Computer Vision
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Thursday 10 September
Long-video understanding on edge devices must reason over hours of content under tight compute and bandwidth budgets. Subsampling visual tokens loses temporal structure, while text-only video memories lose fine-grained visual attributes. We observe a visual-textual duality: language memories carry long-range temporal structure better than dense frames, while pixels remain decisive for attribute-level perception. Building on this insight, we propose Caption-once, Frames-onDemand (CFD), a budget-aware edge-cloud agentic framework. The edge runs a single offline captioning pass that builds a dual-track narrative index, an event-level story skeleton plus a clip-level micro-log, cached and reused across queries without re-captioning. At query time, a cloud-side MLLM reasons over the index in a story-first loop centered on a lightweight Visual-Need Router: a per-query gating module that triggers bounded keyframe retrieval only for perceptual questions (appearance, on-screen text, attribute disambiguation) and keeps temporal-structural questions in language space. The router turns visual access into a first-class, query-conditioned cost, capping per-query frame consumption regardless of video length. Experiments on long-video benchmarks demonstrate strong accuracy-efficiency trade-offs while substantially reducing online visual processing.
Quantum Physics
Taking Advantage of Noise in Distributed Random Quantum Circuits
Adding noise can make a random quantum circuit look faster without making its unitary dynamics more random. This distinction is especially relevant in modular processors, where local gates randomize each core and scarce inter-core communication must spread that randomness across the full device. In this paper, we study this problem with a reduced second-moment transfer-matrix theory for Pauli second moments in distributed random circuits affected by the amplitude-damping, depolarizing, and dephasing noise channels. The key step is to resolve the noisy spectrum into two branches: a radial branch, describing dissipative loss of non-identity Pauli weight, and an angular branch, describing Haar-like mixing within the surviving nontrivial sector. This separation gives a simple weak-noise criterion: noise is useful for angular randomization only when it suppresses the longitudinal Bloch component more strongly than the transverse plane. Among the three channels considered, this selects amplitude damping as the only locally favorable case, while depolarizing noise is neutral and dephasing is dominated by radial loss. For multicore architectures, we derive a universal first-order law for radial leakage and track the angular branch numerically across different channels, topologies, and core partitions. The results reveal narrow windows of genuine noise-assisted Haar mixing, most clearly for amplitude damping, but rule out a generic speed-up by noise. The framework therefore distinguishes useful noisy randomization from mere dissipation.
Read the full abstract
Machine Learning
CausalArena: Benchmarking Causal Discovery in the Foundation Model Era
Causal discovery aims to uncover causal structures from data and is fundamental to scientific reasoning and intervention-based decision making. Its evaluation relies heavily on structural causal models (SCMs), which specify a causal graph together with the mechanisms that generate data, yet existing studies differ substantially in graph families, mechanisms, and evaluation protocols. The emergence of causal discovery foundation models (CDFMs) further complicates evaluation: performance may reflect not only causal discovery ability, but also overlap between pretraining environments and test SCMs, making results on fixed synthetic benchmarks difficult to interpret. We introduce CausalArena, a unified and evolvable benchmark for causal discovery under a common protocol. Synthetic SCMs supply controlled breadth over structures and mechanisms; semantic operational SCMs provide human-auditable, semantically grounded environments beyond standard synthetic generators; and formula-grounded SCMs test discovery under explicit scientific mechanisms. Public real-world datasets provide an additional external-validity check. Experiments across classical, neural, and pretrained methods reveal substantial ranking shifts across SCM families and protocols, showing that strong performance in one benchmark regime does not reliably transfer to others. These results highlight benchmark diversity and pretraining–evaluation overlap as central challenges for evaluating causal discovery in the foundation model era.
Read the full abstract
High Energy Physics (Phenomenology)
Mathematical inverse problem for the world data inference of the parton distribution functions of the proton
We show that the inference problem of constraining the parton distribution functions of the proton from deeply inelastic scattering data can be formulated as a linear tensor reconstruction inverse problem. This means that instead of fitting model parameters to the world data, a reconstructive approach to solve a system of coupled linear functional integral equations can be formulated. We leverage the mathematical structures of the integral equations defined by perturbative QCD and global analysis to pose the problem of global analysis as a coupled system of linear integral equations, and construct a proof-of-principle methodology to solve it. To formulate this approach concretely, we review all next-to-leading order accuracy results for inclusive virtual photon, neutral current, charged current, heavy flavor production, and neutrino deep-inelastic lepton–proton scattering. This mathematical methodology of inference opens a path towards a model-bias-free extraction of the proton PDFs from the world data of deeply inelastic scattering in the spirit of indirect measurement employed in mathematical inverse problems, including robust estimation of uncertainties with reduced bias from model parametrization, while closely adhering to the established paradigm of perturbative QCD and global analysis.
Read the full abstract
High Energy Physics (Theory)
Non-invertible Selection Rules from Generalized Discrete Gauging of Finite Non-Abelian Symmetries
We investigate non-invertible selection rules originating from the discrete $H$-gauging of theories with an underlying discrete global symmetry group $G$. To systematically describe these theories, we formulate a general framework for $H$-gauged models that incorporates generalized field transformations. Our approach naturally accommodates non-Abelian groups, for which multidimensional irreducible representations play an essential role. In such models with non-Abelian groups, the transformations induced by $H$ non-trivially mix the internal components of $G$-multiplets, potentially projecting out specific degrees of freedom. Consequently, conventional selection rules based on standard tensor product decompositions or conjugacy classes become insufficient. By analyzing the full semidirect product $G \rtimes H$, we introduce projected characters to derive necessary and sufficient conditions for non-vanishing $n$-point bare couplings. Furthermore, we demonstrate that the remaining field components obey an associative fusion-like algebra governed by their Clebsch-Gordan coefficients. Phenomenologically, these selection rules restrict allowed interactions and impose specific relations among coupling constants. We illustrate our results through concrete examples, including $\Delta(54) \cong \Delta(27)\rtimes \mathbb{Z}_2$ and $S_4 \cong A_4 \rtimes \mathbb{Z}_2$.
Read the full abstract
Computer Vision
3D Point Splatting for mmWave Radar Novel View Synthesis
Solving novel view synthesis (NVS) for millimeter-wave (mmWave) radar requires a renderer that is physically faithful, complex-valued, and multi-viewpoint-tractable. No prior method achieves these three properties simultaneously. Differentiable Monte Carlo (MC) ray tracers implement the radar forward model directly with explicit material modeling and complex outputs, but a joint fit across views takes over an hour per scene. Optical-NVS ports of NeRF, hash grids, and 3D Gaussians train fast but discard phase and replace explicit material modeling with opaque learned features, restricting them to power-only range-azimuth (RA) magnitudes. We propose 3D Point Splatting (3DPS), the first differentiable point splatting renderer for radar, derived directly from the standard solid-angle form of the radar equation. Each oriented 3D point carries a material model built on ITU-R P.2040 permittivity and slab reflection, evaluated in closed form, with the resulting complex phasor splatted into range bins through a precomputed point spread function (PSF). The complex-valued output makes the renderer product-agnostic. The same optimized scene yields analog-to-digital converter (ADC), complex range profile (CRP), and RA outputs through standard fast Fourier transform (FFT) pipelines without retraining for each format. On six outdoor ColoRadar scenes, 3DPS reaches 0.587 mean Pearson correlation on held-out RA images. This is between 1.7x and 5.2x the three optical-NVS baselines (RadarSplat, Radar Fields, DART) and 1.3x the mmIR ray tracer at 1/25 of its wall-clock time. Training takes approximately 3 minutes per scene on a single RTX 4090. Project page: https://3d-point-splatting.github.io/
Read the full abstract
Logic in Computer Science
A Lumpability-Driven Taxonomy of Strong and Weak Stochastic Bisimilarities with Their Congruence Properties
We study the relationships among the stochastic bisimulation-style equivalences over PEPA - Performance Evaluation Process Algebra definable according to the well known notions of lumpability for the continuous-time Markov chains (CTMCs) underlying process terms. Lumpability is a central tool in the analysis of a CTMC, because it results in aggregations of the state space enjoying properties that are useful for efficiently computing the state probability distribution of the original chain. At the level of process terms, various stochastic bisimilarities accounting for activity types and cumulative rates can be defined over PEPA, which induce different kinds of lumping. Since the formalisations of some of them are scattered across the literature, where they appear under different, and sometimes clashing, names, we collect them within a single, uniform framework, renaming each bisimilarity in a consistent way after the kind of lumping it induces. We present strong and weak variants of what we call ordinary, exact, and strict bisimilarities and show that they respectively induce ordinary, exact, and strict lumpings. We then organise the six bisimilarities into a taxonomy establishing all and only the inclusions holding among them. We also analyse how the taxonomy changes in three special cases: process terms whose underlying CTMCs are time reversible, process terms with no activities of unobservable types, and process terms with no recursion. The paper concludes by investigating the compositionality properties of the six bisimilarities. Some of them are not congruences with respect to the prefix and/or choice operators of PEPA. In that case we single out either a set of process terms over which congruence with respect to those operators is achieved, or the coarsest congruence with respect to them that is contained in the considered bisimilarity.
Read the full abstract
Computation and Language
Nuha-Speech: Building General-Purpose Arabic Speech-LLMs
As Speech Large Language Models (speech-LLMs) become increasingly multilingual, Arabic remains significantly underrepresented, highlighting the need for dedicated infrastructure to train and evaluate Arabic speech-LLMs. To address this gap, we introduce Nuha-Speech, a comprehensive initiative to develop general-purpose Arabic speech-LLMs spanning dataset construction, model training, and systematic evaluation. Specifically, we constructed a large-scale Arabic Speech Question-Answering (SQA) corpus comprising over 1.5 million training samples to allow instruction tuning over a broad range of core speech tasks. Then, the corpus was used for supervised fine-tuning based on Qwen-Omni model variants at different scales. Finally, we designed an evaluation framework featuring diverse tasks and tailored metrics. Through this work, we aim to establish foundational infrastructures for Arabic Speech-LLMs under constraints imposed by limited Arabic speech resources.
Read the full abstract
Cosmology and Nongalactic Astrophysics
Reconstructing Early Primordial Black Hole Domination from Gravitational-Wave Backgrounds
Primordial Black Holes (PBHs) with masses below $\mathcal{O}(10^9)$g occupy an interesting region of parameter space that is largely inaccessible to conventional observations. Despite evaporating before Big Bang Nucleosynthesis (BBN), these PBHs can naturally generate a period of early matter domination in the early Universe. A primordial gravitational-wave background (GWB) provides a window onto this otherwise inaccessible regime, since the modified expansion history leaves a characteristic spectral feature associated with the onset and end of PBH domination. The locations of these characteristic frequencies can be used to reconstruct the underlying PBH parameters, in particular the PBH mass. For a freely propagating GWB, the location of these frequencies additionally allow the initial PBH abundance to be determined. We derive simple numerical relations that map these characteristic frequencies directly onto the PBH mass and initial abundance. Future GW experiments span a vast frequency range, providing sensitivity to PBH masses from the BBN bound of $\mathcal{O}(10^9)\,\mathrm{g}$ down to the lower bound of $\mathcal{O}(10)\,\mathrm{g}$. We find that the nanohertz signal reported by NANOGrav, if primordial in origin, is already probing PBH masses in the range $4$–$90\,\mathrm{Mg}$. GW observations therefore offer access to a vast region of PBH parameter space that is otherwise beyond the reach of current experiments.
Read the full abstract
Functional Analysis
Improved Discrete Dual $p$-Hardy and Weighted Discrete $p$- Birman Inequalities
In this paper, we establish a new version of one dimensional generalized discrete dual $p$-Hardy inequality with a shift. Using this generalized discrete dual p-Hardy inequality, we obtain improvements of two discrete dual $p$-Hardy inequalities. To be specific, for $p>1$ and $A\in C_c(\mathbb{N}_{0})$ satisfying $A_{0}=A_{1}=0$, we first improve the discrete dual p-Hardy inequality \begin{align*} &\displaystyle\sum_{n=2}^{\infty}(n-1)^{p}| A_{n}-A_{n-1}|^{p}\geq\frac{1}{p^{p}}\displaystyle\sum_{n=2}^{\infty}|A_{n}|^{p}, \end{align*} where the associate constant term is sharp. Subsequently, we improve its power-type weighted discrete dual p-Hardy extension \begin{align*} &\displaystyle\sum_{n=2}^{\infty}(n-1)^{\alpha}|A_{n}-A_{n-1}|^{p}\geq\Big(\frac{\alpha+1-p}{p}\Big)^{p} \displaystyle\sum_{n=2}^{\infty}\frac{|A_{n}|^{p}}{n^{p-\alpha}} \end{align*} for $p-1<\alpha\leq p$, where the associated constant term is also sharp. We also establish a discrete $p$- Birman inequality with power weights. Furthermore, we establish a multivariable dual $p$-Hardy inequality with a sharp constant. The proof proceeds by first establishing the inequality for two variables and then extending the argument to multiple variables, while preserving the sharpness of the constant.
Read the full abstract
Game Theory
ABRA: An algorithm which cannot converge to low-quality Nash equilibria
We consider a game theoretic approach to solve multi-agent coordination problems with submodular objectives. It is known for such problems that the Nash equilibria for the corresponding game are always within 50% of the optimal. A recent work further shows that the equilibria which achieve this worst-case bound are not stable. Leveraging this, we design an Approximate Best Response Algorithm (ABRA) governed by a noise parameter and a rationality parameter. The noise allows ABRA to escape the bad equilibria and the rationality parameter balances any degradation in the objective function caused by the noise. We show for any two-player game that if ABRA converges to a Nash equilibrium, its system objective value is strictly more than 50% of optimal plus a term controlled by the noise parameter. Otherwise, ABRA converges to some recurrent class: if a recurrent class contains any action profile yielding system objective less than 50% of the optimal, the class must also contain either the optimal action profile or an action profile yielding system objective strictly more than 50% of the optimal by the same amount in addition to a factor controlled by noise parameter. The time that ABRA spends in such action profiles can be controlled using the rationality parameter. Using numerical simulations, we show that the minimum expected objective function is typically well above half of the optimal.
Read the full abstract
Number Theory
The asymptotic in Waring's problem over function fields beyond twice the degree
We prove the expected asymptotic in Waring's problem over $\mathbb F_q[T]$, with a power-saving error, whenever $n>2d$, the characteristic is greater than $(d-1)^2$, and $q$ satisfies an explicit lower bound. This range is sharp in general for the expected asymptotic uniformly in the target polynomial. Our main new input is an aggregate minor arc estimate: we count functionals according to the codimension of their associated singular loci and use intersection theory to bound the degrees of the resulting parameter spaces. In particular, if $n\ge (2+\varepsilon)d$, the required lower bound on $q$ is polynomial in $d$ of degree $2+4/\varepsilon$.
Read the full abstract
Atomic Physics
SI-Traceable Calibration and Performance Benchmarking of a Terahertz Photomixer Transmitter-Receiver System Using a Rydberg Atomic Sensor
Accurate calibration of electromagnetic field strength in the terahertz (THz) frequency regime remains challenging due to the limited availability of SI-traceable field sensors. Here we demonstrate SI-traceable calibration and performance benchmarking of a commercial photomixer-based THz transmitter-receiver system operating near 204 GHz using Rydberg electric-field sensing in a thermal atomic vapor. The THz electric field is extracted from the Autler-Townes (AT) effect of the cesium $17D_{5/2}\rightarrow18P_{3/2}$ Rydberg transition. We measure the strength of AT-split lines as a function of THz detuning from resonance to find the on-resonant Rabi frequency, which, together with atomic transition dipole moments and fundamental constants, yields the THz electric field. The atomically measured field calibrates the photomixer transmitter field and power, while simultaneous measurements with a commercial InGaAs photomixer receiver calibrate the receiver's current-to-field responsivity and convert its current-noise floor into an absolute noise-equivalent THz electric-field sensitivity. Our study demonstrates that Rydberg atomic sensors provide a practical method for SI-traceable calibration and benchmarking of THz transmitter and receiver systems, and for the establishment of a quantitative link between state-of-the-art and absolute atom-based THz sensors.
Read the full abstract
Computer Vision
Guided Super-Resolution of Digital Elevation Models with Diffusion-Based Image Generators
High-resolution digital surface models (DSMs) play an important role in urban analysis, 3D building reconstruction, and infrastructure monitoring, yet their availability remains limited due to the high cost and complexity of data acquisition. In contrast, coarse DSMs from commercial satellite missions are widely accessible, and high-resolution optical imagery is increasingly available from aerial and satellite platforms. We address the resulting mismatch in spatial resolution and propose a DSM superresolution approach that enhances 5 m DSMs to 0.5 m resolution, using guidance from high-resolution spectral images. Our method employs denoising diffusion to transfer information that is visible only in the image, like crisp outlines and detailed roof structures, into the elevation maps. In this way, surface details are reconstructed more accurately than with conventional interpolation or filtering techniques. Experiments on several cities in Central Europe demonstrate that the proposed approach produces high-quality DSMs with improved structural detail and accurate surface geometry. Our results highlight the potential of guided super-resolution with foundational image priors as a means of reconstructing high-resolution surface models.
Read the full abstract
High Energy Astrophysical Phenomena
Hierarchical Population Inference with Normalizing Flows for Binary Black Holes
Low-dimensional parametric mass models are standard in gravitational-wave population inference, but their rigidity can bias the recovered distribution and the conclusions drawn from it. We present a pipeline in which the source-frame binary-black-hole population in $(z,m_1,m_2)$ is represented by a normalizing flow trained directly through the hierarchical likelihood, with event posterior samples propagating measurement uncertainty and detected injections accounting for selection effects. We restrict population statements to the region supported by the injection campaign; conditioning on this region can itself induce an apparent association between mass and redshift. We therefore introduce a mutual-information diagnostic that compares the reconstruction with a redshift-independent reference under the same support restriction, testing for dependence beyond that induced by the support geometry. We validate the method on two simulated catalogues differing only in whether the characteristic primary-mass scale evolves with redshift. The reconstruction recovers the injected mass structure in both, and the benchmarks establish how the diagnostic behaves in the presence and absence of intrinsic evolution. Applied to the GWTC-5.0 catalogue, the method recovers features near $10\,M_\odot$ and $35\,M_\odot$, consistent with the LVK population analysis of the same catalogue, and finds no evidence for intrinsic evolution of the primary-mass spectrum with redshift. More broadly, the pipeline addresses a problem common to many observational sciences: recovering a population distribution from noisy, indirect, and selection-biased measurements of its members.
Read the full abstract
Machine Learning
CoRA-NAS: Coarse Ranking and Anchor-Residual Refinement for Neural Architecture Search
Zero-cost proxies rank architectures cheaply, but their reliability varies across search spaces. We introduce CoRA-NAS (COarse Ranking + Anchor-residual), a two-stage framework combining a static ranking prior with low-cost learning-curve refinement. CoRA-Rank aggregates capacity and structure-at-initialization proxies through an equal-weight log-rank consensus and a target-free consensus gate. CoRA-Refine samples anchors across this prior, extrapolates their early validation curves, and propagates a learned residual correction with an ExtraTrees model. The refinement uses approximately 1% of the cost of fully training the candidate set. Fully trained architecture-accuracy labels are not used to fit the ranker. One configuration is used across spaces, with space-specific architecture encodings. Across NAS-Bench-201, NAS-Bench-101, TransNAS-Bench-101, and NATS-SSS, CoRA-Refine achieves mean Spearman correlations of 0.946, 0.715, 0.786, and 0.894, respectively. Its worst-space correlation of 0.715 is the highest among the compared methods. On NAS-Bench-201/CIFAR-100, its selected architecture reaches 73.32% accuracy, near the reported ground-truth best of 73.37%. On the pure size space, refinement recovers the static prior's shortfall relative to parameter count, while remaining tied with the strongest capacity proxies within noise. The resulting framework combines cross-space ranking robustness with low-cost architecture selection.
Read the full abstract
Complex Variables
Analytic and Algebraic Oka-1 Approximation for Smooth Projective Morphisms with Rationally Connected Fibers
Let $\pi:Z\rightarrow Y$ be a smooth projective morphism of complex manifolds with connected rationally connected fibers. We prove holomorphic approximation on arbitrary compact sets and finite-jet interpolation on arbitrary closed discrete sets for continuous liftings defined on open Riemann surfaces and holomorphic near those sets. For smooth projective morphisms of smooth complex algebraic varieties and algebraic base maps from smooth affine curves, the approximating liftings can be chosen algebraic, with interpolation on any finite set. In both cases the resulting lifting is homotopic to the initial one through continuous liftings of the fixed base map. For connected smooth projective complex manifolds, this gives the equivalence between the algebraic Oka-1 property and rational connectedness. Every rationally connected smooth projective complex manifold is also Oka-1.
Read the full abstract
Cryptography and Security
From Specs to Apps: Verifying and Monitoring Models of Signal and WhatsApp
The Signal protocol is a prominent messaging protocol that secures communication for billions of users. It powers WhatsApp, the most widely used messaging application worldwide, and the Signal app, popular among privacy-conscious users. Extensive research in the computational and Dolev-Yao settings provides strong formal security guarantees for the protocol itself. However, a gap remains between the guarantees of the protocol specification and the implementation's actual behavior at runtime. In this work, we bridge this gap by applying SpecMon, a recently proposed runtime monitor, to check whether observed executions conform to formal protocol models. To this end, we instrument two applications (WhatsApp Web and Signal Desktop) to capture their interactions with the network and the cryptographic components. Using this instrumentation, we develop two multiset-rewrite models that are compatible with Tamarin, thus enabling verification. We derive the first model of WhatsApp Web's implementation of the Signal protocol and the most detailed model to date of Signal's original protocol. Monitoring establishes that observed executions conform to these models, relative to the trusted event extraction and the symbolic abstraction. For the core components of the Signal protocol, we verify authentication and secrecy properties. Finally, monitoring reveals previously undocumented differences between the original libsignal library and WhatsApp's fork. We evaluate our methodology and demonstrate its reproducibility. Developing the WhatsApp Web model, instrumenting the app, adding fuzzing, and running the experiments took three person-weeks. We also demonstrate efficient monitoring of real-world applications and detection of deliberately injected security faults, with low overhead in our measured setting.
Read the full abstract
Strongly Correlated Electrons
Low temperature thermodynamics of $S_{\mathrm{eff}}=1/2$ triangular lattice quantum spin liquid candidate TlYbS$_2$
Geometrically frustrated triangular-lattice antiferromagnets exhibit a delicate competition between magnetic order and quantum spin liquid (QSL) behavior, with the Yb-based delafossite family $A$Yb$X_2$ providing a structurally clean platform for exploring this physics. Here, we report a comprehensive study of single-crystal TlYbS$_2$ using DC magnetization, AC susceptibility, electron spin resonance (ESR), and specific heat measurements extending from room temperature to the millikelvin regime. Single-crystal X-ray diffraction confirms a trigonal $R\bar{3}m$ structure comprising well-separated triangular layers of Yb$^{3+}$ ions with no detectable site disorder. At zero field, a weak thermodynamic anomaly is observed near $530$ mK, which may indicate the onset of a weakly ordered state similar to that reported in KYbSe$_2$. Below $300$ mK, the zero-field magnetic specific heat follows $C_{\rm m}\propto T^{1.8}$, close to a quadratic temperature dependence, in contrast to the linear temperature dependence reported for the sister compound TlYbSe$_2$, which has been described in terms of the interplay between spinons and thermally excited gauge-flux excitations. Magnetization and ESR measurements establish pronounced easy-plane magnetic anisotropy, with $g_\perp/g_\parallel \approx 6.9$. The anomaly near $530$ mK exhibits a strongly anisotropic response to magnetic field. For $H\perp c$, it remains visible up to approximately $2$ T and continuously evolves toward a field-induced ordered phase above approximately $2.5$ T, followed by a sequence of field-induced phases that includes a $1/3$ magnetization plateau between approximately $5$ and $8$ T and full polarization near $17$ T. In contrast, for $H\parallel c$, the anomaly weakens and is suppressed near $3$ T, consistent with the recently reported confinement–deconfinement transition from an ordered state to a field-induced QSL.
Read the full abstract
Group Theory
Conjugator length in finitely generated groups
We describe all functions $\mathbb{N}\rightarrow \mathbb{N}$ that can be realized, up to the standard equivalence, as conjugator length functions of finitely generated groups. Furthermore, we show that any two increasing functions $f,g\colon \mathbb N\to \mathbb N$ can be simultaneously realized as conjugator length functions of finitely generated, commensurable (in particular, quasi-isometric) groups.
Read the full abstract
Astrophysics of Galaxies
Learning JWST. I. A Foundation Model for New Population Discoveries and Morphology-Aware Photometric Redshift Measurements in the JADES Survey
We present FM-JADES-v1, a self-supervised foundation model for James Webb Space Telescope (JWST) deep-field science, trained with 482,444 objects from the JWST Advanced Deep Extragalactic Survey (JADES) Data Release 5 using multi-band imaging and the photometric catalog. The shared embedding space is trained without class labels. We demonstrate that FM-JADES-v1 can serve as a powerful tool for object discovery and improving property measurements using two experiments, blind active discovery and few-band photometric redshift. For blind object discovery, FM-JADES-v1 identifies rare object populations such as high-redshift galaxies and Little Red Dots (LRDs) without any prior population labels or population-specific selection criteria. These rare populations emerge as isolated islands in the embedding space, which can be identified without prior astrophysical knowledge. For few-band photometric redshift, FM-JADES-v1's learned embeddings achieve $\sigma_{\rm NMAD}=0.157$ in a strictly controlled three-band (F115W/F200W/F356W) photo-$z$ benchmark, compared to $\sigma_{\rm NMAD}= 0.44$ for template fitting. These results demonstrate the potential of self-supervised multi-modal representations as scalable discovery spaces for large astronomical surveys. Applied to ongoing and future wide-field surveys from JWST, Roman, Euclid, and Rubin/LSST, this framework could enable systematic searches for rare populations, as well as enabling multiple downstream tasks such as improving astrophysical property measurements.
Read the full abstract
Computation and Language
Domain-Specific Hallucination Detection in Large Language Models
Large language models generate fluent text that can contain unfaithful claims – a phenomenon known as hallucination. We present a multi-signal detection pipeline combining fine-tuned DeBERTa-v3 classification, Monte Carlo (MC) Dropout uncertainty quantification, and temperature-scaled calibration for response-level hallucination detection. Evaluated on the HaluEval benchmark, our pipeline achieves F1=0.915 and AUROC=0.977 on general-domain tasks, with per-task F1 scores of 0.97 (QA), 0.96 (Summarization), and 0.82 (Dialogue). MC Dropout inference further improves accuracy to 93.2%. A context ablation study confirms the model performs genuine entailment reasoning rather than exploiting surface patterns, with summarization F1 dropping 24% when knowledge context is removed. Learning curve analysis reveals that 25% of training data captures 77% of full-data performance. Beyond detection, we apply Direct Preference Optimization (DPO) to a Qwen2.5-0.5B generator, reducing its hallucination rate from 85.5% to 37.7% (55.9% relative reduction) as measured by our detector. Cross-domain evaluation on the SciFact biomedical benchmark shows that general-domain training transfers poorly (F1=0.52), motivating domain-specific fine-tuning. PubMedBERT fine-tuned on SciFact achieves F1=0.63 and AUROC=0.81, demonstrating that domain-matched pre-training is the strongest adaptation strategy. Code and models are available at https://github.com/varunteja99/hallucination-detection-nlp
Read the full abstract
Quantitative Methods
Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens
Many biological discovery problems require experiments to be selected sequentially under constrained budgets. CRISPR screening is a prominent example, as exhaustive perturbation testing is often infeasible and candidate perturbations must instead be prioritized over multiple experimental rounds. Despite the importance of this problem, existing benchmarks for adaptive hit discovery remain limited in scale and diversity. Here, we introduce AssayBench-Loop, a large-scale benchmark for adaptive hit discovery comprising 1,389 CRISPR screens across five phenotype categories. Beyond enabling systematic evaluation, its scale makes it possible to learn acquisition strategies across historical experiments. Building on this resource, we introduce AssayLoop, a sequential experimental design framework combining AssayFormer, a transformer-based amortized acquisition policy trained across historical screens to adapt from experimental feedback, with LLM-derived biological priors through an adaptive handoff. In this view, completed experiments become training data for learning how accumulated evidence should guide what to test next, while LLMs provide prior biological knowledge to seed the search. We further introduce AssayLLM, showing that the same principle can be extended directly to an LLM through task-specific post-training. On temporally held-out screens, AssayLoop achieves a 5.67-fold enrichment over random selection and recovers 27.7% of hits after assaying approximately 5% of the candidate library, outperforming existing adaptive-design methods and standalone LLMs, and AssayFormer alone. Performance improves with increasing historical training data and transfers to phenotype categories excluded from training. These results demonstrate the value of learning acquisition policies across historical experiments and combining them with broad biological priors for efficient adaptive hit discovery.
Read the full abstract
Artificial Intelligence
On the Regularization Landscape for the Linear Recommendation Models
Recently, a wide range of recommendation algorithms inspired by deep learning techniques have emerged as the performance leaders on several standard recommendation benchmarks. While these algorithms were built on different DL techniques (e.g., dropouts, autoencoder), they have similar performance and even similar cost functions. This paper studies whether the models' comparable performance are sheer coincidence, or they can be unified under a single framework. We find that all linear performance leaders effectively add only a nuclear-norm based regularizer, or a Frobenius-norm based regularizer. The former ones possess a (surprising) rigid structure that limits the models' predictive power but their solutions are low rank and have closed form. The latter ones are more expressive and more efficient for recommendation but their solutions are either full-rank or require executing hard-to-tune numeric procedures such as ADMM. Along this line of finding, we further propose two low-rank, closed-form solutions, derived from carefully generalizing Frobenius-norm based regularizers. The new solutions get the best of both nuclear-norm and Frobenius-norm world.
Read the full abstract
Robotics
UniMPA: A Unified Memory-Prediction-Action Model via Action-Grounded Transition Modeling
Recent advances in Vision-Language-Action (VLA) models have improved robotic manipulation, yet observation-to-action learning remains limited by a fundamental transition realizability gap, manifested in three tightly coupled problems: (i) Transition ambiguity. Visually similar current observations may correspond to different manipulation phases and imply different subsequent transitions. (ii) Prediction–execution mismatch. A visually plausible predicted future observation does not necessarily correspond to a physically realizable transition. (iii) Experience–realization mismatch. A historically executable action pattern may not necessarily realize the intended transition in the current scene and therefore requires context-aware adaptation. Accordingly, we propose UniMPA, a Unified Memory-Prediction-Action model that addresses these problems through a shared action-grounded transition interface. (i) UniMPA introduces Persistent-Selective Future Prediction to resolve transition ambiguity by modeling the intended future state evolution. A persistent latent stream continuously tracks task-level progress, while a transition-critical pixel stream selectively resolves fine-grained interaction changes through memory-grounded prediction. (ii) To assess the physical executability of the anticipated transition, the predicted transition queries a temporal Visual-Action Memory Bank. The bank retrieves historically realized visual-action experience, grounding future prediction in executable evidence. (iii) To adapt executable experience to the current scene, an Action-Visual Memory Bank retrieves visually grounded action prototypes from historical action evolution. Prototype-Biased Flow then shifts the flow source toward a historically supported action manifold for context-aware refinement.
Read the full abstract
Accelerator Physics
Physics-Informed Drift Diagnosis for Laser-Plasma Accelerator Operations
Laser-plasma accelerators (LPAs) sustain accelerating gradients of order $100\,\mathrm{GV/m}$, but routine operation remains difficult: electron beam metrics drift over an operating shift, and the root physical cause is often invisible to the available diagnostics. We formulate LPA operation as a latent state-space model in which three effective interaction-point variables, the normalized laser amplitude $a_0$, the normalized plasma electron density $\tilde n_e$ and the residual pulse chirp $\mathcal{C}$, are inferred from routine electron beam observations by an extended Kalman filter. The emission model, which maps the latent state to the diagnostics, is kept structurally separate from the {transition} model, which describes how the latent state evolves between shots. The separation supports diagnosis in two stages, one asking which latent variable moved and one asking what moved it. The implemented emission model is a toy model, yielding an expected performance in line with current facilities and using 3D blow-out regime dependencies where relevant. We conduct synthetic sessions to test the effectiveness of the detection and attribution protocols, finding that attribution is limited by excitation rather than by shot count or diagnostic resolution. Because the construction needs only a set of physical latent variables, an emission model and a family of hardware-derived transition models, it transfers to other drift-prone subsystems. We argue that the accuracy of the whole procedure is limited by the emission model rather than by the inference method.
Read the full abstract
Machine Learning
The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
Recursive self-improvement (RSI) enables AI systems to turn experience and feedback into persistent changes that improve both their capabilities and the process of future improvement. We first use the Headroom-Closed Index (HCI) to reveal the problems of existing LLMs, then introduce the RSI concept and its development roadmap: from improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement. Next we examine RSI across scenarios (e.g., scientific discovery, embodied intelligence, software engineering), highlighting their distinct requirements and development speeds. Drawing on diverse industry practices and preliminary empirical evidence, we connect RSI research with practical systems and identify key challenges to achieving genuine RSI.
Read the full abstract
Machine Learning (Statistics)
Evaluating Time-Series Foundation Models and Multimodal Dietary Context for CGM Forecasting
Continuous glucose monitoring (CGM) provides high-frequency measurements of glucose dynamics and enables short-term glucose forecasting for diabetes management. Although time-series foundation models have shown strong general forecasting ability, their effectiveness for CGM prediction and the added value of multimodal dietary context remain unclear. We conduct a comprehensive empirical study using eight public CGM datasets spanning Type 1 diabetes, Type 2 diabetes, and non-diabetes populations. Under a unified protocol across multiple context lengths and prediction horizons, zero-shot foundation models did not consistently outperform strong task-specific baselines such as Elastic Net and PatchTST. In contrast, lightweight fine-tuning substantially improved forecasting performance. For example, fine-tuned Chronos-Bolt reduced RMSE by 6.5%-18.4% in the T1D cohort and by 8.6%-18.2% in the non-diabetes/T2D cohort, with comparable improvements in both in-distribution and out-of-distribution test settings. We further evaluate multimodal dietary context using CGMacros, which provides temporally aligned CGM signals, food images, and macronutrient records. A residual-based fusion framework reduced overall RMSE by approximately 3% and postprandial RMSE by approximately 15% relative to the CGM-only baseline. Moreover, Chronos-based CGM representations were more strongly correlated with observed postprandial glucose increments than representations from LSTM and CatBoost, even after those models incorporated additional dietary modalities, suggesting that pretrained temporal representations better preserve meal-induced excursion patterns. These findings show that foundation models require CGM-specific adaptation for reliable forecasting and that dietary context provides clinically meaningful signals beyond CGM alone, especially during postprandial periods.
Read the full abstract
Robotics
Learning Agent-based Model Predictive Control for Holistic Vehicle Performance
Agent-based model predictive control (AMPC) has recently been proposed as a distributed scheme that collaborates with all agents to achieve optimal holistic performance. However, its optimality highly depends on the prediction accuracy that requires all agents or their contributions to be known, which is too idealistic for actual implementation. This research proposes a novel practical hybrid control scheme - learning agent-based MPC (LAMPC), combining the model-based AMPC approach and data-based learning methods to improve the holistic vehicle performance for multi-agent systems. The Gaussian process regression (GPR) enhanced by an online data management strategy serves as the learning core to predict unknown contributions. A novel multi-step prediction mechanism leverages the GPR learning potential along the horizon. The predicted mean, representing the learned unknown contributions, completes the system model in the MPC for more accurate control. Meanwhile, a stochastic framework is formulated to guarantee control safety and feasibility using soft chance constraints based on the prediction variance. Both simulations and experiments show that, with the learning capability, LAMPC outperforms the traditional AMPC. LAMPC can achieve higher tracking performance in well-learned scenarios and always guarantee constraint satisfaction even in less-learned scenarios. Moreover, the proposed hybrid control scheme is efficient for real-time implementation and is flexible to any control agent topology.
Read the full abstract
Computation and Language
Augustinian BabyLM: What Ostensive Definition Can and Cannot Teach a Small Language Model
A language model normally begins training with random word embeddings: whatever 'banana' means must be learned from training corpora. I implement St. Augustine's picture of word learning, meaning by ostension, for a small masked language model (DeBERTa) trained on 10M words: before training, visually grounded tokens receive embeddings derived from the image regions they label; other tokens start random. Visual initialization leaves a measurable imprint that lasts until the end of training. At the same time, the effect remains invisible under most BabyLM benchmarks, which probe abstract grammatical knowledge: visual initialization does not affect performance there. The only zero-shot exception is object-property knowledge (COMPS), where seeding helps in every configuration. To follow up on this result, I build a corpus-tailored version of the Visual-Property Swap benchmark, which tests color, material, size, and shape knowledge, with per-item training frequency and seeded status. Here, vision-seeded models have a persistent, seed-replicated advantage. Function words and abstract vocabulary also receive strong visual seeds and retain them throughout training, and the training objective draws on them: held-out mask-prediction loss falls for these words in every seed. However, no benchmark I run registers this. What evaluation would pick this up remains an open question.