Computer Vision
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Tuesday 15 September
Multi subject video editing modifies designated subjects while preserving non target content, but faces cross subject attribute leakage, and occlusion ambiguity. Existing approaches rely on masks and struggle to distinguish overlapping subjects or ensure consistent generation. To address these limitations, we propose MDN-Control, a training free framework jointly controlling target localization, occlusion geometry, and appearance initialization. Specifically, mask-guided localization provides consistent target localization, while depth-aware occlusion control resolves ambiguous boundaries between overlapping subjects. We further introduce noise latent prompting, which retrieves Gaussian initializations from a noise library for prompt relevant priors. Experiments on MSVBench show that MDN-Control achieves the lowest CM-Err and the highest Q-Edit, while maintaining competitive text alignment and temporal consistency, demonstrating the effectiveness of combining spatial, geometric, and latent priors for multi subject video editing.
Optics
Deterministic Ground-State Search in a Spatial Photonic Ising Machine by Phase Retrieval
A spatial photonic Ising machine (SPIM) solves large-scale combinatorial optimization problems by computing the Ising Hamiltonian through an optical Fourier transform. However, the ground-state search relies on annealing, in which spins are optimized stochastically and sequentially. We propose a ground-state search based on phase retrieval (PR), in which all spins are updated simultaneously and deterministically. Imposing an amplitude constraint in the Fourier plane guides the modulated phase distribution, which encodes the spin configuration, toward the optimal solution. We numerically demonstrate that the proposed scheme reaches the ground state of rank-one Ising Hamiltonians with $10^4$ spins in a single iteration for all trials. Moreover, a radial rearrangement of the amplitudes and an appropriately designed target pattern alleviate search stagnation and improve the optimization of spin configurations. The collective and deterministic spin update by phase retrieval enables a fast ground-state search for large-scale combinatorial optimization.
Read the full abstract
Medical Physics
Closing the gap: A two-site MLC matching study for Varian TrueBeam Linear Accelerators
Purpose: To characterise and harmonise multileaf collimator (MLC) gap calibrations across four beam-matched TrueBeams and assess a common planning-system model. Methods: Four Millennium 120 TrueBeams were studied: TS1 (installed 2016), TS5 (2021), TS6 and TS7 (2024). Gap calibration was assessed by RayStation sweeping-gap measurement, feeler gauge, the EPID-based Stakitt Fence test, and a leaf-resolved Machine Performance Check (MPC). TS1 and TS5 were adjusted in service mode, TS5 also for MLC centreline offset. Ten clinical VMAT plans, and six acquired earlier on an unadjusted reference machine, were measured with an ArcCHECK 1220 against doses computed at 0.053 and 0.065 cm x-offsets in RayStation. Results: The 2024 machines had wider baseline gaps. All four methods registered the TS1 and TS5 changes in the same direction and within 0.060 mm of predictions. After adjustment TS1 and TS5 lay within the TS6/TS7 sweeping-gap and Stakitt ranges but read wider on the feeler gauge. A conventional picket-fence test on TS5 registered no change. Raising the RayStation x-offset parameter increased mean local 3%/2 mm pass rates from 92.3% to 94.8% on TS1 and 93.2% to 95.6% on TS5; all 24 plan-criterion pairs improved on the unadjusted machine. Conclusions: Mechanical, EPID, and sweeping gap MLC gap measurements responded consistently to the MLC adjustment, moving TS1 and TS5 towards the local reference range. Every global 3%/2 mm ArcCHECK measurement exceeded the TG-218 95% limit.
Read the full abstract
Machine Learning
Online Gradient Computation for Warping Gaussian Process Transformations
Warped Gaussian processes (GPs) handle non-Gaussian observations by mapping them into a latent standard GP via a parametric transformation called warping. Existing streaming variants, however, either optimize the warping parameters periodically or sacrifice analytical tractability for a higher model capacity. To bridge this gap, we show that the gradient of the instantaneous negative log-likelihood of a warped GP admits an exact recursive computation. Based on this result, we propose a novel online method for warped GPs that jointly updates the latent GP moments and optimizes the warping parameters.
Read the full abstract
General Relativity and Quantum Cosmology
Study of Null Geodesics and their Stability in Kalb-Ramond Black Holes
We investigate null geodesics and their stability in four-dimensional charged Kalb-Ramond black holes with a cosmological constant. The non-vanishing vacuum expectation value of the background antisymmetric tensor field induces spontaneous Lorentz symmetry breaking, controlled by a dimensionless parameter $l$, and deforms the Reissner-Nordström-(A)dS geometry. We derive the effective potential for massless particles and obtain analytic expressions for the photon-sphere radius, critical impact parameter, capture cross section, and deflection trajectories. The limiting cases $l\rightarrow0$ recover the RN, RN-dS and RN-AdS geodesic structures, while finite values of $l$ shift the turning points, modify the bending of light and deform both first-kind and second-kind photon trajectories. In the AdS branch, we find a special limiting limaçon-type null geodesic whose angular structure is modified by the Kalb-Ramond parameter. We also compute the Lyapunov exponent of the unstable circular null orbit and show that the cosmological constant changes the instability timescale, whereas the Lorentz-violating parameter produces a non-trivial deformation of the photon-sphere instability. Our results identify null geodesics as sensitive probes of Lorentz-violating effects in charged black hole spacetimes.
Read the full abstract
Analysis of PDEs
Extending the Córdoba-Martínez-Zoroa IPM Blow-Up to Uniformly Space-Time Smooth Forcing
By adapting the techniques used in the IPM blowup result of Córdoba-Martínez-Zoroa with spatially smooth force, we prove finite-time blow-up for the IPM equation on $\mathbb T^2$ with a uniformly spacetime smooth force. In particular, we show there exist a smooth odd initial density, a smooth odd force $F\in C^\infty([0,1]\times\mathbb T^2)$, and a classical solution $\rho$ on $[0,1)$ whose density gradient and spatial velocity gradient diverge in $L^\infty$ as $t\uparrow 1$. Nevertheless, $\rho(t)$ converges in $C^\eta$ for every $0\leq\eta<1$.
Read the full abstract
Physics and Society
Unified framework for measuring segregation resolves how social and geographical space jointly shape connections
Our understanding of how geographical and social segregation interact remains limited, as relatively few studies investigate them jointly, and existing approaches often lack a framework distinguishing geographical, social, and total segregation. Additionally, large-scale individually resolved geo-social network data are rarely publicly available. We address both. Conceptually, we develop a unified framework that measures segregation in geosocial networks by comparing network models to appropriate null models and recovers the Theil index, dissimilarity index, and network modularity as special cases. Empirically, we turn to privacy-preserving aggregated relational data (ARD): we combine the Facebook Social Connectedness Index for the US with US Census and Pew data, and introduce an ARD-compatible joint geosocial intervening-opportunities model to infer link probabilities between region–group cell pairs. Applying our segregation framework, we find that social segregation predominates over geographical segregation, with notable separation for White–Black, college-degree–no-degree, and high-income–low/middle-income across both segregation types. We find increasing social homophily with geographical distance and group-specific geographical connectivity patterns, suggesting that geographical segregation may affect cross-group connectivity not only directly but also by amplifying social segregation.
Read the full abstract
Quantitative Methods
GPCR Ligand Bioactivity Prediction with Physics-Informed Dual-State Query Learning
Predicting the bioactivity profiles of small molecules against G protein-coupled receptors (GPCRs) is a challenge in drug discovery. Although deep learning has accelerated the prediction of binding affinities, existing approaches often struggle to distinguish between functional efficacies because they neglect dynamic conformational equilibria. Furthermore, structure-based methods are frequently limited by the scarcity of high-resolution active-state crystal structures and the indistinguishability of conformational states in static representations. To bridge the gap between black-box prediction and biophysical reality, we propose Dual-State Query (DSQ), a physics-informed multimodal architecture that explicitly embeds the Monod-Wyman-Changeux (MWC) model of allostery within a neural network. Unlike conventional models that rely on explicit 3D structures, DSQ utilizes learnable orthogonal queries to extract disentangled representations of active and inactive receptor states. These latent representations are governed by a novel neural MWC gating module, which mathematically derives the probability of receptor activation from thermodynamic competition between ligand-state affinities and the receptor's intrinsic conformational energy barrier. A contrastive ranking objective is also introduced to enforce differential affinity constraints, ensuring physical consistency. Extensive experiments demonstrate that DSQ outperforms other baselines, particularly for the agonist subset. Additional homology-stratified, temperature-sensitivity, perturbation, clustering, and efficiency analyses show that DSQ provides useful thermodynamic inductive bias, while also exposing a clear limitation on low-homology receptors. The code is available at https://github.com/jiankliu/DSQ.
Read the full abstract
Analysis of PDEs
On the Stability of Type II Blowup for the Keller-Segel System in High Dimensions
We study finite-time blowup for the parabolic–elliptic Keller–Segel system on $\mathbb{R}^d$ in dimensions $d\geq11$, where the problem is mass supercritical. For every integer $l\geq2$, we construct smooth radially symmetric solutions whose radial mass variable concentrates the normalized stationary state $Q$ at a quantized scale. Each blowup regime can be realized by solutions with nonnegative population density throughout their classical lifespan. More precisely, near the blowup time $T$, \[ u(t,r)=\frac{1}{\lambda^2(t)}\left[Q\left(\frac{r}{\lambda(t)}\right) +\epsilon\left(t,\frac{r}{\lambda(t)}\right)\right], \qquad \lambda(t)=c(T-t)^{\frac{l}{\gamma(d)}}(1+o(1)), \] where $c>0$ and $\gamma(d)=\frac12\bigl(d-2-\sqrt{(d-2)(d-10)}\bigr)$. The remainder $\epsilon$ converges to zero in local $L^\infty$ norms and in a range of high-order homogeneous Sobolev norms. Since $2l>\gamma(d)$, the concentration scale is strictly smaller than the parabolic scale $\sqrt{T-t}$, and the resulting blowup is of type II. The $l$-th regime has exactly $l-1$ unstable radial modulation directions and is stable within a codimension-$(l-1)$ class of suitably regular radial initial data. The proof combines a generalized-kernel expansion driven by the algebraic tail of $Q$, modulation analysis, coercive weighted high-order energy estimates, and a finite-dimensional topological argument. This yields a quantized hierarchy of stationary-state concentration rates for the high-dimensional Keller–Segel flow.
Read the full abstract
Materials Science
Scaling LLM Agents for Materials Design through Hierarchical Collective Reasoning
Materials design must reconcile competing functional requirements with stability and synthesis constraints. Generative models can produce stable, novel crystals, but accommodating detailed natural-language design instructions remains challenging. Here we introduce HiMatGen, a framework for scaling large language model agents through hierarchical collective reasoning. Built from GPT-5.6 Terra, HiMatGen connects 100 investigators in discussion pods across ten scientific domains with tool-enabled domain representatives. Representatives investigate proposals, exchange evidence across domains and return findings and unresolved questions to their pods. This bidirectional exchange turns specialist disagreements into structural revisions and alternative designs. Across six chemical systems, HiMatGen produces 1.9 times as many final crystal candidates and 1.7 times as many stable, unique and novel (SUN) structures as a tool-enabled GPT-6 Astra single-agent baseline. In property-directed tasks, the complete HiMatGen workflow outperforms same-model single-agent and independent-generation controls. Comparisons with MatterGen and Chemeleon2 demonstrate its ability to develop diverse crystals satisfying joint functional and chemical requirements. In a paired comparison initialized from the same proposals, hierarchical discussion reduces debate-stage token usage by a factor of 22 relative to full debate. HiMatGen provides an approach to materials design that sustains the exchange of scientific knowledge and computational evidence throughout iterative exploration as the agent population grows.
Read the full abstract
Graphics
HairCS: Reconstructing Strand-Based Hair from Hair Cards
We present an automated pipeline that converts hair-card models into high-quality strand-based hairstyles. Given a collection of textured triangular or quad strips as input, our method produces a strand-based representation that preserves the original hairstyle while enriching it with fine-scale geometric detail and adhering to standard production requirements: strands originate from the scalp, roots are uniformly distributed, and the hair volume is plausibly filled. The resulting assets are directly compatible with strand-based rendering, physics-based simulation, and common grooming modifiers (e.g., clumping, curling, noise) for enhanced realism and artistic control. We validate our approach on a large and diverse set of hairstyles, including short and long hair, curly styles, and complex styles such as buns and ponytails.
Read the full abstract
Computer Vision
A multimodal large language model for evidence-based autism spectrum disorder screening
The clinical management of autism spectrum disorder (ASD) faces a bottleneck in early screening, mainly because trained specialists are scarce and conventional assessment tools are subjective. Here, we introduce ASDchat, a multimodal large language model designed for evidence-based ASD screening, which takes video, audio, and dialogue as input. ASDchat adopts a dual-branch architecture, where the decision branch generates screening probabilities and the evidence branch generates traceable, timestamped behavioral evidence aligned with the Autism Diagnostic Observation Schedule (ADOS-2). The model was trained and evaluated on a dataset of 1,035 participants from 27 sites in China, which covered typically developing (TD) children, children with ASD, and children with other developmental disorders. For ASD versus TD, ASDchat reached an area under the receiver operating characteristic curve (AUC) of 0.953 $\pm$ 0.021. On 6 held-out sites that were not used for training, the mean AUC was 0.961. Furthermore, unsupervised clustering of the behavioral dimensions split the ASD cases into six subtypes with different phenotypic profiles, and ASDchat suggests an intervention for each subtype. ASDchat provides a feasible path for large-scale, evidence-based ASD screening in clinical practice.
Read the full abstract
General Relativity and Quantum Cosmology
Weak in the Presence of Beauty: Gravitational waves from the mergers of black holes and neutron stars as a messenger
In 2025, the 4th Gravitational Wave Transient Catalog reported the discovery of over 200 gravitational wave events since 2015. These gravitational waves (GWs) are emitted by compact binary coalescenses (CBCs) between neutron stars and black holes, and detected with a global network of ground-based laser interferometers. GWs represent a new `messenger' – ripples in spacetime that propagate across cosmological distances, and carry information about the physical properties of compact objects and their location in the Universe. In this review, we consider what information we can learn about black holes and neutron stars using GWs as a messenger. We explore how the GW observation process allows us to transform the outputs of laser interferometers like LIGO, Virgo and KAGRA into knowledge about black holes and neutron stars. We show that the GW source population acts as a messenger that can lead us to breakthroughs in our understanding of stellar evolution, how the Universe has changed across cosmic time, and fundamental physics. Finally, we discuss three key opportunities to learn more about our Universe using GWs as a messenger in the coming decades.
Read the full abstract
Cryptography and Security
Not All Relations Are Equal: Relation-Balanced and Calibrated Graph Learning for Provenance-Based Intrusion Detection
Provenance-Based Intrusion Detection Systems (PIDSs) detect Advanced Persistent Threats (APTs) by analyzing system interactions. However, existing methods largely treat relations uniformly, overlooking statistical heterogeneity; in CADETS, relation frequencies differ by approximately $140{,}000\times$. This may cause PIDSs to focus more on frequent relations and overlook differences in normal error levels across relations, increasing the risk of false alarms and missed detections. We present RECAL, an unsupervised framework using relation-balanced masked graph learning to better capture rare interaction patterns. It further calibrates reconstruction errors against each relation's benign error distribution to produce comparable anomaly evidence, helping distinguish attacks from benign behavior and reduce false alarms. On three DARPA E3 datasets, RECAL achieves F1 scores of 99.99%, 99.93%, and 99.99%, outperforming the best baseline on each dataset by 0.88, 0.82, and 0.42 percentage points, respectively. Compared with the baseline reporting the lowest FPR, RECAL reduces mean FPR by approximately $105\times$, $4\times$, and $41\times$.
Read the full abstract
Software Engineering
Protocol-Preserving Context Trimming for Agentic Workflows: Benefits, Failure Regimes, and Budget Guardrails
Agentic large language model (LLM) systems rely on long interaction histories to preserve instructions, tool states, intermediate decisions, and unresolved dependencies, but unrestricted context growth increases computational cost and can reduce efficiency. This study evaluates protocol-preserving context trimming as a reliability-constrained approach for multi-step agentic workflows. Five trimming strategies - recency-based, relevance-based, summarization, protocol-aware trimming, and adaptive budget guardrails - were compared across retained-context levels and workflow-complexity classes using task success, protocol adherence, valid tool calls, token savings, latency reduction, cascading failures, and critical context thresholds. Conventional strategies achieved about 60% mean token savings but lower task success (66.6-77.3%) and protocol adherence (85.5-88.6%). Protocol-aware trimming improved task success to 92.2%, while adaptive guardrails achieved 96.0% task success, 96.3% protocol adherence, and 1.0% cascading failure with 56.0% mean token savings. Retained-context budgets of 25% or less increased failure odds 10.92-fold relative to budgets of 50% or more (p < 0.001). Protocol-aware trimming produced 5.24-fold greater odds of successful completion than conventional methods under aggressive budgets, while adaptive guardrails further increased success odds 2.11-fold versus fixed protocol-aware trimming (p < 0.001). Critical context thresholds also increased with workflow complexity. These findings indicate that reliable context reduction depends more on preserving protocol-critical state than on maximizing token removal, and that adaptive guardrails can improve efficiency, scalability, and reliability in long-horizon agentic systems.
Read the full abstract
Quantum Physics
Environmental records unlock universal quantum computation from thermal decoherence
At a fixed thermal exposure, the same stabilizer processor can be classically simulable or quantum universal, depending on which environmental records its controller retains. We give an exact computational classification of energy-counting thermal-idle instruments in a quantum processor with ideal stabilizer control and independent local Markov baths. The relaxation time $T_1$, homogeneous coherence time $T_2$, and equilibrium excited-state population $p_e$ determine an exact computational boundary at $(1-p_e)T_2/T_1=1$. If every location lies at or below it, branchwise nonnegative stabilizer decompositions give an explicit efficient classical sampler for the adaptive circuit and its full time-resolved exchange record. Above it at a single repeatedly accessible location, a suitable idle duration and no-exchange conditioning supply distillable ancillas and enable universal quantum computation with polynomial overhead. At finite temperature on the resource side, erasing the record at a sufficiently long, unsplit exposure makes the averaged channel stabilizer measure-and-prepare, and even a terminal parity check then yields only simulable branches. Yet at that same exposure, retaining only the bit recording whether any exchange occurred still heralds distillable ancillas, because a thermal round trip restores the parity after its first exchange has already removed the coherence.
Read the full abstract
Machine Learning
OPD-Aha: From Linguistic Momentum to Visual Reflection in Multimodal On-Policy Distillation
Privileged on-policy distillation improves multimodal reasoning by allowing a teacher to evaluate student trajectories using rich, training-only visual evidence. Both models score these trajectories while conditioning on the same student-generated prefix. When a student misinterprets an image early in a response, this accumulating erroneous rationale eventually pulls the teacher away from its visual evidence. The teacher and student converge on the same hallucination, causing standard cross-model supervision to collapse precisely where correction is most needed. We find that the teacher's visual corrective preference is not lost under this misleading agreement. Comparing the predictions of the identical teacher given the real image and a visual null reveals that the privileged evidence still pushes the model toward the correct interpretation. We introduce OPD-Aha, which reconstructs the distillation target directly from this isolated visual preference rather than relying on the fragile teacher-student discrepancy. This reconstructed target aggressively suppresses continuations that contradict the image. Trained with this objective, students learn to naturally interrupt their own flawed reasoning with reflection tokens such as wait and actually. After reflection, subsequent generation relies less on the accumulated erroneous text and more on the visual evidence. Correcting these trajectories mid-generation fundamentally alters the reasoning process, yielding broad and consistent improvements across diverse fine-grained perception and complex multimodal reasoning benchmarks. Our code and models are available at https://github.com/Echochef/OPD-Aha.
Read the full abstract
Audio and Speech Processing
Language Orthogonalization for Zero-Shot Cross-Lingual Audio Deepfake Detection
Audio deepfake detectors need to transfer to languages absent from training, as multilingual speech synthesis outpaces labeled anti-spoofing resources. While detectors increasingly rely on self-supervised speech models (S3Ms), these backbones encode language-dependent structure that confounds spoof cues. We address this confound through language orthogonalization, a target-free ridge map that removes S3M variation projected onto continuous language-identification (LID) embeddings. Across six languages, six S3M backbones, and all Leave-N-Out settings, it consistently reduces EER across unseen languages. Cross-lingual EER correlates with LID-space distance, where orthogonalization yields larger gains for more distant transfers.
Read the full abstract
Differential Geometry
Cyclic Riemannian nilmanifolds are naturally reductive
We study the structure of cyclic Riemannian nilmanifolds, that is, connected nilpotent Lie groups $N$ endowed with a left-invariant cyclic metric $\langle\cdot,\cdot\rangle$. We prove that a Riemannian nilmanifold $(N,\langle\cdot,\cdot\rangle)$ is cyclic if and only if its Lie algebra $\mathfrak{n}$ is at most two-step nilpotent and the associated family of skew-symmetric endomorphisms $J_{\mathfrak{z}}=\{J_Z : Z\in \mathfrak{z}\}\subset so(\mathfrak{a})$ is Abelian. As a direct consequence, every cyclic Riemannian nilmanifold is naturally reductive. We also determine the full isometry group of a connected and simply connected cyclic Riemannian nilmanifold without Euclidean factor.
Read the full abstract
Solar and Stellar Astrophysics
The Variation of Circumstellar Parameters through Early Stellar Evolution
We report on the evolution of extinction, mass accretion rate, and spectral index from analysis of SpeX observations and AllWISE spectral energy distributions, compared to stellar parameters (mass, age, and veiling) derived from analysis of previously reported and newly analyzed iSHELL observations of the same young stellar objects (YSOs). By using stellar gravity as an age indicator, we are able to infer age without any assumptions about the circumstellar environment. Our sample consists of 109 Class I, II, and III YSOs. The spectral index broadly decreases with time but with sufficient scatter that Class I-III YSOs can all be found at the same age. There is a wide range of extinction (A_V) from nearly 0 to greater than 30 mag for ages up to 5 Myr, showing no clear trend with age. The mass accretion rate declines from a median of 10^-8 solar masses per year at ~1 Myr to less than 10^-10 solar masses per year at ages greater than 10 Myr, with a scatter of the mass accretion rate of +/- 1 dex. Accretion seems to effectively end by ~ 5 Myr, by which time the median mass gained is 0.05 solar masses. We find a wide range of veiling (rK) from zero to three (limited by our need to detect photospheric lines) for Class I and II YSOs, with the median veiling declining after ~2 Myr. Class I YSOs appear to be physically very similar to Class II YSOs, with similar accretion rates and overlapping age and mass distributions. There are no objects with a high spectral index and low extinction, suggesting that once starved of circumstellar material, the spectral index rapidly declines.
Read the full abstract
High Energy Physics (Theory)
Stringy Gauge Structures
A field theory action, able to incorporate stringy gauge symmetry enhancement-breaking effects, which occur at particular points of moduli space in a heterotic string toroidal compactification framework, was previously constructed from three-point interactions and duality symmetries. Via doubled periodic coordinates and a non-commutative product, it encodes the non-Abelian structure in a background-independent form. Moduli-space dependence becomes manifest upon performing a generalized mode expansion. Here we further elaborate on this construction. In particular, we propose field transformations for the retained sectors that make up the foundations of a potential symmetry of the action and that, when mode expanded at enhancement points in moduli space, contain the usual non-Abelian gauge transformations. We also examine the covariance of the corresponding scalar derivative structures. Although the proposal must be regarded as a truncated version of a full, presently unknown field-theory description, it already encodes representations of massless and massive states containing up to one oscillator excitation. Finally, we relate the construction to gauged DFT with an extended tangent space by associating generalized Kaluza-Klein modes considered here with additional generalized-frame directions. At enhancement, their brackets reproduce the enhanced gauge algebra; away from enhancement, the same algebraic structure is recovered by rotating the Cartan basis.
Read the full abstract
Artificial Intelligence
Fine-Tuning Fixes Mode Collapse and Over-Dispersion in LLMs
Recent work by Doshi and Hauser (2024), Bisbee et al. (2024), and Xie et al. (2026) raises concerns that outputs from large language models (LLMs) tend to be under-diverse: they repeat or resemble one another more often than responses from the population they are meant to represent, a phenomenon known as mode collapse. In this work, we show that whether mode-collapse, or its opposite, occurs depends on the specific model and dataset used. Further, with sufficient supervised fine-tuning (SFT) data, LLM output diversity converges toward that of the target distribution from which fine-tuning data are sampled. To quantify this comparison, we measure the probability that two responses sampled independently from the same fixed prompt coincide (collide), or their expected similarity under a kernel. We derive a bias-variance decomposition of the expected gap between the model's and target's collision probabilities, showing that SFT is not inherently biased toward mode collapse or its opposite: finite-sample SFT can leave a model either under- or over-dispersed, depending on the model and dataset. Finally, we show that the absolute gap is bounded by the square root of the Kullback-Leibler (KL) divergence from the target distribution to the model. Consequently, a model sufficiently close to optimal under population cross-entropy cannot exhibit arbitrarily miscalibrated diversity. We test the decomposition and the bound in three experiments: small transformers on synthetic languages, four LLMs fine-tuned on human surveys, and these LLMs fine-tuned on CodeNet, a dataset of human code solutions. More target data moves model diversity toward the human (or synthetic target) level in all experiments, consistent with our theoretical predictions. These results show that diversity miscalibration can arise from finite-sample error and shrink as SFT better approximates the target distribution.
Read the full abstract
Information Retrieval
Predicting Partial Answer Quality and Utility in Agentic Retrieval-Augmented Generation
Agentic Retrieval-Augmented Generation (RAG) has become a promising paradigm for multi-hop question answering, where a reasoning model iteratively issues queries to a retriever and incorporates newly retrieved context into subsequent reasoning steps. While this iterative process can improve final answer quality, current evaluations of agentic RAG largely focus on end-to-end outcomes and provide limited visibility into how a model's answer state changes during generation. In this work, we introduce an in-trajectory probing framework to study intermediate answer states in agentic RAG. Specifically, after each retrieval-reasoning iteration, we force an agentic model to stop reasoning and generate an intermediate answer based on its current state. This allows us to define two iteration-level measures: partial answer quality at each iteration, and partial utility as the change in partial answer quality across iterations. Our analysis across multi-hop QA benchmarks reveals that partial answer quality often plateaus before natural termination, with many later iterations contributing only small measurable improvements. Accordingly, we formulate two prediction tasks, partial answer quality prediction and partial utility prediction, and study trajectory-derived signals from intra-iteration, inter-iteration, and query-iteration perspectives. Experiments show that partial answer quality is more predictable than partial utility, with supervised models achieving Pearson's r above 0.43 for quality prediction. Finally, using predicted answer quality and utility for early stopping reduces average iteration count by about 11% while preserving about 98% of the final answer quality achieved by natural stopping.
Read the full abstract
Information Retrieval
PCap: Personalized Retrieval-Stage Diversity Capping in Facebook Marketplace
We propose a personalized capping framework (PCap) to improve the diversity in Facebook Marketplace by introducing user-level diversity constraints at the retrieval stage. PCap models individual diversity preferences using Shannon entropy-based scoring, segments users into diversity buckets, and applies personalized category caps during multi-source candidate retrieval. To navigate the high-dimensional parameter space of per-bucket caps, we leverage an automated online optimization method called Parameter Tuning Sequence. Large-scale online experiments demonstrate that PCap significantly improves users' browsing experience shown in engagement metrics. This work provides practical insights into integrating personalized diversity into industrial retrieval systems.
Read the full abstract
Statistics Applications
Observational constraints on net radiative forcing confirm aviation contrail warming
Contrail cirrus represents a critical component of aviation's non-CO2 climate impact, but its net radiative forcing, the balance between longwave warming and shortwave cooling, remains poorly constrained by direct observations. As a result, current assessments rely almost exclusively on microphysical models such as CoCiP and global climate simulations. Existing empirical estimates are largely restricted to young, linear tracks, because satellite detection masks have a poor recall of contrails once they spread and merge with natural cirrus, leaving a structural gap in our understanding of long-lived, non-linear contrail cirrus. To address this we use a causal framework that isolates the net radiative contrail effect of flight traffic over the Americas. Building on recent progress that quantified the longwave warming contrail effect using advected flight paths as a proxy for contrails, we expand this continuous treatment approach to capture the highly skewed shortwave cooling impact, delivering a 12-hour lifespan net observational radiative forcing. Our analysis reveals a statistically significant net warming energy forcing of 33.7 (95% CI: 20.8, 47.8) GJ per km flown from April 2019 to April 2020, providing a large-scale empirical quantification of long contrail lifespan impact of the same order as, though somewhat larger than, previous bottom-up simulation estimates. This observational benchmark offers an independent line of evidence on the sign and magnitude of the climate impact of contrails.
Read the full abstract
Computation and Language
Early-Bird Decoding: Accelerating Diffusion LLMs with Learnable Block Sizes and Parallel Sampling
Diffusion large language models (dLLMs) offer a promising parallel decoding paradigm as an alternative to autoregressive generation through iterative unmasking. However, dLLMs typically require many steps before token confidence reaches the decoding threshold, resulting in inefficient inference even with block-wise KV caching. To accelerate dLLM inference, we for the first time propose an "early-bird (EB)" decoding framework, motivated by the observation that tokens with similarly low entropy tend to cluster and can be jointly decoded earlier, before reaching the confidence threshold. In particular, our EB-Decode framework integrates two key enablers: (1) a learnable network that adaptively groups tokens with similar uncertainty into variable-length blocks, rather than relying on fixed block sizes; (2) a position-aware sampler that learns to unmask tokens in parallel using fewer decoding steps within predicted variable-length blocks. Both components are developed without modifying pretrained dLLM weights and can therefore be directly deployed as plug-ins during serving, with negligible training and inference overhead. Extensive experiments across three models and four benchmarks consistently validate our observation and the effectiveness of EB-Decode, achieving 3.53-18.76$\times$ higher throughput than the vanilla decoding method and up to 1.58$\times$ higher throughput over the strongest baseline, Fast-dLLM, with comparable accuracy.
Read the full abstract
Optimization and Control
Beyond Phase Reduction: Amplitude Collapse in Optimal Control of Coupled Oscillators
We solve the four-dimensional Hamilton-Jacobi-Bellman (HJB) equation for two diffusively coupled Stuart-Landau-like oscillators to obtain full-state optimal feedback control. A sweep over coupling strength reveals a sharp change in the numerically optimal strategy: below a threshold coupling value, the controller steers the phase difference toward anti-phase while keeping both oscillators near the limit cycle, as reduced-order models would suggest. Above this threshold, the HJB solution changes qualitatively; the controller transiently collapses one oscillator's amplitude to near zero, thereby enabling large phase repositioning near the origin before rebuilding its amplitude. Direct gradient-based and stochastic optimization do not recover this lower-cost collapse trajectory from the initializations considered, suggesting that it occupies a region of the control landscape that is difficult to access by direct search. A joint sweep over nonisochronicity and coupling shows that collapse can occur even for an isochronous oscillator: phase repositioning near the origin can favor an off-cycle strategy. Nonisochronicity provides an additional energetic benefit through a phase-velocity surplus at small amplitude, quantitatively accounting for the observed reduction in control cost. Comparisons with uncoupled and coupled phase-reduced baselines show that phase models become increasingly inaccurate and cost significantly more energy for strong coupling. Results for coupled Van der Pol oscillators further demonstrate that exploitation of off-cycle dynamics is not specific to the Stuart-Landau-like oscillators.
Read the full abstract
Computer Vision
Decentralized Gossip Learning and Federated Averaging for Histopathology Image Classification
Breast histopathology analysis increasingly relies on distributed learning because direct data pooling across institutions is often restricted by privacy, governance, and communication constraints. This study compares server-based Federated Averaging (FedAvg), fully decentralized gossip learning, and Hybrid Gossip-FedAvg for invasive ductal carcinoma (IDC) patch classification. Experiments used 277,524 color image patches with patient-disjoint training, validation, and test partitions and a workload-balanced, Dirichlet-guided allocation across six nodes. Ring, random degree-3, and fully connected gossip topologies were evaluated together with sensitivity analyses for statistical heterogeneity, mixing coefficient, learning rate, model drift, prediction disagreement, calibration, clinically motivated operating points, communication payload, and patient-level IDC burden, together with auxiliary backbone robustness analyses. In the principal alpha=0.3 experiment, Hybrid Gossip-FedAvg achieved a test area under the receiver operating characteristic curve (ROC-AUC) of 0.8811, closely followed by FedAvg at 0.8801 and fully connected gossip at 0.8751. Across three independent patient-level repetitions, FedAvg and Hybrid Gossip-FedAvg obtained the same mean ROC-AUC of 0.9082, with standard deviations of 0.0037 and 0.0043, respectively. Hybrid achieved the highest mean area under the precision-recall curve of 0.8240, whereas FedAvg produced the lowest mean Brier score of 0.1335. Denser gossip graphs improved discrimination but increased theoretical model payload, while ring gossip remained sensitive to learning rate and mixing strength. Overall, FedAvg provided the most consistently reliable server-based baseline, topology-aware gossip offered a viable decentralized alternative, and Hybrid Gossip-FedAvg provided a balanced compromise between peer-to-peer diffusion and periodic global coordination.
Read the full abstract
General Relativity and Quantum Cosmology
Charging of rotating black holes: kinetic simulations of black hole magnetospheres
Whether a rotating black hole (BH) immersed in an external magnetic field charges up to the Wald value $Q_{\rm W}=2aMB$, where $M$ and $a$ are the BH mass and spin parameter, and $B$ is the strength of the external field, is a long-standing open question in BH electrodynamics, with consequences for charge separation, particle acceleration and the structure of BH magnetospheres. We address it with axisymmetric general-relativistic particle-in-cell simulations performed with GRZeltron, including self-consistent pair creation. For representative values of the BH spin, we evolve the same asymptotically uniform Wald field from two opposite initial horizon charges, $Q_{0}=0$ and $Q_{0}=Q_{\rm W}$, and track the accumulated charge through the BH horizon. We find that both branches relax within a few tens of gravitational times to the same saturated charge value. Thus, the equilibrium is a dynamical attractor of the kinetic magnetosphere rather than a memory of the initial data. The attractor lies well below $Q_{\rm W}$, at $\xi_{\rm eq}\equiv Q_{\rm eq}/Q_{\rm W}\approx 0.3$ for $a\lesssim0.7$, then falling almost to $\xi_{\rm eq}\approx0$ for large spins $a\approx1$. We derive an analytic expression for $\xi_{\rm eq}$ by requiring equal magnetic fluxes through the horizon associated with positive and negative charges, and find its spin dependence to be in agreement with the simulations. The charge state of an astrophysical BH is therefore driven by kinetic plasma processes, and the Wald charge is an upper bound, not a general equilibrium value. Since $\xi_{\rm eq}<1$, charging does not quench the horizon-infinity potential drop that powers energy-extraction processes such as the Blandford-Znajek mechanism.
Read the full abstract
Machine Learning
Adaptive Bayesian Partner Selection for Federated Clinical Centers
Federated learning (FL) in healthcare faces pronounced heterogeneity and temporal concept drift across clinical centers, where evolving patient populations and care practices shift data distributions. Existing approaches rely on persistent global communication, incurring substantial bandwidth overhead while risking negative transfer from poorly aligned peers. We propose Adaptive Bayesian Partner Selection (ABPS), a peer-to-peer framework that governs who collaborates, when, and at what cost. Each center maintains a Beta-Bernoulli posterior over prospective peers' Shapley marginal utility, ranks candidates with an Upper Confidence Bound (UCB) criterion, and forms collaborations through a lightweight propose-reject mechanism, with the option to abstain from communication when no mutually beneficial partner exists. The framework admits a stochastic decision interpretation, yielding finite-sample concentration guarantees and O(kappa log T) regret in partner selection, along with conditions under which intentional isolation is optimal under negative transfer. Lightweight extensions (head personalization, bfloat16 quantized communication, and a tunable active-set size) further improve efficiency, and a goal-aware metadata filter enables institution-specific collaboration strategies. On binary in-hospital mortality prediction over the first 24 hours of an ICU stay, with 230 non-IID clinical centers drawn from MIMIC-IV, the full ABPS-X variant matches the strongest federated baseline (FedDyn, AUROC 0.758) at 0.09x the communication cost of FedAvg, with reduced variability. A diversity-driven configuration activates intentional isolation for a substantial fraction of centers. These results show that adaptive, utility-aware collaboration reduces communication without sacrificing accuracy when centers are numerous and small, offering a scalable paradigm for healthcare FL.