Papers

Multimodal encoders such as CLIP underlie many downstream systems, but their web-scale training data embed harmful associations that safety alignment must suppress without unnecessarily changing benign representations. Because ethical and practical constraints prevent collecting real unsafe content at scale, existing datasets pair safe real samples with generated counterparts, but label every generated sample unsafe, even when one modality is individually safe. To address this, we introduce ShieldCLIP, the first framework to condition safety alignment on the observed safety state of each modality rather than the origin of a sample, preserving safe content while redirecting only what is unsafe. We also introduce ViSUv2, a 195k-quadruplet dataset with independent per-modality safety labels across 578 concepts and 28 categories. Using these labels, ShieldCLIP defines a four-way conditional objective beyond pair-level supervision: safe content is anchored, unsafe modalities are redirected to their safe counterparts, mixed pairs update only the unsafe branch, and coherence is enforced when both are unsafe. We evaluate ShieldCLIP on cross-modal retrieval, text-to-image generation with Stable Diffusion v1.4 and SDXL, and image-to-text generation with LLaVA. Across these settings, ShieldCLIP consistently reduces harmful outputs over prior safety-aligned encoders and strong mitigation baselines, while preserving the utility of the original embedding space. Extensive ablation studies further show that both modality-specific supervision and the selective alignment objective contribute to these gains. Source code, trained models, and ViSUv2 (under a controlled-access protocol) will be made publicly available at https://aimagelab.github.io/ShieldCLIP/.

On-policy self-distillation (OPSD) trains mathematical reasoning models using a privileged teacher that sees a reference solution and supervises student-sampled prefixes. Standard OPSD uses one fixed parameter setting at every state, but nearby settings may offer additional supervision. We find that local parameter perturbations reveal complementary reference-aligned corrections under the same reference context. Different experts supply these corrections at different reference positions. Their pool covers more such positions than the unperturbed privileged teacher. We introduce Neighborhood OPSD (N-OPSD) to turn these corrections into supervision at student-visited states. Offline, greedy selection builds a compact pool of frozen experts by rewarding filtered reference-token gains beyond the pool's current best at each position. The highest-peak expert need not provide the best training target. Online routing therefore separates the anchor direction from its level of support. MaxPeak selects the anchor token, and quantile selection chooses among experts whose top token matches it. The student learns from the chosen expert's full next-token distribution through the clipped forward-KL objective inherited from OPSD. We evaluate on AIME 2024, AIME 2025, and HMMT February 2025. Across three independent runs per method, Neighborhood OPSD improves the three-benchmark Average@12 over OPSD by 2.75, 1.67, and 1.94 points on Qwen3-1.7B, 4B, and 8B, respectively. Student-prefix continuations support using the pool beyond the reference trajectories used for selection. Matched ablations support filtered reference-token gains as a selection criterion. Accounting for overlap within the pool and routing by state further improve student accuracy. Inference uses only the distilled student.

Melting temperature calculation is one of the important topics in computational materials science. In high-throughput in silico screening and artificial intelligence assisted design of materials, it usually requires a rapid and autonomous assessment of the melting temperature of the target. Unfortunately, molecular dynamics (MD) simulations of the melting point require many cumbersome and manual operations, making large-scale calculation of the melting point challenging. In this work, we introduce MyTm, a toolkit that employs MD to automatically determine the melting point. The method is fully modularized, and by combining these modules, the program enables fully automated melting calculations by using commonly adopted approaches, including the direct-heating method, the void method, the modified void method, the solid-liquid coexistence method, and the Z method. Moreover, a machine learning (ML) method is proposed and employed to recognize and classify the solid like and liquid like atoms, which effectively resolve the low accuracy issue in conventional classification approaches, thus making the automated high throughput pipeline of melting-point calculation possible. The robustness and efficacy of MyTm have been demonstrated by several well studied systems.

Long-horizon robot manipulation reuses skills across many task compositions, but improving these compositions with additional end-to-end demonstrations is costly. A practical self-improving system must decide both what to teach next and where to apply that supervision. We present ROBOCOACH, a world-model-guided coaching framework that uses imagined failures to guide demonstration requests and expert updates. Its Route-Imagine-Diagnose-Improve (RIDI) loop executes reusable skill experts inside COACHWORLD, our shared action-conditioned world model, and uses a progress judge to record the first subtask that fails to complete. Aggregated records select which subtask demonstrations to acquire and which expert adapters to update. Across two simulation suites and two real-robot platforms, imagined and deployed success correlate over 22 task-policy pairs (rho = 0.840). Controlled comparisons show that our coaching method outperforms matched baselines under matched data budgets and update schedules. With only 150 additional subtask demonstrations, success rises from 13.3% to 75.0% on Franka and from 40.0% to 83.8% on AgileX. The coached experts also transfer to four held-out compositions, achieving an average success of 35.0%, compared with 0% for a shared-policy baseline updated with uniformly acquired demonstrations. Together, these results show that world models can serve as active coaches, turning imagined failures into targeted supervision for modular policy improvement. Project Page: https://robocoach-ai.github.io/

Privacy has become an increasingly important concern in the Document Analysis community, to the extent that in many environments such as archives, governmental institutions, and local businesses, the adoption of automation is restricted by legal and policy constraints. While federated learning has often been regarded as a “necessary evil”, implying an unavoidable performance trade-off in exchange for decentralization and privacy, many prior works overlook its potential to improve robustness to out-of-distribution data. In this paper, we present Unapologetically Distributed, the first comprehensive study evaluating distributed learning in Document Analysis along three key axes simultaneously: the tasks addressed, the architectures employed, and the fine-tuning strategies applied. Specifically, we demonstrate how various distributed training approaches enhance generalization capabilities across diverse tasks such as Table Recognition, handwriting recognition, and Word Spotting, particularly during transfer learning stages. Our results provide strong evidence that decentralization is not merely a constraint, but a valuable opportunity to improve model robustness and adaptability in real-world Document Analysis scenarios.

We establish a multiplicity-one refinement of the non-archimedean type-II Howe correspondence: whenever the small theta lift is non-zero, it occurs in the big theta lift with Jordan-Hölder multiplicity one. Furthermore, we apply this result alongside the Godement-Jacquet zeta integral to compute big theta lift for Arthur-type representations. Based on these computations, we also propose a conjecture regarding the irreducibility of the big theta lift.

Effective spatial organization of post-disaster shelter camps is essential for ensuring access to basic services while making efficient use of limited space and resources. In this paper, we propose a graph-theoretic framework for shelter-camp facility placement based on rainbow $k$-domination and Roman domination. Rainbow $k$-domination models the simultaneous accessibility of distinct facility types, such as sanitation units, kitchens, water points, and schools, whereas Roman domination is used to represent services with different capacity levels, illustrated through Wi-Fi deployment. We present an $O(nk)$-time algorithm for finding a minimum rainbow $k$-dominating set of a tree with $n$ vertices, which is linear in $n$ for fixed $k$, and computational experiments confirm its scalability on large instances. For general graphs, we establish new lower and upper bounds on the rainbow $k$-domination number. We further investigate its relationship with Roman domination, derive structural properties of graphs attaining the extremal equality between the two parameters, and prove that recognizing such graphs is NP-hard. These results provide a theoretical foundation for domination-based approaches to facility placement in post-disaster shelter planning.

Unsupervised domain adaptation (UDA) reduces the annotation burden in panoptic segmentation by leveraging a cost-effectively labeled source domain (e.g., synthetic) and an unlabeled target domain to bridge the distribution gap. Existing panoptic UDA methods rely on teacher-student consistency learning built upon suboptimal per-pixel segmentation architectures. In contrast, state-of-the-art mask transformers are rarely adopted due to their pronounced vulnerability to confirmation bias in consistency learning, where erroneous teacher predictions are reinforced during training. Our earlier approach, MC-PanDA, mitigates this issue through fine-grained confidence estimation, which suppresses gradients from unreliable masks while sampling informative yet reliable locations for loss computation. However, this method entails a complex multi-stage training and requires careful hyperparameter tuning. This work presents MC-PanDA++, which addresses these limitations by introducing: (i) self-supervised vision encoders that provide a stronger and more robust initialization, further reducing the reliance on human annotations, (ii) per-class, self-adapting mask-wide loss scaling that stabilizes training and enables the usage of a single set of hyperparameters across domains, and (iii) a single-stage training pipeline that decreases overall conceptual complexity. Together, these improvements result in a conceptually simpler, better-performing, and more robust method for domain-adaptive panoptics. Source code: https://github.com/martinovicivan/MC-PanDA

We study the critical exponent of additive codes through an integer polymatroid associated with the code. We give a coding-theoretic proof of Whittle's Critical Theorem in this setting, a geometric description of the critical exponent in terms of $h$-projective systems, and general bounds, including an analogue of Kung's girth bound. We then study additive codes whose polymatroid is the hypergraphic polymatroid of a hypergraph $H$. For these codes the critical exponent turns to be determined by the weak chromatic number of $H$. If the code is faithful, then the minimum folded Hamming weight of the dual code is equal to the Berge girth of $H$. If $H$ is connected, the minimum distance is equal to the edge-connectivity of $H$. As a consequence, for $h\geq2$ we determine all such codes with connected $H$ that attain the Singleton bound, that is, all faithful hypergraphic additive quasi-MDS codes. We specialise the Griesmer and linear programming bounds to hypergraphic codes, we derive a lower bound on the minimum distance from the Laplacian eigenvalues of the weighted $2$-section of $H$, and we compare all these bounds computationally.

Audio deepfake detection (ADD) must remain effective when new spoofing attacks emerge after deployment. Emerging audio language model (ALM)-based ADD methods are built on predefined supervision from ground-truth labels or verified forensic rationales. However, this paradigm overlooks an ALM's own mistakes, which indicate where targeted supervision is most needed. To this end, we first introduce evolving spoofing environments for ALM-based ADD, where a new attack becomes dominant while previously observed attacks persist. Motivated by the above learning-from-mistakes perspective, we further propose SE-ADD, a self-evolving framework that iteratively adapts an ALM via low-rank adaptation (LoRA) using mistake-driven supervision built from its verdicts and self-generated forensic cues. All training samples receive direct authenticity supervision, while misclassified ones receive additional cue-augmented supervision. As verdicts and cues are regenerated by the updated ALM, the resulting supervision evolves accordingly. Experiments on two ALMs demonstrate the effectiveness of SE-ADD in generalizing to unseen attacks, reducing the equal error rate (EER) from $36.72\%$ to $7.52\%$ for Qwen2-Audio and from $19.93\%$ to $3.97\%$ for MOSS-Audio.

Coding agents powered by Large Language Models (LLMs) read repository rules (e.g., AGENTS.md and Cursor rules) to learn project conventions and development procedures. An attacker can exploit the same files to mislead the agent into requesting excessive permissions or performing unrelated operations, such as accessing credentials or transferring data. We present Aletheia, a framework for permission-minimality testing of coding-agent rules. It tests whether the permissions requested in these rules are necessary to complete specified tasks. An LLM first translates the rule files into a typed permission language. Aletheia then removes one permission declaration at a time and synthesizes the corresponding restricted sandbox configuration using a trusted task environment. It executes the unchanged rule and task under both the full and restricted configurations. If both executions produce artifacts that pass the functional tests, Aletheia identifies the removed permission as unnecessary for the tested task and flags the corresponding rule file as suspicious. We evaluate Aletheia on AIShellJack, which provides 314 attack rules containing prompt-injection payloads and five benign rules, as well as 80 manually verified benign rules from GHAgentFiles. Aletheia detects all 314 attacks while reporting only three false positives among the 85 benign rules (3.5%). On average, the evaluation takes 44.35 seconds and costs US$0.22 per rule.

Finite free convolution is obtained by averaging characteristic polynomials over Haar unitary conjugation. We ask when two maximal abelian $\ast$-subalgebras of the complex matrix algebra ${\sf M}_n$ can be placed in finite free position: that is, when their relative position realizes this averaging exactly for every pair of elements, one from each subalgebra. Writing such a pair as ${\sf D}_n$ and $U{\sf D}_nU^*$ with $U$ unitary, we characterize finite free position, for either additive or multiplicative convolution, by the condition $|\det U[I,J]|^2=\binom{n}{r}^{-1}$ for every $1\le r\le n$ and all $I,J$ with $|I|=|J|=r$. We show that this condition holds if and only if $n\le3$ and $\sqrt{n} U$ is a complex Hadamard matrix. To quantify the failure of exact realization for $n\ge 4$, we introduce the uniform-minor discrepancy $\delta_r(U)$. We identify it with the mean-square error in the $r$-th coefficient of finite free multiplicative convolution for two diagonal matrices whose diagonal entries are independent and uniformly distributed on the unit circle. We establish the symmetry $\delta_r(U)=\delta_{n-r}(U)$ and the monotonicity $\delta_1(U)\le\delta_2(U)\le\cdots\le \delta_{\lfloor n/2\rfloor}(U)$. For flat unitaries, we derive an explicit formula for $\delta_2$, yielding $\delta_r(U)\ge\frac{n-3}{2n}$ for $2\le r\le n-2$. Equality for $r=2$ holds precisely when the entrywise square of $\sqrt{n}U$ is also complex Hadamard.

We disprove a conjecture of Liu, Luo, and Luo (2020) concerning the Landau radius for bounded planar harmonic mappings. We construct an explicit harmonic mapping $F_M$ satisfying the standard normalization conditions, which fails to be locally univalent strictly within the classical holomorphic Landau disk for all $M > M^* \approx 4.451$. By combining this counterexample with the lower bound established by Ponnusamy, Kalaj, and Vuorinen (2014), we demonstrate that the exact asymptotics of the harmonic Landau radius is $\frac{\pi}{8M}$.

For any $t\in [0,1/2)$, consider the process $\eta_t:[0,1/2]\mapsto\mathbb{R}$ defined by $\eta_t(s)=\Pi(t+s)-\Pi(t)$, where $\Pi$ is the directed geodesic from $(0,0)$ to $(0,1)$ in the directed landscape. Let $0\leq t<u<1/2$. We show that the laws of $\eta_t$ and $\eta_u$ are mutually absolutely continuous if and only if $t>0$. This proves Conjecture 14.5 in Dauvergne, Ortmann, and Virág, 2022 in the affirmative.

We establish sufficient conditions for preparing quantum thermal states of noncommuting local Hamiltonians with polylogarithmic circuit depth in arbitrary fixed spatial dimension. Our conditions combine locality and stability bounds on the effective interactions of reduced density matrices of a Gibbs state with a quantitative high-temperature condition and access to coarse classical reference Hamiltonians. When these references can be generated locally, the classical preprocessing cost is $N^{1+o(1)}$ for $N$ sites at inverse-polynomial global accuracy. We construct the preparation circuit from spatially localized Petz recovery maps. Quantum corrections derived from the microscopic Hamiltonian enable accurate recovery while the classical references remain coarse. After preprocessing, the resulting circuit prepares the canonical purification using $N\operatorname{polylog}(N/\varepsilon)$ gates and qubits, where $\varepsilon$ is the preparation error. These results link two fundamental questions: how correlations are organized in thermal equilibrium, and how efficiently the corresponding states can be realized through operations allowed by quantum mechanics. By translating static equilibrium structure into explicit preparation circuits, they give equilibrium locality a constructive computational interpretation and a physically grounded role in quantum algorithm design.

Can a pretrained graph model replace training and tuning a separate predictor for each dataset? Answering this requires evaluating prediction quality alongside computational cost. We present NodeGround, a node classification benchmark that puts graph foundation models (GFMs) and dataset-specific supervised learning under a common evaluation framework. The benchmark spans 51 datasets and evaluates six GFMs alongside 15 supervised methods under two label-availability regimes. Shared data partitions, validation-only model selection, controlled hyperparameter searches, and multiple predictive metrics make comparisons systematic, while workflow measurements account for adaptation, training, tuning, and inference. The results favor carefully tuned graph neural networks overall. GraphPFN reaches third place by Elo when more labels are available, yet its relative strengths vary substantially with dataset properties. Efficiency comparisons further qualify the benefits of pretrained reuse: GVT and GraphPFN appear on the Pareto frontiers when supervised methods are represented by their default and fully tuned configurations. Adding intermediate tuning budgets removes this advantage for GVT and leaves GraphPFN extending the estimated frontier in the label-rich setting alone. Thus, reusing pretrained parameters does not yet provide a broadly reliable route to either stronger predictions or cheaper workflows. We release the evaluation pipeline, run-level records, and an open leaderboard at https://github.com/nums-ai/nodeground.

We investigate the scattering and diffraction of scalar plane waves by static conformal anomaly black holes. We determine the truncation requirements for the partial-wave series (PWS) method at finite distances and compute the full waveforms, which display a clear diffraction pattern: on the negative $z$-axis the field reduces to the incident plane wave, while on the positive $z$-axis a distorted plane wave coexists with a scattered spherical wave, and a bright Poisson spot appears at $\theta=0$ surrounded by concentric diffraction rings. We obtain the on-axis intensity of the Poisson spot as a function of the radial coordinate. The intensity is most sensitive to the near-horizon geometry: close to the horizon it differs markedly from that of the RN black hole, whereas in the far field the two curves converge. We trace this to the metric functions, where different values of $f'(r_+)$ produce different phase increments that compress and shift the interference fringes, while the conformal anomaly correction falls off as $O(\tilde{\alpha}M^2/r^4)$. By analyzing the potential barrier we study the absorption cross section, and use the series reduction method to accelerate the PWS convergence for the differential scattering cross section. The low- and high-frequency absorption cross sections are strongly correlated with the $\ell=0$ and $\ell$-dependent parts of the potential barrier, respectively. The differential cross section and glory scattering arise mainly from scattering in a finite region excluding a small neighborhood of the outer horizon. Charge and $\tilde{\alpha}$ have opposite effects on the width of the glory peak but the same effect on its height.

Classical phase-field models typically use the Allen–Cahn evolution law to model interface motion as a consequence of energy-gradient descent. This implies a linear kinetic relation between the velocity of an interface (e.g., a domain wall in a ferroelectric) and its driving force. When nonlinear kinetic relations are to be modeled, as commonly found in ferroelectric ceramics, an alternative evolution law and hence an alternative model is needed. Here, we propose an electromechanically (stress-driven) coupled multiphase-field framework with a general kinetic formulation, which integrates a prescribed kinetic relation directly into the evolution law. We show that this model correctly evolves domain walls with the assigned nonlinear kinetics, e.g., of mixed exponential-power law type, following the so-called Merz–Stadler law. The multiphase formulation further enables distinct kinetic relations and interfacial energies to be assigned to different order-parameter pairs, which reflects the distinguishable properties of the different types of ferroelectric domain walls. The proposed framework is broadly applicable for simulating kinetic relations in materials with applications beyond ferroelectrics.

Vision-language-action (VLA) models enable task-conditioned interaction, but extending them to scene-scale aerial manipulation remains challenging due to costly whole-body demonstrations, latency-induced action-state misalignment, and cross-site behavior composition. We present a unified framework for synthetic policy training and scene-scale execution on articulated uncrewed aerial manipulators (UAMs). A scene-reconfigurable pipeline synthesizes task-conditioned, kinodynamically feasible trajectories and synchronized multiview observations for VLA training without physical-platform demonstrations. Measured-progress-aligned realization (MPAR) aligns asynchronously returned action chunks with measured execution progress and realizes them as continuous, dynamically feasible trajectories. A relational Scene Graph grounds language goals to object instances and feasible interaction regions, while topology-guided transfer connects local behaviors across sites. Local VLA skills achieve 39/60 successes (65.0%) in simulation under oracle target and feasible-handoff conditions. Under 500-ms added latency, with and without a transient command-update stall, MPAR reduces median takeover phase error by 0.212 s over nominal-time alignment. The complete system completes 21/50 simulated multi-site missions (42.0%) and is further validated on a physical articulated UAM.

We prove that every bipartite state $\rho$ of a qubit coupled to an $n$-level system whose operator Schmidt rank equals three can be written as a mixture of $\text{rank}(\rho)$ pure product states, and as a mixture of at most $n+1$ mixed product states in general. The second result is tight in the sense that there exist states that cannot be decomposed with a smaller number of terms. Our proof gives constructive methods to obtain the decompositions. It involves eigendecompositions of suitable unitary dilations of an operator explicitly determined by the state. Our framework recovers previously known results for the length of decompositions into pure product states and is also able to deal with decompositions into mixed states. For pure product states, our decomposition is different from existing ones.

In $\mathsf{QMA}(2)$, two computationally unbounded provers submit unentangled quantum witnesses to a polynomial-time quantum verifier. We ask whether allowing limited shared entanglement between the two provers changes the power of $\mathsf{QMA}(2)$. To this end, we formalize $\mathsf{QMA}(2;h)$, a variant in which the provers initially share $h$ EPR pairs, as suggested by Aaronson, Beigi, Drucker, Fefferman, and Shor (ToC~2009). We show that polynomially many shared EPR pairs give the provers no cheating advantage: $\mathsf{QMA}(2)\subseteq\mathsf{QMA}(2;h)$ for $h=\operatorname{poly}(n)$. For logarithmically many pairs, the power of $\mathsf{QMA}(2)$ remains unchanged, strengthening this inclusion to $\mathsf{QMA}(2)=\mathsf{QMA}(2;h)$ for $h=O(\log n)$. Finally, we introduce a variant that also allows classical communication between the provers when preparing the witnesses. This setting is particularly revealing for logarithmic-size witnesses: assuming $\mathsf{NP}\nsubseteq\mathsf{BQP}$, we show that superlogarithmically many shared EPR pairs strictly reduce the model's computational power.

In this paper, we study the exact distributions of the age of information (AoI) and peak AoI (PAoI) in a bufferless setting in which time-stamped updates, or processed tasks, generated by one or several information sources according to a Poisson process for each source, are submitted to a shared pool of $c$ homogeneous servers with exponentially distributed service times, i.e., the so-called M/M/c/c or the Erlang loss system. We consider three update management policies which come into play when a new update arrives to find all the servers busy: the arriving update is blocked (non-preemptive, NP) as in the Erlang loss system, or it preempts a randomly chosen update of its own source in service (preempt at random, PR), or the stalest such update (preempt the stalest, PS). All three policies preserve the same birth-death structure of the server occupancy process associated with the underlying Erlang loss system. However, their AoI and PAoI distributions can be very different. The approach we take is the absorbing Markov chain (AMC) method, in which a single AoI cycle, rather than the entire sample path of the system, is modeled by an absorbing Markov chain. Via the AMC method, we derive the exact distributions of AoI and PAoI in matrix-exponential form. The method extends to multiple sources sharing the same pool of servers, with the AoI of a given source being affected by the remaining sources only through their aggregate update rate. Numerical examples illustrate the implications of our findings, including distribution-based server provisioning under age violation constraints.

We study Connected Dominating Set on graphs whose closed-neighborhood set systems are $d$-semi-ladder-free. This structural condition strictly generalizes the biclique-free setting and provides a natural regime for connectivity-constrained domination. We obtain both a fixed-parameter algorithm and an approximate kernelization framework for the problem on this class. Our algorithmic result is based on a new compact representation theorem for inclusion-wise minimal set covers in $d$-semi-ladder-free set systems. Although the number of minimal set covers of size at most $k$ may be as large as $n^{\Omega(k)}$, we show that all such set covers can nevertheless be encoded by a family of at most $k^{kd+1}$ tuples, and that this family can be enumerated in time $\Oh(k^{kd+2}\cdot nm)$. Combining this representation with a Group Steiner Tree subroutine, we obtain an algorithm for Connected Set Cover, which in turn yields an algorithm for Connected Dominating Set running in time $k^{kd+2}\cdot 2^k \cdot n^{\Oh(1)}$ and polynomial space. For the preprocessing result, we introduce grouped domination cores and dominator cores, and prove polynomial upper bounds on their sizes in $d$-semi-ladder-free graphs. Using these structures, we obtain, for every fixed $d$ and $\varepsilon>0$, a polynomial-time $(1+\varepsilon)$-lossy compression for Connected Dominating Set to an equivalent reduced instance of size $k^{\Oh(d^2/\varepsilon)}$. The reduced instance is a Connected Dominating Set instance on a $(d+2)$-semi-ladder-free graph.

Embodied agents must determine where to act, anticipate the resulting scene changes, and interpret observed outcomes to guide subsequent actions. This requires connecting 4D interaction understanding, which explains how past actions changed the scene, with spatially grounded planning, which determines how and where to act toward a goal and anticipates the resulting scene changes. We introduce ChronoGraph, a functional 4D scene graph that links actions on affordance parts to semantic and geometric state changes. By representing observed and anticipated transitions in the same form, it provides a shared basis for understanding and planning. We construct ChronoGraphBench through an automatic data engine that converts human-interaction videos and simulated robot trajectories into graph-annotated questions for training and evaluating Vision-Language Models (VLMs) on both tasks. Using these annotations, we train ChronoGraphVLM by adapting pretrained VLMs in two stages. Graph-as-Chain-of-Thought supervised fine-tuning teaches the models to reconstruct observed transitions and predict future ones as graph traces before answering. Subsequent joint 4D graph reinforcement learning directly rewards graph properties and answer correctness. Experiments across model scales show improvements over the corresponding pretrained baselines and zero-shot transfer to VLM4D. Real-world demonstrations further show that graph-based planning and affordance grounding support mobile manipulation through existing robot skills without additional fine-tuning.

We construct spherically symmetric thin shells in unimodular gravity (UG) and analyze their dynamical stability under radial, symmetry-preserving perturbations. We use the UG junction conditions and adopt non-conservative charged exterior solutions sourced by a radial electric field. As an application, we build charged vacuum bubbles: a Minkowski interior surrounded by a charged thin shell and matched to an exterior UG geometry. The resulting spacetimes are free of event horizons and central singularities, while in de Sitter-like cases a cosmological horizon may exist outside the shell. We compare the stability regions and the matter content with their counterparts in general relativity, finding some differences controlled by the UG non-conservation parameter. This work extends previously published results.

The parabolic Anderson model is the Cauchy problem for the heat equation on the integer lattice with a random potential $\xi$. We consider the case where $\{\xi(z): z\in \mathbb{Z}^d\}$ are independent and identically distributed Pareto random variables with parameter $\alpha$, and assume that the solution is initially localised at the origin. We establish the full asymptotic behaviour of the total mass of the solution as time tends to infinity in the heaviest-tailed regime $\alpha\in(d,2d)$. In particular, we find qualitatively different behaviour in dimension $d=1$ and in dimensions $d\ge 2$.

Vision-language models (VLMs) are increasingly used for AI-generated image (AIGI) detection, providing natural-language explanations for authenticity judgments. However, their ability to interpret text within images may also expose these judgments to misleading semantic cues. We systematically evaluate typographic attack strategies across detection-oriented, open-weight, and commercial VLMs, considering both real-to-fake and fake-to-real attacks. Our results show that reasoning modes generally exhibit greater vulnerability than direct modes and that attack effectiveness exhibits pronounced directional asymmetry. Moreover, larger models tend to exhibit higher clean detection accuracy but also higher attack success rates. We further examine attack robustness under image and text transformations and investigate whether overlays indicating the correct class can aid error correction. Together, these analyses characterize how typographic attacks influence authenticity judgments and expose limitations of current VLM-based AIGI detection systems.

Self-attention gives LLMs fine-grained, query-dependent access to context, but dense token interactions incur quadratic prefill cost and a key–value cache growing with context length. Research thus spans explicit-memory compression, sparse access, recurrent state construction, structured state dynamics, and heterogeneous mechanism composition. This survey analyzes these developments as model-internal contextual memory. We introduce a five-dimensional lens—Memory Representation, Memory Update, Access, Readout, and Integration—describing what is represented, how it changes, what is query-eligible, how it is read, and how readouts form outputs. This lens compares overlapping research lines without imposing one computational model. We reconstruct mechanism-level developments and architectural adoption using 59 release-level records from 14 major model lineages and 11 high-performing open-weight endpoints. First, explicit-memory and recurrent-state methods retain distinct interfaces but increasingly control overlapping memory functions. Second, heterogeneous architectures increasingly coordinate across network depth: layer-wise composition distributes complementary memory processing across representational stages, while cross-layer reuse carries selected memory and routing artifacts forward. Depth thus becomes a dimension along which contextual memory is constructed and managed. Third, these developments motivate a stateful multidimensional memory-routing hypothesis: persistent memory is organized across temporal scope, network depth, substrate type, and representation granularity, while coordinated Sparse Write and Sparse Read determine what is maintained and what contributes to each query. Overall, efficient sequence architecture design increasingly concerns the organization, lifecycle, and selective use of contextual memory rather than an isolated Attention operator.

Generative models are transforming Bayesian computational imaging, yet the field still lacks physics-aware foundation models. Current practice falls into two camps. Large foundation image models are deployed as plug-and-play priors with zero-shot approximate likelihood guidance, which introduces significant bias and computational cost. Physics-aware generative models avoid this bias, but each is tied to a specific dataset, task and instrument. We introduce BAM (Bayesian Anything Model), a lightweight foundation model for few-step, physics-aware posterior sampling that generalises robustly to unseen data and tasks, zero-shot or with minimal finetuning. BAM upgrades the operator-conditioned Reconstruct Anything Model (RAM) backbone (Terris et al.) into a conditional flow map, so instrument physics is specified at inference time rather than fixed during training. BAM has just 36M parameters and is pre-trained jointly on large image corpora and libraries of forward operators. A single network then draws posterior samples in a few steps, with no likelihood approximation and no guidance weights to tune. Across linear inverse problems on FFHQ, AFHQ, LSUN, DIV2K and the Kohler camera-shake benchmark, BAM outperforms in just 3 steps both specialised models and leading zero-shot methods in sample quality, at a fraction of their computational cost. BAM gives the community an accessible entry point to generative computational imaging, lowers the economic and environmental cost of training imaging models, and opens a new path for research on physics-aware Bayesian computational imaging. Official page: https://bayesian-anything-model.github.io/

For Hermitian matrices $A$ and $B$, we prove that $(A\join B)\oplus(A\meet B)$ is unitarily equivalent to $A\oplus B$, where the join and meet are taken in Olson's spectral order. Taking traces answers a question of Bourin and Lee: for positive semidefinite matrices, $\Tr(A\join B)=\Tr(A+B)$ if and only if $A\meet B=0$. Iterating the direct-sum identity gives a trace formula for a finite family of positive matrices. The trace of the spectral supremum equals the trace of the sum precisely when the ranges form an algebraic direct sum. For each fixed $1<p<\infty$, the spectral supremum and the sum have equal Schatten $p$-norms if and only if the ranges are pairwise orthogonal. Finally we establish an elegant formula for the Frobenius inner product and the spectral order.