generalized complex tori using a non-commutative version of the Poincaré line bundle. This lays the foundation for categories of generalized complex branes on generalized complex tori. Examples include complex tori, symplectic tori as well as their non-commutative and B-field deformations.

Few-shot personalization enables large vision–language models (LVLMs) to learn user-specific visual concepts for applications such as personalized retrieval and subject-aware querying. However, it also creates a privacy risk: an adversary can bind a target identity from a few reference images and subsequently detect that identity in new images through natural-language queries. We introduce Anti-Persona, an image-level defense against unauthorized identity binding and recognition in personalized LVLMs. Our key insight is that identity personalization relies on visual features shared across multiple reference images. We aggregate these features into an identity prototype and optimize visually subtle perturbations that disrupt prototype alignment in the vision-encoder space. Spatial smoothing and low-frequency preservation further promote visual fidelity and practical resilience to image compression. The resulting protection does not depend on a specific prompt and supports both proactive anti-personalization and reactive image protection. Experiments on two representative personalized LVLMs demonstrate protection rates of up to $95.0\%$ while preserving visual fidelity. The method remains stable across prompt variations and evaluated identity-query tasks, and improves black-box transfer under encoder mismatch.

Robotic simulation and virtual reality increasingly require object assets that capture not only visual geometry but also the physical cues underlying tactile and thermal interaction. Existing 3D datasets and reconstruction methods primarily represent object-scale geometry and visual appearance, overlooking microscale surface structure for high-fidelity haptic rendering and transient temperature dynamics for temperature-aware interaction. We present TouchTherm, a framework for constructing simulation-ready visuo-tactile-thermal object assets from real-world objects. For visual and tactile reconstruction, we combine structured-light scanning with multiview normal maps obtained from photometric stereo. The normal maps are registered to the scanned geometry and transformed into tangent space to recover local micro-height fields for optical tactile rendering, while the coarse mesh handles collision detection. For thermal reconstruction, we capture synchronized multiview infrared videos of natural cooling following controlled heating and reconstruct a physics-regularized dynamic thermal field. Experiments on 20 objects show that the reconstructed micro-height fields preserve dominant surface structures and recover higher-frequency details beyond the coarse geometry, while the thermal fields achieve held-out surface-temperature MAEs of 0.465 degrees C and 0.592 degrees C at 30 s and 45 s, respectively. The resulting tactile assets support synthetic-to-real object recognition from tactile observations, while a glove-based VR system demonstrates spatially and temporally varying thermal feedback. These results highlight the potential of TouchTherm for multimodal sensory simulation and temperature-aware virtual interaction.

Predicting the future evolution of a scene is a fundamental capability for world modeling. Recent work has shown that operating in the feature space of Vision Foundation Models (VFMs) yields semantically rich representations that support diverse future scene understanding tasks. However, existing approaches rely on two-stage pipelines, where VFM features are first compressed using fixed dimensionality reduction (e.g., PCA) or independently trained autoencoders, and a separate predictor is trained on top of the resulting frozen latent space. This decoupling between representation learning and temporal prediction, as well as approaches that apply predictors directly on raw VFM features, provides no guarantee that the latent space is structured for predictable dynamics. In this work, we propose Latent-Foresight, an end-to-end framework that jointly learns a latent tokenizer and a flow-based generative dynamics model, explicitly shaping the representation to support temporal predictability. To enable stable joint optimization, we introduce several key design choices that prevent latent collapse and align reconstruction with generative objectives. Extensive experiments show that our approach learns more temporally coherent latent representations and consistently outperforms two-stage baselines across multiple future scene understanding tasks and prediction horizons, while eliminating separate training stages, including during high-resolution adaptation. We provide the implementation code and model weights at https://github.com/Sta8is/Latent-Foresight

Universal channel coding enables communication without a complete description of the channel. For classical channels, universal codes can attain both capacity and the optimal high-rate reliability. We show that this compatibility fails for classical-quantum channels in general; that is, the optimal reliability in the channel-aware scenario is not always achievable with universal coding due to the ignorance of the unitary rotation of the output system. We exhibit a family of classical-quantum channels for which one cannot achieve the channel-aware optimal reliability by a fixed coding scheme. We further derive a converse bound on the reliability for unitary-invariant decoders, a natural assumption for the universal coding scheme, that can be strictly smaller than the optimal channel-aware error exponent. Conversely, we construct a channel-independent encoder-decoder pair and establish a universally achievable bound on the reliability that matches this converse bound in the high-rate regime, thereby characterizing the optimal universal reliability. Specifically, the channel-aware and universal exponents are governed by the Petz and sandwiched Rényi divergences, respectively. These divergences coincide for commuting outputs but differ for noncommuting ones, explaining why universality preserves optimal reliability classically but can reduce it quantumly. Our results showcase the fundamental reliability cost of performing the classical-quantum channel coding task universally.

Given oracle access to an unknown unitary $U=e^{iH}$ , the fractional query problem asks how many queries are required to implement a noninteger power $U^t=e^{itH}$, $0<t<1$, when the spectrum is separated from the branch cut by a gap $\delta$. Quantum singular value transformation gives an upper bound of $O\!\left(\frac{1}{\delta}\log\frac{1}{\varepsilon}\right)$ queries for approximation error $\varepsilon$. We prove a matching lower bound for arbitrary query algorithms. Our argument reduces any $N$-query circuit to the approximation of $e^{it\theta}$ by a trigonometric polynomial with degree bounded by $O(N)$, together with Remez inequality. This allows us to establish the lower bound of $\Omega_\tau\!\left(\frac{1}{\delta}\log\frac{1}{\varepsilon}\right)$. Consequently, the optimal query complexity for fractional query problem is $\Theta_{\tau}\!\left(\frac{1}{\delta}\log\frac{1}{\varepsilon}\right)$, showing that the known QSVT construction is asymptotically optimal. We also give an alternative lower bound proof based on constructing a linear functional that annihilates the approximant space, yielding a $\Omega_{\tau}\!\left(\log\frac{1}{\varepsilon}\right)$ bound uniform to $\delta$.