Computer Vision
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Sunday 8 November
Streaming 3D foundation models (3DFMs) estimate camera poses and dense geometry online, but they fall short of SLAM: their predictions are fixed once emitted, so loop closures cannot correct accumulated drift. With such models already providing strong geometric capability, a central question for SLAM becomes how to schedule their inference and how to keep its results as revisable state. We present Map2SLAM, which keeps LingBot-Map frozen, trains only a lightweight retrieval head, and builds SLAM on the model's inference interface, turning its predictions into observations of an external persistent SLAM state. The front-end tracks every incoming frame with the model's causal streaming inference and decides which frames are written to the persistent state as keyframes. The back-end asynchronously decides when and which past frames to re-observe, re-running the model's joint inference over them with fresh context; verified loop observations correct the trajectory online, and segment observations refine dense geometry after the sequence. All invocations share a single frame encoder pass per image, whose cached features also drive loop retrieval, so frequent re-invocation incurs no additional encoding cost. Map2SLAM achieves state-of-the-art performance on KITTI, VBR, and Oxford Spires, reducing mean ATE by over 40% compared with the best uncalibrated baselines, while maintaining real-time tracking at over 28 FPS at a resolution of $518\times154$ on an NVIDIA RTX 5090D GPU. Please visit our project page at https://robbyant-research.github.io/Map2SLAM/.
Rings and Algebras
A unified approach to Manin triples of Lie algebras and the corresponding bialgebra structures
We introduce self-module Lie bialgebras as a common framework for bialgebra structures associated with different types of Manin triples of Lie algebras. A self-module Lie algebra is a Lie algebra with a specified representation on its underlying vector space; equivalently, it is an NS-Lie algebra expressed in different notation. We define a generic Manin triple as a finite-dimensional Lie algebra equipped with a nondegenerate bilinear form and two complementary isotropic Lie subalgebras, and establish the equivalence among standard generic Manin triples, suitable matched pairs of Lie algebras, families of matched pairs of self-module Lie algebras and self-module Lie bialgebras. The framework recovers Lie, pre-Lie and anti-pre-Lie bialgebras as special cases and admits examples outside these three families. We also introduce the self-module Lie Yang-Baxter equation and show that its solutions yield self-module Lie bialgebras and admit the $\mathcal{O}$-operator characterization. Finally, we obtain solutions in semi-direct product self-module Lie algebras, including constructions arising from self-bimodule pre-Lie algebras.
Read the full abstract
Systems and Control
Numerical Stability Issues and a Mitigation Method in Switching Events Calculation Considering State Variable Discontinuities for EMT Simulation
Some existing methods for switching events calculation considering state variable discontinuities for EMT simulation can encounter numerical stability issues, either over-stabilizing or destabilizing the studied power systems. This paper documents these issues in detail with a concrete example. The underlying mechanism of the issues are revealed via systematic analysis. A mitigation method for the issues is proposed. The convergence behavior of the method is analyzed. Numerical examples demonstrate the effectiveness of the method. It is observed that the method also helps to improve the accuracy of the simulation results.
Read the full abstract
Robotics
DexForesight: Grounding Dexterous Policies with Implicit Spatial Foresight
Dexterous manipulation requires rich spatial foresight over both the surrounding scene and the robot hand. Existing world-action models often acquire such predictive knowledge through explicit future prediction. We introduce DexForesight, a framework that learns spatial foresight implicitly by injecting dynamics-aware world representations into the latent space of a dexterous policy. During training, DexForesight learns an action-conditioned spatial world model from offline scene depth, kinematic hand geometry, and demonstration trajectories. The world model predicts 3D hand-point motion and visual feature changes, capturing both articulated hand dynamics and scene changes under demonstrated actions, including the motion of fingers occluded from the cameras. We then freeze the world model and align policy queries with its predictive world queries, transferring the learned spatial dynamics into policy latents and using them to condition action generation. At deployment, only the policy is retained, operating on RGB observations, language, and robot state without depth input or explicit future prediction. On 11 DexJoCo simulation tasks, DexForesight improves the overall success rate of the base policy from 49.4% to 60.5%. On three real-robot dexterous manipulation tasks, DexForesight achieves 58.3% success, compared with 50.0% for the strongest baseline. Project page https://dexforesight.github.io/
Read the full abstract
Machine Learning
The Race Is Decided Early! A Kernel Law Predicts Lock-In in Majority-Vote Test-Time RL
Test-time reinforcement learning with majority-vote rewards trains a language model on unlabeled questions by rewarding the samples that agree with its most frequent answer. During training, each question locks into one answer, on hard questions often a wrong one, but which answer it locks into is known only after training. In this paper, we show that this outcome is decided early, in a short race driven by sampling noise between the answers that are close at the start, and we derive the probability of each outcome. With a separate logit for each answer, the runner-up wins with a normal tail probability of the initial gap, scaled by an explicit noise constant. In a network, however, the answer logits depend on shared parameters, and through the kernel of these parameters the updates of the trailing answers also change the gap of the two leaders. This spillover shifts the boundary between the starts won by each leader, and a closed form computes the shifted probability from a kernel probe and the start logits. In LoRA experiments with three models from two families, predictions made before training with the kernel frozen at its start value forecast which answer each question locks into better than a logistic model refitted on the outcomes, although they use no fitted constants. The closed form matches these predictions on the runs in which a check before training finds that the two leaders stay above the other answers, and it explains why a law that ignores the spillover needs a strongly inflated step.
Read the full abstract
Robotics
Tutorial: Vision-Based Formation Control with Differential-Drive Robots
This tutorial presents a step-by-step introduction to multi-robot formation control using differential- drive robots, an overhead camera, and Python. ArUco markers are used to estimate robot positions and orientations, while desired pairwise displacement vectors define a circular formation. By comparing measured and desired relative positions, the controller computes movement vectors that are converted into differential-drive wheel commands, with basic proximity-sensor-based obstacle avoidance. The tutorial guides readers through hardware and software setup, robot pose estimation, formation geometry, feedback control, system integration, and formation-error evaluation. Motivated by the need to bridge the gap between theoretical understanding and practical implementation in robotics education, we emphasize how mathematical concepts are translated into coordinated behaviors on physical robots. Although the implementation uses a specific robotic platform, the underlying perception and formation-control principles can be adapted to other differential-drive robots and simulation environments. The accompanying open-source GitHub repository provides the Python implementation, setup instructions, and demonstrations to support hands-on learning and further exploration of multi-robot systems.
Read the full abstract
Systems and Control
Learning the NMPC with Stability Guarantees
Although nonlinear model predictive control (NMPC) is widely used, its online optimization is computationally expensive. Neural network approximations of NMPC can effectively reduce the online computational burden but generally lack stability guarantees. This paper proposes an approximate NMPC framework with guarantees of constraint satisfaction and closed-loop stability. The approximation error is addressed through constraint tightening, and the resulting tightened NMPC is used to generate the training dataset of the approximate NMPC. The maximum allowable approximation error is prescribed a priori for the design of tightened NMPC and satisfied in a probabilistic sense through an iterative trial-and-error procedure alternating between network training and a posteriori verification. Finally, the effectiveness of the proposed framework is demonstrated through a numerical example of a constrained nonlinear system.
Read the full abstract
Atmospheric and Oceanic Physics
Analysis increments for model bias diagnosis and correction: Theoretical foundations
A common method for diagnosing and correcting model bias in weather and climate prediction is based on analysis increments: the difference between the analysis and the forecast at each data assimilation cycle. Their time average is used to estimate and correct the bias, and machine learning is increasingly used to map the forecast state to the increments, giving a state-dependent correction. Despite its widespread use, the method lacks theoretical foundations. We first analyse a Kalman filter whose forecast model has a constant bias. We show what the averaged increment measures, and prove that iterating the correction removes all of the bias seen through the observations; for linear dynamics, no method can identify the rest. We then extend the analysis to state-dependent biases. For linear dynamics, each increment records the model error over the filter's memory, as seen through the observations. We use this to determine what the increments measure when averaged over a long period, averaged over periods with similar slowly varying conditions, and used without averaging, as in machine-learned corrections. Numerical experiments confirm the theory for a linear model and suggest that its main conclusions carry over to the nonlinear Lorenz '96 model, including for a correction learned by a neural network.
Read the full abstract
Dynamical Systems
Proximal Myhill-type theorems for $m$-expansive amenable semigroup actions
The Garden of Eden theorem relates surjectivity to a weak injectivity condition for cellular automata on amenable groups. Li extended the Myhill implication to expansive actions of countable amenable groups with weak specification. We consider two simultaneous relaxations: the group is replaced by a cancellative amenable semigroup, and expansiveness by m-expansiveness. The loss of ordinary expansiveness removes homoclinic rigidity, so we introduce a proximal analogue of pre-injectivity. Our main result shows that, for a surjective action of an infinite countable cancellative left-amenable semigroup on a compact metric space, m-expansiveness plus weak specification forces every continuous equivariant map injective on each proximal set to be surjective. The proof uses entropy detection at a fixed scale, a strict entropy gap from weak specification, and finite multiplicity to produce proximal pairs from uniformly close points. We also give a one-dimensional variant under shadowing and examples clarifying the assumptions.
Read the full abstract
Medical Physics
Design and Performance Evaluation of a Quad-panel Dual-isotope PET Imaging System
In simultaneous dual-isotope PET, prompt-gamma-based separation is challenged by the low statistics of triple coincidences and the limited flexibility of coincidence logic for multi-energy event extraction. We designed and evaluated a compact quad-panel dual-isotope PET system without dedicated prompt gamma detectors. The system comprised four orthogonally arranged detector panels, each providing an effective area of 107.0 * 145.5 mm. GPU-accelerated software coincidence processing extracted double- and prompt-gamma-tagged triple coincidences from raw single-photon data. Triple-coincidence images were reconstructed with correlated-random-triple correction using the consistent delayed expectation-maximization method, followed by image-domain isotope separation using a spatially varying scaling factor. System performance was evaluated using phantom and in vivo studies, with emphasis on 68Ga/18F imaging because of the low prompt gamma yield. The system achieved energy resolutions of 15.7%, 12.9% and 10.2% at 511, 1077, and 1275 keV, and a timing resolution of 1.35 ns. In the 68Ga Derenzo, 1.6-mm rods were resolved in double-coincidence images, with 1.2-mm rods partially distinguishable, while 1.6-mm rods remained discernible in triple-coincidence images. The mean prompt gamma detection efficiencies along the three axes ranged from 8.19% to 9.31%. The system sensitivity was 1.12% for doubles with an energy window of 350-650 keV, and 0.068% for triples after random correction with a dual-energy window. The scatter fraction was 27.81%, and the NECR reached 83 kcps at 9.38 MBq for doubles and 106 cps at 21.42 MBq for triples. Phantom and in vivo studies demonstrated successful separation of 18F/68Ga and 18F/124I tracer distributions. By maintaining high-resolution and quantitative imaging performance, the system provides a preclinical platform for multitracer PET studies.
Read the full abstract
Instrumentation and Methods for Astrophysics
Foreshortening of Polarizer Angles in Wide Field Polarizing Imagers
The Polarimeter to Unify Corona and Heliosphere (PUNCH) is a very wide field polarizing imager covering a 90 degree diameter region of the sky, stitched together from four individual fields of view (FOVs), three of which are supported by Wide Field Imagers (WFIs) with 40 degree diameter individual FOVs. The science objective requires precise measurement of linear polarization. The WFI FOV is sufficiently wide that polarizer angles are foreshortened by projection effects relative to the line of sight of each pixel in the instrument, deviating from their nominal angular positions of -60, 0, and +60 degree relative to the instrument frame of reference. These angles are foreshortened by the very wide WFI field of view, changing their effective angle across the image plane - an effect that is analogous to similar challenges of polarimetric right/left channel separation in large-format 3D cinema, such as the IMAX format. Using a simple geometrical approach we estimated the perceived polarizer angle in the WFI FOV will vary by +/-2.5 degree for +/-60 degree and +/-3.5 degree for 0 degree polarizer angles. We present the effect of this deviation on the polarized brightness to be recorded by WFI, along with how optical distortion will shift the location of foreshortened angles. We also discuss the implications of this effect on upcoming technologies based on polarimetry across in wide fields.
Read the full abstract
Numerical Analysis
An Arbitrary-Order Distributional Finite Element Method for the Curl–Dev Grad Stokes System
We develop an arbitrary-order distributional finite element method for the three-dimensional curl–dev grad Stokes system governed by the Hessian complex. On tetrahedral meshes, the method couples degree-$(k+1)$ nonconforming symmetric-tensor virtual elements, degree-$k$ tangential-normal continuous traceless tensor elements, and Raviart–Thomas elements of index $k$, without additional virtual-element stabilization. Building on zero-extension curl–div theory, we derive a well-posed distributional saddle-point formulation with an $H(\operatorname{div})$ vector multiplier. A divergence-augmented discrete norm equivalence controls the full Raviart–Thomas space modulo translations and underpins mesh-uniform stability for each fixed $k\ge0$. We establish optimal $O(h^{k+1})$ energy error estimates and an optimal $O(h^{k+2})$ $L^2$ error estimate for the full symmetric-tensor approximation, under suitable regularity and load approximation assumptions. As an application, the method supplies the Stokes block of a decoupled triharmonic scheme with optimal Hessian error estimates. Numerical experiments support the predicted convergence rates for computable quantities.
Read the full abstract
Quantum Physics
Adaptive Quantum Metrology of Gravitational Waves
We develop a feedback-controlled quantum metrology scheme to probe weak gravitational waves (GWs) through their direct energy exchange with light. Adaptive SU(2) direction control, based on current estimates of the GW frequency and phase, allows the frequency shifts acquired in successive interaction intervals to add rather than cancel. The resulting persistent frequency difference is converted into an optical phase difference for interferometric readout. We further extend the scheme to elliptically polarized GWs using two transverse sensing channels separated by $45^\circ$ to probe the $+$ and $\times$ components independently. After the GW interaction, the two signals are mapped onto orthogonal polarization modes, enabling phase accumulation along a common path. We derive the corresponding quantum Fisher information matrix and find a $T_{\rm int}^2T_{\rm ph}^2$ scaling in the ideal coherent regime. For a fixed-photon-number probe, the sensitivity shows Heisenberg-like $1/n$ scaling in the ideal case.
Read the full abstract
Networking
Federated learning for remote IoT over LPWAN: Preliminary insights from structural health monitoring
Realizing federated learning (FL) in remote regions lacking reliable cellular infrastructure is challenging, as it requires dependable communication among participating devices. Low-power wide-area networks (LPWANs) offer an attractive communication solution for such scenarios. However, their limited data rates necessitate careful investigation of the communication-learning trade-offs and mechanisms to reduce communication overhead. This paper presents an investigation of the Long Range (LoRa) LPWAN technology for FL-based structural health monitoring (SHM), a representative remote IoT application. We consider a SCAFFOLD-based FL framework for multi-class anomaly detection and analyze the latency-accuracy trade-off. We also examine strategies for allocating the LoRa spreading factor (SF) during model-update exchange. Numerical results show that, with appropriate SF allocation, LoRa-supported FL-SHM can achieve competitive learning accuracy. In particular, jointly accounting for link distances and clients' dataset sizes when selecting the SF provides a favorable balance between convergence performance and communication latency. These preliminary findings offer useful insights into the design of FL systems over LPWANs for remote IoT deployments.
Read the full abstract
Optimization and Control
Data Center Power Flexibility Characterization: A DeepSets Reinforcement Learning Approach
Data centers can provide power flexibility through the scheduling of computing workloads, cooling systems, and batteries. Characterizing this flexibility online is useful but challenging because of the complex coupling among these resources. To address this challenge, this paper proposes a role-conditioned cross-head DeepSets reinforcement learning approach to characterize the power flexibility interval of a data center at each time step. First, we formulate the flexibility characterization problem as a Markov decision process that jointly models the three resources under a carbon budget, with actions specifying the upper and lower endpoints of the power flexibility interval. We then design DeepSets actor and critic that represent delay-tolerant workloads as an unordered set, enabling the proposed approach to accommodate varying numbers of workloads with a fixed number of trainable parameters. Finally, because maximizing the width of the flexibility interval requires moving its endpoints in opposite directions, we develop a role-conditioned cross-head decoder that distinguishes their respective roles while keeping them coordinated. Case studies on the Google Cluster Data show that the proposed approach provides more than twice the daily flexibility of the standard soft actor-critic method with nearly zero carbon overshoot while using 87.7% fewer trainable parameters.
Read the full abstract
Representation Theory
Finite groups with consistent subgroup lattices
Motivated by the problem of Gragg and Kung on characterizing finite groups with consistent subgroup lattices, we investigate how the consistency of the subgroup lattice $L(G)$ constrains the structure of the underlying finite group $G$. To this end, we first establish several equivalent characterizations of consistency for lattices of finite length in terms of operation-preserving subposets and forbidden sublattices. More interestingly, for complete atomistic lattices of finite length, we introduce the consistency exchange property (CEP) for the canonical closure operator and show that consistency is equivalent to CEP, and that CEP is strictly weaker than the Steinitz-Mac Lane exchange property. We then apply this framework to semidirect products $G=V\rtimes_{\alpha} C_m$, where $V$ is an elementary abelian $p$-group, $m$ is square-free and $p\nmid m$. Writing $d=o(\alpha)$ and letting $C_\alpha=\langle\alpha\rangle$, we prove that $L(G)$ is consistent if and only if either $d=1$, or $d$ is prime and $V$ is isotypic as an $\mathbb F_pC_\alpha$-module. For prime $d$, this latter condition is equivalent both to the irreducibility of the minimal polynomial of $\alpha$ over $\mathbb F_p$ and to $\mathbb F_p[\alpha]$ being a field. More generally, we verify that if $G=V\rtimes_\theta H$ where $H$ acts faithfully on $V$ and $p\nmid |H|$, then $L(G)$ is consistent if and only if either $H$ is trivial, or $H$ is cyclic of prime order and $V$ is isotypic as an $\mathbb F_pH$-module. We thus obtain a complete solution to the problem of Gragg and Kung for this class of coprime semidirect products.
Read the full abstract
Representation Theory
Inhomogeneous deformations of $\mathfrak{osp}(N|2n)$: triviality for any number of odd parameters
We treat a symmetrized mixed-oscillator deformation family of the orthosymplectic Lie superalgebras $\mathfrak{osp}(N|2n)$, $N,n\geq1$, with even mixed coefficients $\beta_{u,i}$ and one odd square-zero parameter. We derive the deformed bracket from an oscillator realization, show that its coefficient is determined by the realization, and exhibit an explicit odd primitive, normalized by its vanishing on $\mathfrak{sp}(2n)$. Every odd primitive differs from it by a unique inner derivation. For $N=1$ this is the case $\mathfrak{osp}(1|2n)$ treated earlier. With multiple odd parameters the deformed bracket is the sum of the single-parameter deformations and is again trivial. The trivialization, however, is not additive: it is the product of the single-parameter trivializations followed by a nonsymplectic change of the bosonic generators. The main results are formalized in Lean~4, with the algebras constructed by generators and relations; what is not formalized is stated in the appendix.
Read the full abstract
Machine Learning
Full Fine-Tuning is not the Problem: Why Adam Fails and SGD Succeeds in Few-Shot CLIP Adaptation
Few-shot CLIP adaptation is often treated as a parameterization problem. Full fine-tuning is assumed to overfit, so parameter-efficient fine-tuning (PEFT) methods are preferred. In this paper, we show that this conclusion conflates parameterization with optimization. Across 11 datasets, 3 CLIP backbones, and 5 shot levels, dense full fine-tuning with plain SGD outperforms representative PEFT methods under the reported evaluation protocol, whereas Adam/AdamW collapses. We trace this collapse to unreliable support-set preconditioning. The few-shot second moment systematically underestimates the corresponding full-data statistic, already at initialization. Adam's inverse-square-root normalization converts this error into an oversized adaptive update. Matching Adam's first-step update budget to SGD resolves most of the collapse, whereas numerator-side alignment does not explain the failure. Shampoo and SOAP also fail, so richer preconditioning alone does not resolve the problem. Theory formalizes how few-shot estimation error is amplified into unsafe updates. SGD succeeds because it avoids unreliable preconditioners and stays close to pretrained CLIP. Interpolation probes show low-loss connections along tested directions, and curvature probes show favorable directional curvature. These results identify optimizer-induced preconditioning, not dense full fine-tuning itself, as the main source of failure in few-shot CLIP adaptation.
Read the full abstract
Data Structures and Algorithms
Fast Approximate $\ell_p$ Chamfer Distance via Lopsided Embeddings and Structured JL
For two $d$-dimensional point sets $A, B$ of size $n$, the Chamfer distance from $A$ to $B$ is defined as $\text{CH}(A,B)=\sum_{a \in A} \min_{b \in B} \mathsf{dist}(a, b)$, where $\mathsf{dist}$ is some underlying distance measure. We present efficient algorithms for approximating the Chamfer distance under the $\ell_p$ norm for $1 < p \le 2$. For the general case $1 < p < 2$, we utilize lopsided $\ell_p$ to $\ell_1$ embeddings with weak guarantees. We show that they are sufficient for preserving the Chamfer distance. In the high-dimensional regime, we use Fast Matrix Multiplication techniques to speed up the lopsided embeddings. These lead to $O(nd\cdot \mathrm{poly}(\log \log n, \log(1/\varepsilon))/\varepsilon^2)$ total time. For the specific Euclidean case ($p=2$), we leverage the Fast Johnson-Lindenstrauss Transform based on Toeplitz matrices and re-analyze the previous $\ell_1$-specific Chamfer algorithm in $\ell_2$. These achieve a runtime matching the recent state-of-the-art $\ell_1$ result of $O(nd(\log \log n + \log (1/\varepsilon))/\varepsilon^2)$, improving upon the previous $O(nd \log n/\varepsilon^2)$ bound for $\ell_2$. We also give additional results for the angular distance and upper and lower bounds in the streaming setting.
Read the full abstract
Geophysics
Combining rock typing and NMR logging to estimate permeability in complex carbonate formations
Estimating intrinsic permeability from nuclear magnetic resonance (NMR) logs is difficult in heterogeneous carbonates because a single calibration may not represent distinct pore systems. We evaluated whether flow-zone-indicator (FZI)-based rock typing improves empirical calibration of the modified Schlumberger-Doll Research (SDR) model in the Arbuckle Group, Kansas. The analysis used 470 depth-matched core-permeability and NMR measurements from the Wellington and Cutter wells. FZI calculated from measured permeability and NMR-derived porosity was grouped using K-means clustering. Six clusters were identified, but one singleton cluster was excluded, leaving five rock types for SDR calibration. Relative to the global model (without clustering), rock-type-specific calibration reduced testing root mean square log-transformed error (RMSLE) from 1.215 to 0.240-0.427 and increased coefficient of determination (CD) from -0.106 to 0.578-0.924. The fitted porosity exponents ranged from 0.053 to 0.521, indicating that porosity provided limited additional information in some rock types and that the SDR parameters should be treated as empirical calibration quantities. Because measured permeability was used to calculate the FZI before the within-cluster splits, these results represent conditional performance for samples with known FZI membership, not blind permeability prediction. Field implementation will require a permeability-independent method for assigning rock types.
Read the full abstract
Machine Learning
ZarrFM: A Canonical Tensor Storage Architecture for EEG Foundation-Model Pretraining
Statistical models for brain-computer interfaces (BCIs) increasingly require large, heterogeneous EEG corpora, yet existing datasets vary in sampling rates, channel layouts, window lengths, and storage formats, creating substantial preprocessing and I/O overhead. We propose ZarrFM, an execution-oriented representation that complements BIDS by storing a corpus as a single Zarr v3 hierarchy with indexed metadata, canonical channel spaces, explicit masks, hash-stable subject splits, verified event tables, and block-aligned sharding. On real PhysioNet recordings and synthetic corpora, we compare ZarrFM with EDF, FIF, NPY, HDF5, and Parquet. Across 200 recordings, its two-level index reduces metadata queries from 601 requests to 14. Preserving native signal codes reduces storage from 42.4 MB/hour in EDF to 22.7 MB/hour, while block-aligned sampling reaches 3,131 windows/s, up to 11 times faster than Parquet. Reads remain bit-identical across independent Zarr implementations, and concurrent failure-injected appends recover without data loss. These results show that ZarrFM can provide a compact, interoperable, and execution-efficient storage layer for large-scale EEG pretraining.
Read the full abstract
Robotics
MOSU2: Long-range Navigation System with Multi-modal Scene Understanding
Robots that travel long distances through public outdoor spaces must do more than avoid obstacles: they must remain on terrain suited to their mobility and decide when to proceed, yield, or detour around pedestrians, cyclists, vehicles, traffic signals, and temporary restrictions. These decisions are difficult because their correctness often depends on how a scene evolves, such as whether an approaching car is slowing or a cyclist will cross the robot's path. Vision-language models (VLMs) can interpret such cues, but they reason over images that lack metric geometry and motion. Supplying this evidence at every step is too slow for real-time navigation, and hand-crafted triggers cannot anticipate the open-ended variety of outdoor situations. We present MOSU2, a long-range navigation system with two-stage trajectory selection, in which first-stage contextual reasoning determines when additional evidence is needed and what to retrieve. Only when an interaction remains unresolved does the first stage trigger the second stage to retrieve geometric and dynamic information and align it with scene context to make navigation decisions. Building on MOSU, MOSU2 follows routes from GPS-based route planning and filters candidate trajectories through multi-modal traversability analysis before selecting a trajectory to move the robot. In real-world experiments spanning stop gestures, construction sites, crosswalks, and roads shared with vehicles, cyclists, and pedestrians, MOSU2 raises mean navigation success from 61.8% to 94.2% and condition-detection accuracy from 72.2% to 91.4% over its single-image predecessor, at about 2.2 times its mean decision latency. Project website: https://jinggm.github.io/mosu2/
Read the full abstract
Computer Vision
Not All Layers Are Equal in Image-to-Video Transfer
Parameter-efficient image-to-video transfer adapts frozen image foundation models to video via lightweight trainable adapters. Existing methods insert the same adapter at every transformer layer, implicitly assuming all layers require equal temporal modeling. A layer-wise sensitivity analysis across four video-native model families and two pretraining paradigms reveals that this assumption is wrong: temporal sensitivity increases with depth, and the pretraining paradigm determines its growth profile. Vision-only models exhibit a gradual rise, while vision-language models show early suppression followed by late recovery. This finding motivates STRAP, a training-free adapter placement rule that derives a spatial-to-temporal transition boundary from a paradigm-matched video model. Spatial adapters are assigned before this boundary and temporal adapters after it, without modifying the adapter architecture itself. On four fine-grained action recognition benchmarks, STRAP improves accuracy by up to 11.8% over uniform placement, reduces trainable parameters by 29 to 48%, and lowers inference time by up to 10%. On standard benchmarks, it matches accuracy with fewer parameters. Code is available at https://github.com/th-nesh/STRAP.
Read the full abstract
Analysis of PDEs
Well-posedness and attractors for the sub-quintic weakly damped wave equation with low-regularity forcing on $\mathbb{R}^3$
We study the weakly damped semilinear wave equation on $\mathbb{R}^3$ with a sub-quintic nonlinearity $f(u)$ and a forcing term of low regularity, $g\in H^{-1}(\mathbb{R}^3)$. By absorbing the forcing term into a stationary solution $h$ and then expanding the solution around $h$, we reduce the problem to a wave equation with potentials. Relying on Strichartz estimates for this equation, obtained through the Besov spaces associated with the Schrödinger operator, we establish the global well-posedness of weak solutions by an iteration argument on the Besov scale. In particular, uniqueness holds unconditionally in the energy class for $3<p<5$. Combining a refined decomposition of the solution with uniform tail estimates for trajectories starting from bounded sets, we then prove that the associated solution semigroup is asymptotically compact and, consequently, possesses a compact global attractor in the energy space $H^1(\mathbb{R}^3)\times L^2(\mathbb{R}R^3)$.
Read the full abstract
Machine Learning
DISCOVER: Online Variance-Guided Data Discovery for Budgeted Multimodal GRPO
Recent advances in reinforcement learning have made verifiable rewards a central post-training paradigm for improving multimodal reasoning. However, methods such as GRPO remain data- and compute-inefficient, spending rollout and annotation budget on prompts that provide little learning signal. We propose DISCOVER, an online data-discovery framework for budgeted multimodal GRPO that enables vision-language models to dynamically identify prompts that are most useful for the current policy. DISCOVER first organizes the multimodal sample pool into semantic clusters and queries a small set of density-diverse representative samples to obtain initial reward observations. During training, it estimates the utility of unannotated samples by transferring sparse reward feedback from queried samples. Since sample utility changes as the policy evolves, DISCOVER dynamically updates these estimates with recency-weighted rollout observations. It then selects only high-utility prompts for annotation using a policy-conditioned score that accounts for reward variance and difficulty. Once the discovery budget is exhausted, DISCOVER continues with utility-based replay. Experiments on ViRL39K and VLAA-Thinking with Qwen3-VL-2B show that DISCOVER substantially improves multimodal mathematical reasoning under a strict 10% unique-sample budget, outperforming budget-matched random selection, hard-example sampling, full-pool GRPO under the same training-step budget, and online selection baselines. Ablations further validate the importance of semantic coverage, online reward transfer, and stable discovery cadence, establishing DISCOVER as a practical framework for data-efficient multimodal RL post-training.
Read the full abstract
Classical Analysis and ODEs
A Quarter-Power Bound for Chowla's Cosine Problem
For every nonempty finite set $S$ of $N$ positive integers we prove $-\min_{x\in\mathbb{T}}\sum_{s\in S}\cos(2\pi sx)\gg (N/\log(2N))^{1/4}$. The proof uses a reversible phase kernel whose one-step and two-step laws are close in total variation. Comparison of rows and a height cutoff produce a partition into pieces of bounded mean. A Fourier rounding estimate then forces many additive triples to cross the partition. Two choices of phase density yield bounds in terms of negative mass and an exact conditional-pair defect, respectively. A logarithmic refinement gives a log-free bound at low triple density. At fixed positive triple density the log-free bound is a corollary of known boundary estimates. We give a separate spectral proof using an integer difference count and a theorem of Yip.
Read the full abstract
Theoretical Economics
Fair Combinatorial Auctions
An efficient and stable auction can leave the seller with no share of the surplus. We study how the gains from a combinatorial auction can be shared as equally as possible between the seller and buyers while respecting coalitional stability. We show that, when buyers are substitutes, the core is described by individual caps on the buyers' payoffs and contains a unique Lorenz-dominant payoff vector. Each buyer receives the seller's payoff capped at her marginal contribution. The seller consequently receives at least an equal share of the total surplus, without compromising efficiency or stability. We then construct a complete-information bargaining game in which the seller proposes a division of the surplus and rejecting buyers receive exit payments. Without discounting, the game fully implements the Lorenz-dominant vector in pure-strategy subgame-perfect equilibrium: an equilibrium exists, and every such equilibrium yields precisely this division.
Read the full abstract
Number Theory
On mod p Hecke algebras of abstract p-adic automorphic forms
We study completed local shallow Hecke algebras acting on residual automorphic forms attached to projective augmented modules for $\GL_2(\Q_p)$. For odd primes $p$, we prove that the maximal ideal has nilpotence exponent $O(p^{r/2})$ on the level-$r$ generalized eigenspace. For $p\ge5$, if the corresponding Hecke eigenspace in the residual smooth dual has finite length as a $\GL_2(\Q_p)$-representation, then the completed local algebra is noetherian of Krull dimension at least two. If, in addition, its cotangent space has dimension at most two, then it is a formal power series ring in two variables. The nilpotence bound follows by conjugating level subgroups. To deduce the dimension bound, we replace the Fourier pairing in the classical nilpotence method with an inequality comparing the dimension of each finite-level generalized eigenspace with that of its Hecke eigenspace. We apply these results to modular curves and definite quaternion algebras.
Read the full abstract
Computational Engineering
TALK: One-Shot Batch Design for Protein Variant Effect Prediction
Protein variant effect prediction is central to understanding protein function and guiding protein engineering, yet experimental labels remain costly. In many protein experiments, the variants that can be assayed are dictated by protocol constraints, while the variants that need accurate prediction may lie in a different or broader region of the landscape. This assay-prediction mismatch makes candidate-local criteria such as predicted function, uncertainty, or exploration value insufficient for measurement design. Given a protocol-defined accessible space and a limited budget, we ask which variants should be assayed so that the resulting data best improves landscape prediction. We propose TALK, which estimates each candidate's contribution through posterior coupling between accessible and target variants and reduces redundant information within the batch. Across retrospective ProteinGym and FLIP protocols, TALK generally improves target-space prediction under limited measurement budgets. These results point to a practical two-stage workflow: first, spend a limited assay budget on variants selected to be informative for the target landscape and train a predictive model from the measured labels; then use this model to score and prioritize high-potential variants. Code is available at https://github.com/joehuang618/TALK.
Read the full abstract
Computer Vision
SurgWorld: Learning Surgical Representations Beyond Observation via Latent Prediction
Surgical procedures are inherently partially observable, requiring surgeons to continuously infer evolving anatomical dynamics and future surgical states beyond the current observation. However, existing self-supervised learning methods primarily learn from observed surgical videos through masked reconstruction or predicting representations within the camera's current field of view, lacking the capability to explicitly model the out-of-sight surgical dynamics. Inspired by latent predictive world models, we introduce SurgWorld, a video-native, self-supervised framework for surgical representation learning, which treats a specified spatial or temporal shift as a pseudo-action and learns to predict the target latent representation from a partial observation. Specifically, we propose Out-of-Sight Dynamics Modeling (OSDM) as the pretraining objective for predicting representations conditioned on relative spatio-temporal shifts. We further regularize the learned latent dynamics from OSDM with a Geometric Visual Similarity (GVS) objective by aligning patch-level relational structures between predicted and target representations, preserving spatial geometry and local visual coherence for dense perception. These two objectives are integrated through an agglomerative training framework that consolidates complementary supervisory signals from multiple teachers to learn a unified latent representation of surgical dynamics and visual geometry. Extensive transfer learning experiments on 17 surgical benchmarks show that SurgWorld consistently outperforms existing vision foundation models across surgical workflow recognition, fine-grained action understanding, and dense visual perception. We further conduct tailored qualitative and quantitative analyses, demonstrating that SurgWorld effectively captures surgical dynamics beyond the directly observed context through latent prediction.