Papers

Mn$_3$Sn exhibits a complex magnetic phase diagram with multiple symmetry-breaking phenomena. Minute electron doping at the Mn site breaks the kagome lattice symmetry in Mn$_3$Sn. In this work, we exploited this to tune the order of the magnetic phase transition using magnetocaloric effect in Mn$_{3-x}$Co$_x$Sn single crystals. Pristine Mn$_3$Sn shows metallic behaviour with a spin-reorientation transition at 278 K, which shifts to 175 K upon Co doping. Additionally, Co doping also induces a weak-ferromagnetic transition near $\approx 210$ K. The effect of Co doping results inverse to conventional magnetocaloric effect, with enhanced magnetic entropy change. Further, the scaled entropy change reveals a field-driven magnetic phase transition crossover from first order to second order in Mn$_{2.75}$Co$_{0.25}$Sn crystal. Based on experimental data, a new phase diagram is constructed, depicting the field evolution of various magnetic orders.

Active matter systems composed of self-propelled entities exhibit a wide range of out-of-equilibrium phase behaviors and collective dynamics. Assemblies of active particles, e.g., in a linear structure, are particularly interesting, as they exhibit an intriguing interplay between activity and geometrical restrictions due to self-avoidance and entanglements. Using Langevin dynamics simulations of an active Brownian ring polymer with translational and rotational inertia, we uncover an inertia-driven re-entrant coil-globule transition in the absence of any attractive interactions. We show that rotational inertia plays a paramount role, leading, in the globular state, to spontaneous, nonzero, radially outward polarization of the active forces. Self-avoidance and entanglements break the rotational symmetry of active beads and effectively stabilize the globular states. The transition becomes increasingly sharp with increasing rotational inertia, yielding highly compact globular conformations at intermediate activities, whereas at high activity the polymer re-expands into a coil state. Scaling of the radius of gyration reveals crossovers from self-avoiding, phantom-like to compact-globular structures. Further, rotational inertia enhances transport properties, with an intermediate activity-induced Rouse-type bead diffusion and a long-time diffusion coefficient displaying a power-law dependence on the moment of inertia. Our study establishes a purely inertia-induced mechanism for controlling the collapse of active polymers, arising from spontaneous active-force polarization and self-avoidance.

Existing cumulative regret guarantees for Bayesian optimization (BO) do not necessarily characterize the quality of the algorithm's final output. This gap is critical under noisy evaluations, where the goal is to return a single high-quality recommendation. To address it, we introduce efficient exploration strategies that can be incorporated into any BO procedure by occasionally replacing the acquisition query with an exploration query. To validate this approach, we bound the Gaussian process posterior standard deviation over the search space via the fill distance of the queries, which measures the largest region left unexplored. We then use this result to derive upper bounds on the inference regret, i.e., the gap between the global optimum and the function value at the algorithm's returned point. Empirically, these exploration strategies improve the final recommendations of existing BO algorithms on synthetic benchmarks, hyperparameter optimization, and a control task.

This paper presents a Mixed Reality (MR) robotics experience that provides an accessible introduction to interacting with a physical robot. The application combines a TurtleBot with a virtual animal-tending activity and three control methods: Joystick, Waypoint, and Path-Drawing. A mixed-methods user study with 10 participants investigated changes in robotics interest and confidence following the experience and how the control methods influenced engagement, workload, and preference. Confidence increased following the experience, with the exploratory analysis indicating a significant pre–post difference, while the increase in interest was not statistically significant. The control methods produced clear differences in perceived workload and ease of use, with Path-Drawing generally more demanding and difficult to use. Joystick was preferred by eight of ten participants and was associated with better perceived performance and lower frustration. However, no statistically significant differences were detected in the other engagement dimensions, suggesting that the most preferred method was not necessarily more engaging across all aspects of the experience. These findings highlight the importance of considering both usability and engagement when designing how users interact with robots in introductory MR experiences.

We study the price of anarchy (PoA) in nonatomic routing games as a function of the full vector of origin–destination demands. We show that the Beckmann value function is continuously differentiable on the positive demand cone, with gradient equal to the equilibrium origin–destination cost vector; hence the PoA is continuously differentiable on the regular demand region. We prove that the nondifferentiability set has Lebesgue measure zero and, when every edge cost is strictly increasing or constant, is covered by countably many codimension-one submanifolds. Under the same cost structure, all equilibrium objects are cellwise as smooth as the edge costs. For affine costs, the exceptional sets become finite unions of lower-dimensional polyhedral cells linked by an exact scaling law, and the maximum of the PoA over a cell closure is attained on the cell boundary, generalizing the one-dimensional endpoint result. The theory is illustrated on the Sioux Falls benchmark network.

The upcoming sixth generation (6G) wireless networks are designed to support advanced applications that the current air interface based on orthogonal frequency division multiplexing (OFDM) cannot accommodate, necessitating innovations in waveform design. Orthogonal time frequency space (OTFS) has emerged as a promising alternative, offering enhanced performance particularly in high-mobility channel environments. However, deploying fundamentally different waveforms over the same frequencies fails to exploit the mutual relationships between them. In this paper, a forward-compatible yet backward-compatible coexistence framework is proposed, where various OFDM waveform options are incorporated while seamlessly integrating OTFS within the same network. The proposed coexistence paradigm involves utilizing a specially designed spreading matrix in the delay-Doppler (DD) domain to process OTFS data. The matrix is tailored to establish orthogonality between OTFS and OFDM waveforms, effectively reducing cross-interference when both coexist within the same time-frequency (TF) resources. Our results demonstrate up to a 6 dB improvement in bit error rate (BER) performance for the OTFS waveform compared to other coexistence schemes. The proposed approach also shows greater interference resilience and approaches the upper bounds of achievable rates. By utilizing the full bandwidth and time duration, it enhances range and velocity resolutions, leading to improved sensing performance. These findings are supported by mathematical analysis and extensive simulations for uplink scenarios.

Flat bands amplify interactions into correlated phases, and topological bound states are robust by principle; combining the two in a model that is both exactly analyzable and experimentally accessible is therefore especially desirable. Here we show that a square-lattice model with a slowly modulated Dirac mass texture, $m(x)=m+m_0\cos(2\pi x/L)$ with $L$ up to hundreds of sites, provides such a setting, without engineered geometries, lattice-scale textures, or long-range hoppings. Exact nesting reduces the problem, independently at each transverse momentum $k_y$, to a $1+1$ dimensional Dirac Hamiltonian with a position-dependent mass: every simple zero of $m(x)$ binds Jackiw-Rebbi zero modes, and the two walls of a period form an effective Su-Schrieffer-Heeger ring. The model interpolates between a chirally pinned flat band at large $L$, a helical miniband whose nodes are enforced by a boundary phase of the ring, and a bias-dimerized regime in which the sign of the bias sets a winding number. A central result is that the miniband bandwidth is not a single scale: a max-min measurement conflates three widths of different origin, the helical height $W_{\rm JR}$ exponential in the period, the zone-edge upturn $W_{\rm edge}\propto L^{-2/3}$ dominating at large $L$, and a bias bump exponential in $L$, so the flat band is robust against bias in the thermodynamic limit. We analyze two lattice realizations, a two-orbital model with modulated inter-orbital hybridization and a single-orbital bond-order texture sharing the nodal physics while leaving the van Hove points gapless, and propose experiments on cold-atom superlattices, Kekule-ordered graphene, STM designer lattices, and microwave arrays.

Flagship large language models (LLMs) offer rich general knowledge and strong reasoning capabilities, but their high inference cost and latency limit their use for large-scale data labeling. We study zero-shot text classification with semantically meaningful labels and task descriptions, without task-specific labeled examples or classifier feedback. We propose SLKR (Semantic Label Knowledge Retrieval), a label-centric retrieval-augmented generation framework that separates offline knowledge construction from lightweight online inference. Offline, a flagship LLM independently generates label definitions, retrieval keywords, and boundary rules for potentially confusable label pairs, forming a reusable label-knowledge store. Because this construction operates over the label space rather than the input corpus, its one-time flagship cost can be amortized across large-scale inference. Online, hybrid semantic and lexical retrieval first identifies the top-$K$ candidate labels and then selects up to $M$ input-relevant boundary rules connecting candidate-label pairs. A lightweight LLM performs the final classification in a single call using the retrieved definitions and rules. Across three public benchmarks and a real-world multilingual book-classification dataset, SLKR achieves the best results on all four datasets with Qwen3.5-4B and on three of four with Ministral-3-3B, while using compact input contexts and no task-specific labeled examples. Boundary ablations further show that retrieving a small set of input-relevant rules preserves performance close to using all boundaries with substantially less context, supporting efficient reuse of flagship-generated label knowledge for lightweight classification.

The spectral form factor (SFF), which is the Fourier transform of the two-point spectral correlation, is a fundamental diagnostics of quantum chaos. Nevertheless, its experimental detection in quantum many-body systems was less unexplored until recently. Here, we establish a relationship between the SFF with the quantum Fisher information (QFI) by means of quantum sensing protocols. Furthermore, when the eigenvalues and eigenstates of the disordered Hamiltonian are statistically independent, we prove that QFI is typical both in short and long timescales while numerics suggest that it is also typical at intermediate timescales, which ensures self-averaging of QFI. The self-averaging property renders the experimental measurement of QFI with our protocol less noisy in contrast with the direct measurement of SFF which is known to be non-self-averaging. Our results reveal an intriguing connection between quantum noise and the signatures of quantum chaos.

We prove a regularity lemma for polynomials with fixed degree $d$ and any number of standard Gaussian input variables. Following modern approaches to the Szemerédi regularity lemma, we deduce our results from a general formulation of the regularity lemma using ultraproducts. Along the way, we give a direct combinatorial proof of the celebrated “fourth moment theorem” of Nualart and Peccati, which characterizes the asymptotically Gaussian polynomials inside degree-$d$ Gaussian chaos as those such that $\mathbb{E} p^2 \to 1$ and $\mathbb{E} p^4 \to 3$.

Latte art is a complex task requiring dexterity and precision, making it a challenging and commercially relevant task. A one-shot learning system is employed in which a single barista demonstration is segmented into five functionally distinct phases, each encoded as an independent Cartesian Dynamic Movement Primitive (CDMP) with its own parameters. Against a single-CDMP baseline, per-phase encoding reduces transition endpoint error by 3.2-4.5 times and art-stroke shape error by 2.7 times, reproducing jug tilt to below 1 degree RMS in the pouring phases. The learnt pour scales from a 280mL to a 400mL cup through a single spatial offset and a scaled recipe, with no re-demonstration. The phase-based CDMP approach generalises beyond latte art to any demonstration task whose phases have distinct spatial or temporal requirements whilst demonstrating the ability to adapt reproductions to suit different task requirements.

In this paper, introduce a new finiteness condition for commutative rings, called $Z$-Noetherian, under which every ideal contained in the set of zero-divisors $Z(R)$ is finitely generated. This condition lies properly between Noetherian and $Nil_{*}$-Noetherian. We establish intrinsic characterizations of $Z$-Noetherian rings and investigate their stability under nilpotent quotients and localizations. Under the assumption that $Z(R)$ is an ideal, we obtain a primary decomposition theorem for ideals contained in $Z(R)$. For $ZD$-rings, we prove a Cohen-type criterion characterizing $Z$-Noetherian in terms of the finite generation of prime ideals contained in $Z(R)$. We further establish $Z$-Noetherian analogues of the Hilbert Basis Theorem and Krull's Intersection Theorem. Finally, we establish a principal ideal theorem for $Z$-Noetherian rings. Collectively, these results show that several fundamental finiteness and structural properties of Noetherian rings extend to the broader class of $Z$-Noetherian rings.

Open-vocabulary 3D Gaussian segmentation typically binds semantic features to the per-Gaussian state via per-scene gradient training (1-4 h) or the more recent closed-form distillation. Both commit each primitive to a single feature; therefore, boundary Gaussians spanning multiple 2D instances are forced into a hard decision that degrades retrieval accuracy and yields unstable mask boundaries at render time. We present DIGS, a training-free framework with two stages, class-agnostic mapping and multi-modal encoding, that decouples instance discovery from linguistic naming. At the Gaussian level, our distribution-informed representation maintains an explicit top-$M$ label distribution per primitive, updated by a tempered Dirichlet rule; preservation of multi-peaked label distributions at boundary primitives allows a late-argmax renderer to recover clean instance boundaries. A bidirectional state machine aggregates these distributions into a persistent class-agnostic instance ontology. The instance-level encoding stage then fuses the CLIP visual embedding of each instance's isolated 3D footprint with the CLIP-text embedding of an LLM-generated attribute description, disambiguating visually similar instances. On a single A100, DIGS completes mapping and instance encoding on LERF-OVS in 1.0-3.2 min per scene, depending on whether LLM fusion is enabled, and serves native 3D queries in 5.4 ms. It reaches 71.95% mean mIoU on LERF-OVS, outperforming both the training-free SFS and training-based LaGa baselines. DIGS also achieves the highest mean mIoU on LERF-Mask and 3D-OVS and surpasses Gaga and ObjectGS on Replica under their respective evaluation protocols. Code is available at https://github.com/ZephyrZos/DIGS.

Situated 3D Question Answering requires models to interpret a 3D scene relative to an agent's pose. Existing methods commonly convey that pose through a natural-language situation description, which may be unavailable or costly to obtain in real-world deployments. While some methods additionally incorporate numerical pose, we find that randomly shuffling the agent pose causes little performance degradation, indicating limited functional dependence on the numerical pose. To address this gap, we introduce EgoAlign, a simple 3D language model that represents a scene in an agent-centric coordinate frame, together with a dedicated egocentric alignment stage comprising four complementary tasks: object category recognition, egocentric direction prediction, metric distance estimation, and spatial relation prediction. This stage teaches the model to consistently interpret scene geometry relative to the agent's pose before task-specific fine-tuning. EgoAlign achieves the strongest performance on SQA3D, Real-3DQA, and MSQA among compared methods while remaining highly efficient. More importantly, removing textual situation descriptions yields comparable or better downstream performance, whereas shuffling EgoAlign's numerical pose causes substantial degradation. These results show that EgoAlign establishes agent-centric spatial context directly from numerical pose without requiring textual situation descriptions.

Real-world studies (RWS) analyze routinely collected healthcare data, such as electronic health records and insurance claims, to estimate treatment effects, monitor safety, and identify risk factors outside randomized trials. Automating RWS with large language model agents is appealing but brittle: valid analyses depend on dataset-specific clinical semantics, including index dates, follow-up windows, patient/event-level units, and outcome boundaries that are often implicit in observational data. This leads to schema mismatches, temporal errors, inconsistent analytical granularity, and unsupported clinical claims. We present ProCARE, a profile-grounded framework that reformulates RWS automation as constrained compilation. ProCARE converts heterogeneous data profiles into Evidence Contracts specifying available variables, temporal attributes, distributions, and evidentiary limits, and compiles study objectives into RWS-IR, a strongly typed intermediate representation for variables, time relations, analysis units, and statistical tasks. Deterministic validation and localized repair then check and revise generated workflows across schema mapping, cohort construction, variable derivation, analysis, and interpretation. On 36 EHR-based benchmark tasks, ProCARE outperforms strong LLM-agent baselines in code executability, semantic consistency, and conclusion traceability, with automated violation diagnoses showing moderate-to-substantial agreement with blinded expert review.

On a complete noncompact Kähler manifold $M^n$ of complex dimension $n$ with nonnegative Ricci curvature, we prove that polynomial growth holomorphic functions of degree $d$ ($d\geq 0$) have an dimension upper bound $C(n)(1+d)^n$. In particular, $\dim_{\C}\OO_d(M)\leq C'(n)d^n,$ for $d\geq1$, and the estimate is sharp.

Parallel wire-driven systems are a framework in which multiple wires are connected to an end effector, and the end effector is driven by the tensions of those wires. Because they can transmit large forces, up to the breaking strength of the wires, over lengths on the order of several tens of meters, they can realize powerful actuation in large workspaces. In recent years, studies have also investigated connecting wires to external environments such as trees and columns and forming parallel wire actuation on site. However, conventional parallel wire-driven systems often assume a single end effector, and because of interference between wires and links in a shared workspace, it has been difficult to handle multiple floating links simultaneously. In this study, we realize multi-floating-link parallel wire actuation by introducing an inter-body wire. In real-robot experiments, the two bodies supported and lifted a mannequin and transferred it onto a stretcher as a rescue motion. This shows that parallel wire actuation can be extended from a single end effector to a multi-floating-link system, and indicates its potential for broader tasks such as the manipulation of large objects and rescue support.

Recent multimodal large language models (MLLMs) achieve strong results on video question answering, yet often fail to recover the temporal structure that makes a video coherent. They may merge neighboring events, miss single-event progression, or infer implausible relations across events. We study structured temporal understanding as the ability to recover event structure from visual evidence, including event boundaries, single-event progression, and multi-event relations. We argue that this gap arises from a mismatch between current post-training objectives and temporal reasoning, where caption-style supervision allows models to rely on language shortcuts rather than visual temporal grounding. To address this, we propose EventLens, a vision-centric reinforcement learning framework that learns video understanding through event-structure recovery. Instead of reconstructing captions, EventLens trains models to recover temporal structure under controlled perturbations, such as altered temporal granularity, reversed progression, and shuffled event order. We instantiate this framework with a three-level temporal ontology and derive three verifiable task families: event segmentation, progression discrimination, and multi-event relation reasoning. These tasks admit deterministic rewards, enabling task-specialized GRPO training without human preference labels or learned reward models. We further introduce multi-teacher on-policy distillation (MT-OPD) to consolidate specialized policies into a unified model. Experiments on Qwen3-VL backbones show consistent improvements on temporal grounding and event-sensitive reasoning benchmarks, while preserving general video-QA performance. Learning curves further demonstrate that EventLens achieves comparable downstream transfer with significantly reduced training cost compared to direct mixed RL optimization.

This manuscript presents DeepThink, a knowledge-addressable framework for recurrent language modeling. A Knowledge Bank stores typed Knowledge Units, a Knowledge Activator selects units from the current execution state, Knowledge Assembly produces an executable configuration, and a Shared Thinking Operator updates that state. An output stage then organizes generation. This organization exposes knowledge capacity, activation width, and recurrent depth as separate design variables, supporting both complete parameter packages and fine-grained expert or low-rank modules. We formalize the interfaces and their execution semantics, including state-dependent selection, parameter-space assembly, and thinking followed by generation. For compatible templates and complete runtime-state configurations, we establish an exact Fixed-Schedule Realization of an existing fixed-depth model. Exact shared-base reparameterization and pretrained output-tail reuse provide complementary initialization constructions. Together, these results define a route from pretrained models to dynamically selected and assembled recurrent computation. A synthesis of conditional computation, recurrent models, modular parameters, and knowledge interfaces positions DeepThink within the research landscape and identifies the design questions arising from their integration.

Vision-language models (VLMs) can answer many geometric and physical reasoning questions correctly, yet still behave inconsistently in interaction, for example, abandoning a correct answer when confronted with a confident but incorrect user assertion. Standard single-turn accuracy cannot distinguish such capitulations from beneficial corrections because it scores each response in isolation. We introduce GRIP, a controlled benchmark with 34 procedurally generated reasoning domains, 100,000 images, and 600,000 questions spanning geometric, spatial, visual, and physical reasoning. Each scene is generated by deterministic code, and every answer is derived directly from the underlying scene state, providing exact ground truth without human annotation. On GRIP, we propose Dual-Loop Evaluation. A closed loop probes each image with a five-step compositional question ladder to localize where reasoning fails. An open loop then re-elicits the model's answer under controlled challenges, including peer-model responses and unsupported user assertions. Because ground truth is known exactly, answer changes can be signed as corrections (wrong-to-right) or capitulations (right-to-wrong), separating two failure modes: rule absence and rule fragility. The dataset is available at https://huggingface.co/datasets/bilaljawaid980/GRIP-Benchmark, with further details at https://bilaljawaid980.github.io/grip-benchmark/.

Let $\mu_q$ be the distribution of the normalized Kloosterman sums $\mathrm{Kl}(a;q)$, $a\in\mathbb{F}_q^\times$, and let $\mu_{\mathrm{ST}}$ be the Sato–Tate law. We prove $W_s(\mu_q,\mu_{\mathrm{ST}})\ll q^{-1/3}$ for $1\le s\le2$, $W_s(\mu_q,\mu_{\mathrm{ST}})\ll q^{-1/6-1/(3s)}$ for $2\le s<\infty$, and $W_\infty(\mu_q,\mu_{\mathrm{ST}})\ll q^{-1/6}$, answering a question of Kowalski and Untrau. More generally, we bound $W_s(\mu,\mu_{\mathrm{ST}})$ for all $s$ whenever the Chebyshev moments of $\mu$ satisfy bounds of Katz type. For $s=2$ the bound also follows from a smoothing inequality of Borda and Cuenin; for $s>2$ and $s=\infty$ our bounds are stronger. The proof uses the quantile coupling, which arises from Stieltjes sums over two different intervals whose cells are paired so as to balance two measures, and which is optimal for all $s$ simultaneously; an explicit $W_2$ smoothing inequality relative to $\mu_{\mathrm{ST}}$; and a sharp comparison between $W_s$ and the Kolmogorov distance for laws with square-root edges. The last result shows that $q^{-1/6}$ is the best rate for $W_\infty$ obtainable from Niederreiter's discrepancy bound. Numerically, for primes $10^3<p<10^6$, all these distances appear to decay like $p^{-1/2}$.

Swap Monte Carlo dramatically accelerates the equilibration of deeply supercooled liquids while preserving the equilibrium Boltzmann distribution. Recent work has argued that swap and other accelerated algorithms act by reparameterizing time: different dynamics follow the same slow relaxation process, but with different clocks. Here we test this idea in a continuously polydisperse glass former. We confirm that mean collective overlaps satisfy an approximate time reparameterization, in agreement with earlier results. However, the associated clock depends systematically on the collective probe length and on the relaxation stage. Thus the observed collective collapse does not define a unique, observable-independent material time. Over the numerically accessible $\alpha$-relaxation regime, swap changes not only the rate of relaxation, but also its length-scale dependence. The resulting collective reparameterization is therefore only approximate and does not establish an exact dynamical equivalence between swap and reference relaxation.

We propose the task of Interactive 4D Volumetric Liquid Forecasting: predicting the spatiotemporal evolution of a future dense 3D liquid velocity field conditioned on how a solid object moves through it. This task provides an upstream physical forecasting primitive for systems that need to reason about motion-conditioned liquid response, including robotic liquid manipulation, digital twins, and rigid-fluid video generation. Existing works have made efforts in predicting flow around static or prescribed solid geometries and synthesizing visually plausible fluid motion. Yet these approaches provide limited coverage for interactive liquid dynamics, as they are not capable of forecasting how a moving solid induces a future full-domain liquid response. In this forecasting scenario, the hydrodynamic effect is condensed near an evolving liquid-solid interface, but the prediction target is a global spatiotemporal rollout. To address this challenge, we propose a two-level recurrent forecasting framework, TIDE/TIDES: TIDE isolates localized interface-driven forecasting, while TIDES adds transport-aware coupling and explicit stabilization to reduce accumulated rollout error. Across adapted neighboring baselines, controlled component ablations, motion/size OOD tests, and long-horizon evaluation, TIDES achieves the best forecasting performance and the ablations identify localized interaction, transport coupling, and stabilization as the key contributors. Together, our task formulation, benchmark, and TIDE/TIDES study establish a controlled foundation for moving-solid-conditioned dense liquid forecasting. The code is publicly available at https://github.com/jongwonryu/TIDE-TIDES.

Interactive world models must capture how multiple participants jointly change a shared environment while observing different parts of it. Competitive games make this challenge concrete: players cooperate and compete, while fog of war, team vision and camera placement determine what each can see. Modeling these interactions requires data that connect individual observations and actions to shared state and events. We introduce the Esports World Model Dataset (EWM), a collection of human gameplay spanning StarCraft II, Dota 2 and Counter-Strike 2, with 3,053 games and maps and approximately 4,880 player-view hours, including estimated hours for the ongoing CS2 extension. It includes 4,628 completed StarCraft II views, all ten perspectives for 303 Dota 2 matches, and up to ten views per round in CS2. Across these views, replay-clock alignment connects rendered video with game-native actions, structured state and events from the same match. The three games expose complementary observation conditions: private fog of war, shared team vision and first-person viewpoints. In StarCraft II, player-specific entity records further distinguish live observations from remembered snapshots and link co-observed entities across opposing views. Together with temporal alignment, these records allow researchers to distinguish legitimate information differences from disagreement about shared state. By connecting human decisions, shared dynamics and player-specific information, EWM supports research toward jointly controlled multiplayer world models and interactive environments populated by teammates and opponents learned from human play.

Large language models (LLMs) have demonstrated remarkable capabilities in solving complex tasks in coding, mathematical logic, and iterative reasoning, especially by coordinating effectively within multi-agent systems (MAS). However, language agents traditionally communicate via text exchange, which introduces a bottleneck in information transfer and adds prefill latency for downstream agents. Recent work proposes that directly sharing the key-value (KV) cache between language models can enable efficient latent collaboration, entirely bypassing standard text-based communication. To enable latent communication between heterogeneous models of different sizes, prior work introduces trained fusion adapters for each model pair, which limits practical deployment. We introduce CacheFuse, a training-free method which demonstrates that augmenting a downstream model with an upstream model's KV cache using linear interpolation consistently enables superior performance than the standalone downstream model can achieve. We systematically analyze CacheFuse across four reasoning and knowledge benchmarks and demonstrate up to 8.6 percentage points higher accuracy compared to single-model baselines and up to 12.2x TTFT speedup compared to text-based communication in multi-model inference.

Cloud providers offer on-demand access to Field-Programmable Gate Arrays (FPGAs), and multi-tenant FPGAs let multiple users share one device to improve utilization. This sharing introduces security risks: an adversary co-located with a victim on the same FPGA can mount power side-channel attacks against the victim's Machine Learning (ML) accelerator. Prior on-chip attacks on such accelerators based on Time-to-Digital Converter (TDC) sensors have recovered only small, MNIST-scale grayscale inputs, including with ML-assisted reconstruction. This work is the first on-chip, TDC-based attack to produce full-scale color reconstructions of ImageNet-scale inputs, by combining ML-based upscaling and recoloring with TDC measurements of a remote victim's ML accelerator. We frame this as a proof-of-concept leakage study: the victim implements the first convolution layer of a CNN processing 224x224 ImageNet inputs on a shared FPGA, which we isolate to characterize the input-dependent leakage, and we quantify how much of each reconstruction reflects genuinely recovered information versus ML-generated post-processing. As a preliminary countermeasure, we study a pixel-substitution defense that replaces each pixel value, within a fixed intensity budget, with one of different Hamming Weight, producing a different TDC signature. We find that this transformation costs the victim's own model more accuracy than it costs the attacker and, because it applies the same fixed, publicly reconstructable mapping to every image, is largely defeated by an attacker retrained on the transformed data; robust defenses will require randomized or input-dependent transformations.

A first-order theory $T$ has the finite model property if every first-order sentence that is implied by $T$ has a finite model. We give a characterization of when a complete $\omega$-categorical theory $T$ has the finite model property. The characterization makes it possible to transfer stability theoretic results in [29] to certain classes of finite structures. This approach also gives information concerning properties that a sequence of finite “base structures” of ever larger size must have in order for its random expansions to a larger vocabulary to have a chance to satisfy a convergence law for first-order logic.

We construct a counterexample to integral Aleksandrov-Bakelman-Pucci estimates for nonlocal inequalities of the form \begin{equation*} L_au(x):=\int_{\mathbb R^n}\delta_yu(x) \frac{a(x,y)}{|y|^{n+\sigma}}\,dy \ge-f(x)\quad\text{in }B_1, \end{equation*} with $u\le0$ in $\mathbb R^n\setminus B_1$, where $\delta_yu(x)=u(x+y)+u(x-y)-2u(x)$, and the measurable coefficient satisfies \begin{equation}\tag{C} \lambda\le a(x,y)\le\Lambda,\qquad a(x,y)=a(x,-y),\qquad x\in\mathbb R^n,\quad y\ne0. \end{equation} More precisely, for every $n\ge1$, $0<\sigma<2$, and $0<\lambda<\Lambda$, there exist a fixed coefficient $a(x,y)$ satisfying (C), a constant $d>0$, and nonnegative functions $u_j,f_j\in C_c^\infty(B_1)$ such that $L_au_j\ge-f_j$ in $B_1$, $0\le f_j\le1$, and $|\operatorname{supp}f_j|\to0$ as $j\to\infty$, while $u_j\ge d$ on $B_{1/4}$ for every $j$. Consequently, this fixed operator admits no $L^p$ maximum estimate for any $1\le p<\infty$.

A fruitful approach towards understanding generalization in machine learning involves characterizing how models separate generalizable structure from idiosyncratic or even corrupted features of training data. Fisher information has emerged as a diagnostic tool for this separation, leveraging loss geometry to distinguish generalization from memorization. Here we investigate this connection in vision models trained on CIFAR-10 with controlled noise that can be learned only through memorization. We show that projecting model weights onto the top Fisher eigenvectors decouples general task performance from noisy sample memorization, and we investigate how this decoupling presents differentially across layers. We show that the same phenomena are present in small transformers trained to solve a synthetic in-context learning task, and are reproduced in closed form in a simple, analytically solvable linear model. Our results reveal that Fisher information captures aspects of the geometric structure underlying how neural networks allocate capacity between generalization and memorization.

Recent advances in 3D reconstruction enable the extraction of temporally consistent 3D motion from RGB observations; however, current Vision-Language-Action (VLA) models still predominantly rely on 2D image-text priors. Therefore, they do not accurately represent the fundamental dynamics of the 3D world associated with physical actions. This over-reliance on static 2D visual cues limits the capacity of VLA models in achieving robust robot manipulation. To address these limitations, we introduce FlowVLA, a reconstruction-guided, pluggable 3D flow learning framework aimed at empowering VLA models with an inherent understanding of action-relevant dynamics. Instead of requiring explicit 3D sensor inputs or altering the inference-time architecture, our method is designed to utilize temporally aligned 3D flow only during training. This allows the model to learn the dynamic evolution of the physical scene while preserving the computational efficiency of conventional 2D VLA backbones. To bridge scene-level dynamics with robotic manipulation, we propose a modular 3D flow learning framework with dual-level flow supervision. Specifically, point-level flow guides the model to forecast the future displacements of task-critical 3D points, while TCP-level flow aggregates these motions to align with the robot's TCP trajectories. This dual-level flow supervision encourages the model to focus on semantically meaningful dynamic regions, effectively coupling scene-level motion with precise control signals. Extensive evaluations across simulation and real-world experiments demonstrate that FlowVLA consistently improves manipulation performance, zero-shot robustness, and optimizes data efficiency. FlowVLA adds only lightweight auxiliary heads during training and discards them at inference. Therefore, FlowVLA introduces no additional 3D sensor inputs, architectural modifications, or inference-time computation.