In briefDriver emotion recognition using traffic context and expectation deviation
SummaryThis paper proposes TADER, a framework that recognizes driver emotion by combining in-cabin facial images with out-of-cabin traffic scenarios. It models emotion as arising from expectation deviation, the gap between predicted and observed vehicle trajectories, fused into the recognizer through attention-based embedding. Transfer learning from a trajectory prediction model addresses the shortage of paired data. The authors report a 5% accuracy improvement on the AIDE and manD datasets over state-of-the-art methods.
In briefPrivacy-preserving federated learning on curved model spaces with projection-based aggregation
SummaryThe paper introduces DP-RFedProj, a record-level differentially private federated learning method for compact submanifolds that aggregates via closed-form projections and adds noise in a fixed ambient geometry. It gives an explicit f-DP guarantee for the whole training transcript and proves convergence when clients run Riemannian SGD and the server uses projected averaging. Experiments on federated PCA and EEG classification with SPDNet show higher utility than a Euclidean baseline, using five to six times less noise than the closest prior method.
Thibault Pautrel, Florent Bouchard, Guillaume Ginolhac et al.
In briefE3Stereo: Better Stereo Matching at Depth Edges Using Learned Geometric Edges
SummaryStereo matching often fails at depth boundaries. This paper proposes E3Stereo, which learns geometric edges from disparity supervision and combines them with backbone disparity and image details to predict full-resolution disparity corrections. On Scene Flow, it reduces endpoint error at depth boundaries by 9.5% to 14.3% across RAFT-Stereo, IGEV, and MonSter, and also lowers overall error.
In briefIris: Typed Visual Option Decisions with Reusable Memory
SummaryThis paper introduces Iris, a framework for agents choosing among discrete candidates that may be images, text, or both, such as UI element selection or chart matching. The authors prove that candidate-order invariance and exact cache reuse hold under Iris, and compare independent and joint candidate encoding on a benchmark built from Multimodal-Mind2Web. A proposed design encodes each candidate conditioned on the request without cross-attention between candidates, reporting 91.7 / 86.5 / 90.0 percent accuracy at about one-third of OneJev's latency.
In briefGender Bias in AI-Assisted ADHD Self-Diagnosis
SummaryThis paper examines how demographic cues shape GPT-4o's outputs when people use large language models for medical self-diagnosis, focusing on ADHD. The authors adapt the Context Association Test to measure gendered ADHD associations in a zero-shot self-diagnostic setting. They find that explicit gender cues consistently amplify certain ADHD subtype associations, while implicit cues produce distributions close to proxy-aligned references.
Matilde Elene Hansen, Ross Deans Kristensen-McLachlan
In briefMemory-Efficient Training for Unrolled MRI Reconstruction Networks
SummaryThis paper proposes a memory-efficient training framework for unrolled variational networks used in accelerated MRI reconstruction. It updates one randomly sampled cascade at a time while keeping supervision from the final reconstruction loss, decoupling training memory from chain length. On fastMRI knee and brain datasets, the method matches or surpasses E2E-VarNet at certain acceleration factors, with fewer parameters, lower peak memory, and faster per-iteration training.
WhyDecouples training memory from cascade length via sampled cascades and light backprop, matching or beating E2E-VarNet on fastMRI; impact mainly MRI reconstruction.
In briefEmbedding-Based Educational Taxonomies: Automatically Building Question Groupings for Learning
SummaryThis paper clusters semantic text embeddings of math questions to build educational taxonomies computationally. It compares these taxonomies with existing CCSS-based knowledge-component taxonomies on two ASSISTments datasets. The embedding-based taxonomies correspond substantially but incompletely to the existing ones, and improve Deep Knowledge Tracing performance by about 3-5 AUC points.
In briefPathKGen: A knowledge graph-assisted multimodal LLM for explainable histopathology reports
SummaryThis paper addresses automated histopathology report generation from whole slide images (WSIs). It proposes PathKGen, which pairs a visual branch built on pathology foundation models with a knowledge branch that queries a pathology-specific knowledge graph for key morphological features. Evaluated on the WSI-Bench dataset, it reportedly outperforms prior state-of-the-art methods and presents its diagnostic reasoning as visual graphs.
Nadun Rajapaksha, Sandun Herath, Devnith Wijesinghe et al.
In briefPACE: A Proof Agent That Revises Conjectures While Researching
SummaryThis paper presents PACE, a research-proof agent that selects mathematical objectives, develops or challenges candidate mechanisms, and builds proofs, with an independent verifier checking them against the original target. Its evidence-tracking memory records assumptions and dependencies and flags evidence for revalidation when conditions change. The authors also introduce ResearchProofBench, an optimization-focused applied-mathematics benchmark with 51 known-proof tasks and six open-problem cases. PACE reaches an aggregate Proof Score of 75.16 and completes 20 proofs on the known-proof tasks, and its conjecture-research submissions receive higher review scores than the compared baselines.
In briefRecovering Pruned Visual Tokens: Post-Training for Multimodal LLMs
SummaryThis paper studies pruning visual tokens to cut inference cost in multimodal large language models and whether post-training can recover the lost performance. The authors post-train 36 pruning configurations on a 282K-sample mixture covering VQA, reasoning, OCR, and grounding. Post-training recovers most of the loss, and the authors find that pruning at the vision encoder rather than the LLM determines prefill speed.
WhyPost-training across 36 configurations recovers most pruning loss, and encoder-side pruning proves more effective than projector-side for latency.
In briefEXE: Egocentric cross-embodiment pretraining with factorized latent actions for VLA models
SummaryVision-Language-Action (VLA) models are limited by scarce robot data. This paper proposes EXE, a framework that learns factorized latent action representations from egocentric human manipulation videos, separating shared manipulation priors from embodiment-specific motion signals. This lets robot policies acquire generalizable manipulation semantics from human demonstrations without being biased by human kinematics, and the pretraining benefits both parallel grippers and dexterous hands. The authors report consistent gains over a no-pretraining baseline on RoboTwin 2.0, RoboCasa GR1, and real-world manipulation tasks.
In briefMORE: Adaptive Reasoning Modes and Effort for Self-Evolving Agents
SummaryThis paper proposes MORE, a self-evolving training framework that lets agents choose both the reasoning form (direct answer, short or long chain-of-thought, or code) and the amount of computation for each task. It uses a two-stage process: first building a prior over these modes, then refining it with MATE-PPO on generated, verifiable task streams. Reported results on coding, math, and general reasoning benchmarks show consistent gains for 3B to 14B models, with shorter responses and lower inference time.
In briefMeta-CLIP: Using scan metadata to improve zero-shot disease detection from medical images
SummaryThe paper introduces Meta-CLIP, a medical vision-language pretraining method that uses acquisition metadata from DICOM headers to reweight cross-modal negatives. On CT-RATE it improves macro-AUROC over CT-CLIP. Inference uses the standard dual-encoder setup and does not require metadata.
Haozhe Luo, Nirosh Sivanesan, Shelley Zixin Shu et al.
In briefBeyond Affine Reconstruction: Transformer DeepONets for Multi-Resolution Radiative Transfer Modelling
SummaryThis paper proposes a transformer-based DeepONet surrogate for the radiative transfer equation, which is a computational bottleneck in fire simulations. A Perceiver encoder maps temperature and absorption fields at different resolutions into a fixed latent representation, and a cross-attention decoder performs non-affine reconstruction of radiative intensity. On a two-dimensional pool-fire dataset, the model outperforms autoencoder-based DeepONets in predicting radiative intensity and incident radiation with fewer parameters. At one matched latent dimension, its training error falls below the theoretical lower bound for affine reconstruction.
In briefSWE-Ladder: Training coding agents to localize, reproduce, and patch bugs in stages
SummaryThis paper proposes SWE-Ladder, a training method for software-engineering agents that resolve repository-level issues. It splits each verified repair into stage-specific evidence for localization, reproduction, and patching, routes each agent step to a stage by its tool call, and distills a frozen teacher's guidance in repair order while optimizing with GRPO. On SWE-bench Verified with Qwen3-4B-Instruct-2507 under the R2E-Gym scaffold, pass@1 rises from 9.6% to 16.4%, compared with 10.4% for GRPO and 12.4% for OPSD with the full patch.
In briefTracking Is Not Sensing: Artifact-Aware Thermal Video Respiration Estimation in Head-Free Marmosets
SummaryThis paper examines whether marmoset respiratory rate can be estimated from thermal video using synchronized contact-sensor recordings. Respiration is recoverable from thermal video but is highly sensitive to the region of interest, and visually plausible nasal tracking under head motion can yield traces inconsistent with contact measurements. The authors propose Selective Temporal Attenuation Regularization (STAR), which applies local temporal imputation to high-motion thermal discontinuities during training, reaching 2.23 bpm MAE versus 20.76 bpm for tracked ROI analysis.
In briefPower-of-Two Leaky Integrate-and-Fire Neurons for Energy-Efficient Vision-Language-Action Robot Models
SummaryThis paper introduces the Power-of-Two Leaky Integrate-and-Fire (PoT-LIF) neuron, which emits signed spikes with power-of-two amplitudes to replace dense continuous computation in vision-language-action models. The authors integrate it into the projection layers of the π0.5 model and test it over 2,000 paired closed-loop LIBERO episodes per configuration. Success rates match or slightly exceed the continuous-activation baseline of 96.70%, and at T=1 with a 60.01% firing rate the modeled arithmetic energy is about 1/7.67 of the dense reference.
In briefFreeView: Robust 3D Reconstruction from Messy Image Collections
SummaryFreeView is a training-free framework for building 3D scenes from uncurated image collections that contain irrelevant or non-overlapping views. It uses the consensus across the collection to pick a representative view as a filtering reference, then rejects distractor views by combining visual and geometric evidence. The method works with VGGT and pi^3 without needing a manually curated reference or manual calibration. Experiments across multiple datasets report high clean-view retention and distractor rejection rates.
In briefSafe reinforcement learning with physics-informed graphs and constraint-specific critics for coordinating community batteries in distribution networks
SummaryThe paper addresses voltage and line-limit violations in distribution networks caused by high rooftop solar and EV charging, proposing a physics-informed safe deep reinforcement learning framework to coordinate community battery storage. It uses a dual-level graph representation shaped by line impedance and capacity, and an algorithm called LDCPPO that uses separate reward and cost critics. In case studies with high photovoltaic and EV penetration, the method reduces the key voltage-violation degree by up to 80.5% versus tested baselines, with the lowest violation count and line overload among compared methods and competitive operating cost.
In briefJointly Modeling LLM Preferences and Imperfect Decision-Making
SummaryThe paper introduces the Preference-Behavior Model (PBM), which combines a preference representation with mechanisms for imperfect utility maximization and estimates both from observed choices as an inverse-game problem. It evaluates the model in three social dilemmas: a Public Goods Game, a Collective-Risk Dilemma, and a Common-Pool Resource dilemma. Preferences estimated from standard behavioral economic instruments do not consistently predict behavior, but models fitted in one dilemma retain predictive value in others.
Axel Abels, Elias Fern\'andez Domingos, Tom Lenaerts
In briefTAD: Training Robot Vision-Language-Action Policies from Teacher Trajectories Alone
SummaryThis paper proposes Trajectory-only Adversarial Distillation (TAD), which trains a student robot policy without access to teacher action scores. A discriminator compares teacher and student joint-position feature distributions and supplies rewards for on-policy learning. On the ManiSkill PutOnPlateInScene25Main task, TAD substantially improves OpenVLA-OFT (7B) but trails white-box on-policy distillation slightly, and the best run uses more interaction and optimization budget.
In briefPreView: Improving GUI Grounding with Preprocessed Views and Cues
SummaryThis paper addresses GUI grounding, which means finding the small on-screen control that a natural-language instruction refers to in a cluttered screenshot. The authors propose PreView, an auxiliary preprocessor that improves the inputs to an existing grounding model without training that model further. A lightweight auxiliary model refines a center point, crops and enlarges a local view, and adds a visual red-circle cue and a textual description of the target. Across five benchmarks, it improves average grounding accuracy for Qwen3-VL-2B, OS-Atlas-4B, and Qwen3-VL-4B.
In briefAlphaIdeaBench: Benchmark for Automated Research Idea Generation Under Resource Limits
SummaryThe paper introduces AlphaIdeaBench, a benchmark of 100 research-ideation tasks, each with an explicit resource budget and a shared proposal format. An LLM judge scores proposals on 12 dimensions, and its novelty scores are checked against reviewer-derived scores from 600 ICLR 2026, NeurIPS 2025, and ICML 2025 papers. The authors evaluate eight published ideation harnesses and several backbone LLMs, and release the tasks, rubrics, and scores for reproducible benchmarking.
In briefLENS: Likelihood-based scoring and neighbor smoothing for person search on unseen cameras
SummaryThe paper addresses attribute-based person retrieval, where accuracy (mADM) drops on unseen cameras because predicted attribute probabilities shift. It proposes LENS, which combines fine-tuned DINOv3 and CLIP models with a distilled student, uses group cross-entropy for mutually exclusive attributes, scores gallery images by the exact query log-likelihood, and smooths gallery logits over nearest neighbors. The authors report that these changes raise mADM and reach 68.17 mADM on the hidden test set of the UPAR 2027 benchmark.
Nguyen Dang Duy Tan, Vo Thanh Nguyen, Linh T. T. Tran et al.
In briefUniQuery: A Joint Query Head Unifying Vision Models for Cross-Domain Pedestrian Attribute Recognition
SummaryThe paper proposes UniQuery, which jointly fine-tunes the DINOv3 and CLIP encoders for recognizing pedestrian attributes such as gender, age, and clothing. Its prediction averages the logits of three query heads, including a joint head that attends to tokens from both encoders. In leave-one-dataset-out experiments on the UPAR Challenge 2027 benchmark, it outperforms an independent ensemble on Challenge_Avg and reaches 0.8001 on the hidden test set.
Nguyen Dang Duy Tan, Linh T. T. Tran, Huynh Nguyen Phat et al.
In briefBuild or Adopt? Comparing In-House and General-Purpose LLM Agent Harnesses for Car GUI Testing
SummaryVolvo Group compared its in-house LLM agent harness with two general-purpose harnesses, Claude Code and OpenCode, for GUI testing of Android in-vehicle infotainment systems, keeping tools and models the same. General-purpose harnesses matched the custom harness on stepwise tests but cost 2.6 to 10.7 times more per task in inference. Including engineering effort, total monthly cost for a nightly regression suite was reported as 3 to 66% lower.
Arsham Gholamzadeh Khoee, Ruyu Fu, Jiaqi Wei et al.
In briefDROA-CLIPSeg: A method for segmenting faint cracks in low-light images using text prompts
SummaryThis paper addresses pavement crack segmentation under low light, where reduced contrast and shadows degrade results. It proposes DROA-CLIPSeg, a parameter-efficient framework that adapts a frozen CLIPSeg vision-language model with two lightweight modules, and introduces CrackLingo-9K and CrackLingo-12K, text-augmented versions of existing crack datasets. Training about 1.93M parameters (1.3% of the full model), the method reportedly outperforms 9 zero-shot and 4 fine-tuned baselines on both splits and generalises zero-shot to MVTec AD.
WhyCombines frozen CLIPSeg with low-rank adaptation and orthogonal residual branches for low-light crack segmentation, a meaningful combination. Extensive baselines, zero-shot transfer, and ablations sup
Vinita Shukla, Partibha Jangra, Vinayak Pathak et al.
In briefPASS Is a Claim, Not an Observation: Re-verifying Self-Reported Fixes in a Multi-Agent System
SummaryThis paper mechanically re-verifies 10 code-level fix claims self-reported by a coordination agent on AITRAP, a live multi-agent knowledge-tree platform. The claims were extracted from 179 coordination broadcasts and checked for file timestamp consistency, presence of the claimed defensive pattern, and rule-based verification from the announcement. Among claims with enough information to re-verify, 50% failed the identifier and call-site check. The authors conclude that self-asserted PASS declarations require independent verification.
In briefA model-theoretic formalization of al-Khwarizmi's inheritance calculations (ilm al-faraid)
SummaryThis paper formalizes al-Khwarizmi's treatment of inheritance using model theory and computation, within the field of Computational Fiqh. It shows that every inheritance case in the studied class can be expressed as a finite system of rational linear equalities and inequalities, which can be exactly decided by Fourier-Motzkin elimination. The authors also define a fixed-dimensional inheritance structure and prove that its algebraic representation is canonical within that class.
In briefONOT/TONO-Print: A paired benchmark testing face-photo quality checks after ID document printing
SummaryThis paper introduces ONOT/TONO-Print, a small paired benchmark for testing face-image quality models on identity-document portraits altered during production. It uses ONOT and TONO portraits with synthetic variants for five production techniques, plus a smaller set of physically printed and scanned portraits. The authors compare biometric and ICAO-related quality components before and after production. They report that production causes substantial domain shift for some quality estimators, while others remain comparatively stable.
In briefAllocation Before the Clock: Graph, Set and Game Theory of Pro LoL Objective Setups
SummaryThis paper studies how professional League of Legends teams assign their five players to the pit, lanes, and commitments before neutral objectives. It models a setup as an allocation on a walled map, using geodesic Voronoi control and a quantal response equilibrium priced in win probability. Across 164 games from MSI and Worlds 2025, trailing teams leave the pit empty far more often than leading teams at the first baron and Atakhan.
In briefRAQ-MLA: Role-Aware Quantization for Multi-Head Latent Attention Caches
SummaryThis paper studies how to compress the key-value cache of Multi-Head Latent Attention (MLA), an attention design used by recent frontier language models. The authors find that the compressed latent mainly supplies values while the positional key mainly shapes attention scores, so each tensor needs its own transform. They propose RAQ-MLA, a training-free method that quantizes the MLA cache to about 2.5 bits per element. The method is reported to give a 6.5x increase in theoretical KV-cache capacity while staying nearly lossless on reasoning and coding tasks across four MLA models.</summary>
In briefRSI Arena: A study of recursive self-improvement guided by human feedback
SummaryThis paper introduces RSI Arena, in which ten frontier AI models act as agents that autonomously post-train NVIDIA Nemotron 3.5 Lightning 30B-A3B Base over a six-day first stage on a shared 64-GPU cluster. Models are assessed on 20 benchmarks and blind human evaluations. In a second stage, agents receive human comparison records and three more days of training, and eight of ten models improve in style-adjusted human preference. Benchmark gains were modest and not clearly linked to preference gains, and some models produced longer responses or more harmful compliance.
In briefGraph-Dependent Pooling: Adaptive Granularity in Hierarchical Graph Neural Networks
SummaryThis paper proposes Graph-Dependent Pooling (GDPool), a method that lets hierarchical graph neural networks choose the pooling granularity for each graph instead of using fixed cluster counts or ratios. It uses self-attention to pick graph-specific anchor nodes and builds coarser graphs through soft node assignment. On metal-organic framework and molecular graph benchmarks, GDPool reports the best predictive performance, higher modularity than existing methods, and a monotonic decrease in active clusters with depth.
In briefGeodesic-SVDD: one-round distributed tabular learning with latent mixup on a hypersphere
SummaryThis paper addresses distributed learning on privacy-sensitive tabular data, where tree ensembles are hard to federate and deep classifiers forget under label skew. It proposes Geodesic-SVDD, which treats each client's task as one-class density estimation on a hypersphere, using a Residual Spherical Encoder and Geodesic Latent Mixup. It also introduces a One-Shot Product of Experts protocol that aggregates local models in a single communication round. The authors report the highest mean AUROC among seven federated methods across five benchmarks and four skew levels, and one-shot results comparable to centralized Deep SVDD under IID and mild skew.
Andrea Moleri, Christian Intern\`o, Markus Olhofer et al.
In briefRecirculated AI Articles: How Errors Persist in Model-Mediated Information
SummaryThis paper studies how errors persist when LLM-generated articles become inputs for later models. In a controlled community writing experiment, recirculating published articles raised the final-round presence of complete false claims by 7.29 percentage points while reducing explicit endorsement. Abstracting error details left reproduction largely unchanged across most targets, with local reductions on one quotation target.
In briefSemantic-guided frequency modulation that cuts image bitrate while keeping edges
SummaryThe paper proposes SG-AFM, which uses a semantic structural prior computed only at the encoder to adaptively attenuate high-frequency subbands of encoder latents. No semantic map is transmitted and no decoder modules are added. Across six datasets, it reduces PSNR-based BD-rate relative to TCM by an average of 13.17%, and achieves the best BD-rate on ExDark, NightCity, and OnePlus with a mean reduction of 16.80%.
In briefError analysis of triangular transport maps for conditional generation using MMD objectives
SummaryThis paper studies theoretical error guarantees for conditional generative models built from triangular transport maps, trained with MMD-based objectives. The authors show that minimizing a joint MMD can generically fail to recover the true conditionals in the misspecified parametric setting. They analyze an estimator that directly minimizes the conditional criterion, proving a non-asymptotic error bound with fast rates independent of ambient dimension. For joint MMDs with concentrating Gaussian kernels, they derive an error bound that suffers from a curse of dimensionality in the conditioning variable.
WhyProves joint MMD minimization is generically inconsistent for conditionals and gives non-asymptotic bounds with dimension-free fast rates for the conditional objective.
Michele Martino, Bamdad Hosseini, Amir Sagiv et al.
In briefSatellite depth mapping of Gulf ports, tested on unseen sites
SummaryThis paper estimates water depths of Abu Dhabi port areas from Sentinel-2 satellite images and compares them with about 285,000 multibeam sonar soundings. It tests empirical, machine-learning and deep-learning models and finds that the validation method changed results more than the choice of model. At an unseen site the Stumpf ratio performed best, and with local control points a hybrid model reached about 1.11 m error over 0–11 m depth.
Muhammad Saeed, Fatema Al Hajri, Mohamed H. Alhosani et al.
In briefMycelium: A Meta-Study of Hallucination Detection Across Benchmark Provenance, Types, and Granularities
SummaryThis paper treats hallucination detection as a benchmark-dependent measurement problem and organizes existing benchmarks along provenance, typology, and granularity. It analyzes how construction, annotation, and evaluation choices shape the prediction task, including how span boundary variation affects results. The authors propose Span Coverage F1, a containment-based complementary metric, to clarify what detector scores capture and the limits of cross-benchmark comparison.
In briefData Governance for Generative and Agentic AI: Technology, Law, Economics, and Security
SummaryThis book presents a unified governance framework for systems that retrieve, generate, infer, remember, and act on information, from large language models and retrieval-augmented pipelines to autonomous multi-agent systems. It covers the full lifecycle, including data quality, provenance, data protection and intellectual property, economics, cybersecurity and zero-trust architecture, ethics, and governance in elections and the public sphere. Anchored in Indian law and institutions, it takes a comparative view and reviews AI regulation across six jurisdictions, ending with a design for a governance control plane for 2030.
In briefH-GLow: Faster Graph Generation via Hazard-Guided Discrete Flow Matching
SummaryThis paper addresses the difficulty of generating high-quality graphs in few steps, given strong dependencies between nodes and edges. It proposes H-Glow, a framework that decomposes CTMC generators into hazards and conditional jump distributions to plan graph edits and adjust sampling time to graph activity. H-Glow can be applied to pretrained graph discrete flow matching models, and the paper also introduces a new model trained via conditional generator matching. Experiments on molecular and synthetic graph benchmarks report improved generation quality with substantially fewer sampling steps.
In briefTesting ML Models Under Data Shift by Separating Population Change from Pipeline Data Faults
SummaryDeployed machine learning models can lose accuracy when their input data distribution shifts, which may stem from population changes or pipeline data integrity issues. The paper proposes reproducing observed changes from both sources on a controlled baseline to estimate current performance and the potential gain from repairing data integrity problems. Experiments on real-world datasets with simulated faults show that separating the two sources generally improves both estimates. A churn prediction case illustrates support for monitoring and repair prioritization.
Joran Leest, Ilias Gerostathopoulos, Claudia Raibulet et al.
In briefAuditing data leakage and evaluation flaws in LLM rare disease diagnosis
SummaryThis paper audits large language model systems for rare disease diagnosis using over 60,000 cases from public and independent clinical cohorts. It finds that candidate-space contamination and data leakage inflate reported performance, with cross-dataset Recall@1 falling from 54.3% to 40.3% under strict data isolation. The authors conclude that current evaluation protocols largely reflect leakage rather than genuine medical reasoning, and propose a framework combining strict data isolation with standardized benchmarking safeguards.
In briefMix Forcing: causal pre-training for interactive video world models using mixed attention
SummaryInteractive video world models need autoregressive generation, but converting pretrained bidirectional video diffusion models directly with teacher forcing can damage their visual priors. This paper proposes Mix Forcing, a causal pre-training method using Mixture-of-Bidirectional-and-Autoregressive (MoBA) attention that jointly trains teacher forcing and bidirectional objectives on a 14B image-to-video model. After distillation into a four-step autoregressive generator, the authors report state-of-the-art visual quality and precise camera control on internal and public benchmarks, and quality preserved over one-hour rollouts.
In brief"Words Are the Voice of the Mind": Estimating Depression Severity from Social Media Posts via Evidence Reasoning
SummaryThis paper proposes UMer, a framework that estimates depression severity from users' social media posts without requiring them to complete questionnaires. It first builds a user-specific mental profile from historical posts, then retrieves posts relevant to each scale item as evidence. Finally, it reasons over this evidence to predict item-level scores with explanations. The authors report competitive results on real-world benchmarks and generalization to other scale-based tasks.
In briefWhen the Weakest Language Keeps Moving: ARMOR-DP for Multilingual BabyLM
SummaryMultilingual BabyLM systems often fix language sampling ratios, yet the weakest language can shift during training. This paper presents ARMOR-DP, a modular framework that diagnoses imbalance from representation, exposure, capacity, and objective pressure, and updates sampling using per-language validation loss. Experiments show equal character budgets severely harm Chinese representation, which byte-level encoding largely mitigates, and the dynamic controller achieves a better score–balance tradeoff than fixed protection.
In briefPatchwork: sharing unchanged simulation resources across randomized robot environments
SummaryThe paper presents Patchwork, a system that represents each simulation environment as a shared template plus a patch of changes, so unchanged resources are reused across many randomized instances. It is implemented on robosuite and MuJoCo and evaluated on LIBERO-plus and RoboCasa365 workloads. With 32 replicas, shared-template execution reduces CPU RSS by 88-92% compared with one process per environment.
In briefUltraFlash: Speeding Up Megapixel Image Generation
SummaryThe paper addresses a bottleneck in cascade refinement for megapixel image synthesis with latent diffusion models: repeated VAE decoding and re-encoding during pixel-space transitions. It proposes UltraFlash, which includes Latent Hyperloop, a learned latent-space shortcut replacing RGB round trips, and Sparse Patch-based Decoding, which reuses cached GroupNorm statistics to reduce patch seams with smaller overlap. Integrated into a state-of-the-art cascade baseline, the authors report about 2x lower 4K generation latency with improved image quality.
In briefZoomSEM: Faster Large-Scale Brain Connectomics Imaging Using AI Super-Resolution for Electron Microscopy
SummaryThis paper presents ZoomSEM, a deep learning method that enhances scanning electron microscopy images to higher resolution. Trained on real low- and high-resolution SEM image pairs, it recovers 4x, 8x, and up to 16x resolution from fast low-resolution scans while preserving fine neuronal structures. The authors report up to 30-fold effective acceleration and robust generalization across diverse biological samples.
In briefVariational Atlas Autoencoders: a paper on learning multiple local latent charts for data manifolds without labels
SummaryThe paper introduces Variational Atlas Autoencoders, a mixture-of-charts model in which local latent representations each cover one region of a data manifold. Chart assignment is treated as posterior inference and is learned in a self-supervised way without labels. Compared with capacity-matched baselines on image datasets including CelebA, the authors report better-structured latent spaces and higher sample quality, with sampling requiring only a single chart's decoder.
Arno Coomans, Merlin Nimier-David, Eugene d'Eon et al.
In briefCADonatello: Rebuilding Multi-Part 3D CAD Models from Real-World Scans
SummaryThis paper presents CADonatello, a framework that reconstructs parametric CAD models from real-world 3D scans of objects with multiple parts. It first splits the scan into components using a foundation-scale 3D part segmentation model, then rebuilds each part by parameterizing cross-section sketches and applying an extrusion. The authors also introduce CC3D Cross-Sections, a dataset pairing scan-derived cross-section images with CAD-derived sketches to support domain adaptation. Experiments reported in the paper show better generalization to noisy scans and reliable reconstruction of complex multi-part objects.”
Ahmet Serdar Karadeniz, Dimitrios Mallis, Parsa Akbari Chianeh et al.
In briefTrajectory2Gym: Building Learnable Tool-Use Interactions Grounded in Agent-Visible Decisions
SummaryThe paper argues that successful tool-use trajectories are not always useful training data, since they may miss multi-step dependencies or include actions unsupported by what the agent can see. It introduces Trajectory2Gym, which builds executable interactions around under-covered compositional structures and supervises only decisions grounded in the agent's visible history. Across five Qwen3 scales, it improves Tau1, Tau2, and BFCL at every scale, raising the equal-weight three-benchmark mean by 16.9 to 29.8 points. The largest gains appear in stateful, multi-turn interactions.
In briefDistilling Dense Trajectory Teachers for Large-Motion Video World Models
SummaryThis paper addresses video world models that generate videos following camera paths; existing models trained on small inter-frame camera motion degrade on sparse, large-step trajectories, producing frozen frames and geometric drift. The proposed Dense Trajectory Teacher Distillation (DTTD) post-training framework interpolates camera poses to form dense trajectories, uses the frozen base model to produce teacher videos along them, and evaluates candidate videos from the sparse trajectory against pose-aligned teacher observations, so the model acts as its own teacher. Geometry-aware rewards for cross-frame consistency and trajectory following are combined with this supervision for offline preference optimization, and a cumulative curriculum gradually adds larger motion steps.
In briefWorldFailBench: Testing Whether World Models Capture How Robot Policies Fail
SummaryRobot policies are often evaluated in generated environments using world models, but existing benchmarks only check whether tasks succeed or fail. This paper introduces WorldFailBench, which tests whether world models reproduce failure reasons, failure stages, and the scene conditions under which policies shift from success to failure. It covers manipulation tasks in the LIBERO and RoboTwin simulators and real-world tasks on the G1 humanoid robot. Experiments on seven world models reveal substantial gaps in how well they understand policy failures.
In briefOODConceptBench: A Benchmark for Open-World Concept Recognition in Vision-Language Models
SummaryThis paper introduces OODConceptBench, a hierarchical benchmark for testing how well vision-language models recognize open-world concepts from image-concept pairs. It covers 560 concept phrases across seven categories and evaluates 37 models. The authors find that recognition is lower on the harder L2 subset for all models, and that newer models do not necessarily perform better.
In briefHybrid-time PINNs: neural networks that learn systems with sudden jumps
SummaryThis paper proposes hybrid-time PINNs (HT-PINNs), which assign a separate network to each continuous flow interval of a hybrid dynamical system and train them jointly with jump times and unknown parameters. Standard PINNs are smooth in time and cannot model jumps, while HT-PINNs find the number of jumps and mode sequence by enumeration. The method recovers the jump count and restitution coefficient of a bouncing ball. In a MuJoCo robotic disassembly simulation, it estimates static and kinetic friction coefficients with errors of 0.1% and 1.7%.
In briefJointly training weighted BP decoder ensembles end to end
SummaryThis paper learns diversity among weighted belief propagation decoders by jointly optimizing trainable Tanner-graph edge weights, rather than relying on code-specific structure. Four ensemble training objectives are compared, with best-member training found most effective. For a (64,32) PEG LDPC code, a 10-member ensemble gains about 0.66 dB over a single BP decoder at a target FER of 10^-4, with similar behavior on a (128,64) CCSDS LDPC code.
In briefGeoLift: Grounding video spatial reasoning in explicit 3D maps
SummaryThe paper finds that multimodal large language models recognize objects in single frames but become much less accurate on route planning as the number of required turns grows. It introduces GeoLift, a training-free framework that uses feed-forward 3D reconstruction to place per-frame detected objects into a shared metric 3D scene, linking the same object across views via shared geometric regions called superpoints. With all weights frozen, Qwen3-VL-4B's VSI-Bench average rises from 55.4% to 64.1%, route-planning accuracy from 30.9% to 53.6%, and MindCube accuracy reaches 59.8%.
In briefEstimating aerosol particle size distributions from PM2.5 using physics-regularized flow matching
SummaryPM2.5 mass concentration is cheap and widely monitored, while particle size distributions need costly instruments found at few stations. The paper uses generative deep learning to produce ensembles of hourly size distributions from a 24-hour window of PM2.5 and eight meteorological parameters. Tested at two Finnish stations, flow matching performed best overall, and adding a physics constraint improved scores, though transfer between sites remained harder than local estimation.
Zhi-Song Liu, Jenni K\"oykk\"a, Wenqing Peng et al.
In briefEnd-to-end prostate lesion segmentation and detection from undersampled diffusion MRI
SummaryThis study proposes VarNet-Seg, an end-to-end network that uses complex multicoil diffusion-weighted MRI data and coil sensitivity information to segment prostate lesions. It uses 269 fastMRI prostate examinations and compares the model with conventional reconstruction-based approaches under simulated undersampling. At eight-fold acceleration, VarNet-Seg achieved a higher Dice score than the comparators, while patient-level AUROC was similar to the RSS baseline. The authors state that prospective evaluation is needed before clinical use.
Lukas T. Rotkopf, Marco Schlimbach, Kevin Kr\"{o}ninger et al.
In briefCross-scale sparse attention for segmenting remote sensing images from text descriptions
SummaryThis paper addresses referring remote sensing image segmentation, where regions in satellite or aerial images are identified from natural-language descriptions. It proposes CSSAM, a model with a cross-scale fusion module and a sparse attention module that align multiscale visual features with complex referring expressions. The authors also introduce M³RRSIS-D, a dataset built around multi-target, multi-hop, and multi-granularity expressions. Experiments on this dataset and two public benchmarks report competitive results.
In briefTracking concrete crack growth across viewpoints and over time
SummaryThis paper frames crack growth monitoring as a cross-view and cross-time crack-instance tracking task and proposes a geometry-assisted, morphology-aware framework. It aligns images from different viewpoints into a shared reference, assigns each crack a physically interpretable identity, and separates newly developed growth from existing segments. A temporally annotated dataset covering single crack propagation, mixed-mode branching, and multiple-crack interaction was created for evaluation. Results show loose cross-view identity correspondence under changing viewpoints, while strict spatial overlap and growth localization remain limited by early micro-crack segmentability and projection drift.
In briefLearning When Agents Should Collaborate: Selective Multi-Agent Medical Reasoning via Uncertainty-Aware Routing
SummaryThis paper proposes SMART-Med, a fully local multi-agent method for medical visual question answering that treats collaboration as a learned decision rather than a default. An uncertainty-driven router decides whether a single agent answers directly or a small set of expert agents is recruited. The authors build the MedVR-3K dataset with reasoning traces and difficulty levels, and fine-tune local agents with GRPO. On four benchmarks, SMART-Med is competitive with existing multi-agent approaches while cutting inference time by more than 3.8x and token usage by 72%.
In briefSEDATE: Whether LLMs share one aggression circuit across languages
SummaryThis paper proposes SEDATE, a training-free method that makes large language model responses safer using a detect-edit-verify process. It first locates neurons responsible for aggressive content, edits them when needed, and then checks the model's responses. The authors evaluate it on 7 languages and 4 model families, and further test it on 4 out-of-distribution languages and 4 additional datasets. They report substantially higher safety scores than the best baseline, with general capabilities retained and negligible overhead.
In briefPatch-Derived or Register Tokens? Comparing Variable-Length Image Tokenizers
SummaryThis paper compares two variable-length image tokenizer strategies: efficient patch-derived tokens with explicit spatial organization, and register tokens with no predefined spatial layout. Patch-derived tokens reconstruct better under extreme compression, while register tokens do slightly better at intermediate budgets and perform better on generative tasks. Adding representation alignment with pretrained visual encoders substantially improves patch-derived tokens' generative performance.
Philipp Schenkel, Tobias Sch\"urmann, Omer Sahin Tas
In briefCoEvolve: Counterfactual-guided co-evolution of memory and teamwork in AI agent systems
SummaryThis paper proposes CoEvolve, a framework that lets memory and collaboration in LLM-based multi-agent systems evolve together rather than separately. It represents reusable local collaboration patterns as skills and compares original with counterfactual trajectories to detect evolution signals that coordinate updates to skills and memories. Experiments on the ALFWorld and WebShop benchmarks report the best overall performance among the evaluated baselines, and ablation studies support the value of aligned co-evolution and counterfactual-guided evolution.
In briefFraudSociety: An LLM-Agent Simulator of the Financial Fraud Economy
SummaryThis paper introduces FraudSociety, an LLM-agent-based simulator that jointly models information flow and financial dynamics in fraud. It combines a dynamic social environment, persistent financial ledgers, and a Need-Intent-Action architecture for civilian decisions, with fraudsters adapting to interaction histories and ledger states. The simulator is validated against macro-level outcomes, micro-level loss concentration, and intervention effects. It is presented as a testbed for comparing anti-fraud strategies.
In briefClustering High-Dimensional Human Gene Score Data to Classify Target Cells
SummaryThis study examines whether the full human gene-score dataset and a selected kinase-gene subset yield the same number of clusters using 769 features. Two missing-data strategies are tested, followed by PCA and four clustering methods: K-Means, hierarchical, fuzzy C-means, and graph-based. Across all settings the optimal number of clusters is two, with relatively high Silhouette scores.
In briefSkillShift-Bench: A Benchmark for Web Agent Skill Learning Under Changing Environments
SummaryWeb agents can learn skills from past experience, but these skills are usually tested only in closed environments. This paper introduces SkillShift-Bench, the first benchmark that evaluates skill learning under controlled open-environment shifts across three self-hosted web systems. Experiments show that agents become less robust under these shifts, with lower success rates, higher interaction costs, and more skill failures, and execution shift has the strongest effect. The authors also propose a test-time adaptation method that improves learned-skill execution without relearning the skill library.
In briefUsing implicit beliefs as scalable process rewards to train strategic persuasion in language models
SummaryThe paper addresses two obstacles in training LLMs for multi-turn strategic persuasion: realistic simulation of psychologically grounded persuadees and dense, faithful credit-assignment supervision. It introduces PersuArena, a simulation environment producing non-monotonic dialogue trajectories, and Probe-RPO, an online RL framework that reads three introspective probes from a frozen persuadee's logits to form multi-dimensional process rewards. Across healthcare, financial advising, workplace, education, and personal relationship scenarios, the authors report a full acceptance rate of 52.5% for Probe-RPO, versus 35.8% for the strongest GRPO ablation and 25.8% for the best off-the-shelf LLM.
In briefOptimizing physical neural networks by selecting device input variables
SummaryPhysical neural networks use device physics in electronics and optics to build energy-efficient neural networks. This paper treats the routing between input signals and device input variables as trainable parameters, optimized with Gumbel-Softmax reparametrization. Using experimentally calibrated differentiable models of multivariate electronic RNPUs, MNIST accuracy rose from 94.5% to 96.9% and CIFAR-10 accuracy from 62.8% to 68.0%.
Sjoerd van den Belt, Manuel Escudero, Mohamadreza Zolfagharinejad et al.
In briefCross-architecture distillation for a lightweight medical ViT (LiteMed-ViT): gains and limits
SummaryThe paper introduces LiteMed-ViT, a dual-token vision transformer with a convolutional stem and class-balanced distillation from a ResNet18 teacher, using about 0.234M parameters. On MedMNIST subsets it outperforms some compact CNN baselines on BloodMNIST and PathMNIST and matches or exceeds their robustness under corruptions. On DermaMNIST, however, its balanced accuracy trails the strongest compact CNNs, and higher input resolution or stronger teachers do not help.
Shuaibu Shehu (SR University), Elijah M. Flomo (SR University), Mohammad Safiya (GITAM University) et al.
In briefRotation-trained component pooling for few-shot aerial scene classification
SummaryThis paper addresses few-shot remote sensing scene classification, where models must recognize new scene types from only a few labeled images. It proposes RoCoP, a parameter-free network that decomposes local image features into components, pools them using activation-based soft masks, and trains the backbone on randomly rotated views. Experiments on NWPU-RESISC45, AID and UC Merced report gains over a transfer-learning baseline. The authors note the method does not apply when the base sets are very small.
In briefOne Model, Two Paths: Balancing Single-Pass and Looped Inference in Language Models
SummaryThis paper studies how one language model can support both single-pass and looped inference by reusing middle-layer weights for extra computation. Because both paths share weights, training to improve looped predictions can degrade single-pass quality. The authors propose joint training with weighted losses on both outputs, guided by the frozen original model. Compared with looped-only training, joint supervision substantially reduces single-pass degradation while preserving looped gains.
In briefAffine-invariant Cubic Newton with Weak Learners: a refined optimization method for gradient boosting
SummaryModern gradient boosting relies on Newton's method, which works well in practice but lacks global convergence guarantees. The paper generalizes the Affine-Invariant Cubic Newton (AICN) scheme to convex optimization with weak learners, analyzing it using average cosine values along the optimization path and Hessian-norm directional improvements. Under semi-strong self-concordance, it establishes a global convergence rate that the authors state significantly improves on the prior weak-GRN bound. It also proves local and global linear convergence results and proposes weak-AdaAICN, which adaptively tunes cubic regularization.
Nikita Zozoulenko, Daniel Falkowski, Thomas Cass et al.
In briefCollective Collaboration of Recursive Autonomous Agents: a framework for coordinating smart factory production across multiple levels
SummaryThe paper proposes a framework for recursive autonomous agents (RAAs) in smart factories. Production objectives are decomposed top-down into tasks, while unit capabilities are aggregated bottom-up through peer collaboration. A production network covering machining, inspection, and logistics across multiple workshops illustrates how a common organizational kernel is instantiated at equipment-service, workshop, and factory scales.
In briefOverlap-aware sub-hypergraph pattern frequencies for classifying whole hypergraphs
SummaryThe paper extends frequent sub-hypergraph mining with overlap labels that record binned intersection sizes between pattern hyperedges, and represents each hypergraph by the relative frequency of each mined pattern. The vocabulary is mined only from training hypergraphs and then frozen. On DHG-Bench's six datasets, the method reports higher mean Macro-F1 than the best published mean on five datasets, and lower on RHG-3. Ablations show that replacing relative frequencies with presence, or removing overlap labels, reduces Macro-F1.
Nafiul Hasan Anik, Nazira Jesmin Lina, Md Tanvir Alam et al.
In briefEat Your Precision: A paper on reducing reward hacking in rubric-based reinforcement learning by also rewarding precision
SummaryRubric-based rewards let models learn from tasks without a clear right answer, but they reward only how much of a rubric is covered, which makes responses longer and less accurate. The authors propose the FORK reward, which jointly optimizes recall and a precision objective requiring each response unit to be necessary and factually correct. Experiments report that FORK outperforms existing rubric-based methods and substantially raises win rates against the base model on RubricHub-Medical and HealthBench.
In briefA Learnable LLM Retriever for Textual Graph Question Answering
SummaryThis paper proposes TaLenT, an end-to-end framework for textual graph question answering that pairs an LLM-based retriever with a fine-tuned LLM generator. The retriever is trained with gold answer signals, using a weighted pseudo-label loss that prioritizes gold answers, keeps answer-related nodes for coverage, and discards irrelevant noise. Experiments on three GraphQA datasets show consistent improvements in downstream performance.
In briefFreqPKC: Frequency-Aware Peak Convolution for Radar Image Segmentation
SummaryThis paper proposes FreqPKC, which analyzes the frequency response of the PeakConv operator used in radar semantic segmentation and lets the network adaptively adjust its effective bandwidth. Built on it, FreqPKC-Net combines frequency-guided adaptive modulation with cross-view fusion for multi-view inputs and can also handle single-view data. Experiments on CARRADA, CARRADA-RAC, and the self-built RS-SAR dataset show it consistently outperforms existing state-of-the-art methods on both range-Doppler and range-angle views.
In briefIs It Novel? PatEx-Gym: Agentic Prior-Art Search for Patent Novelty Assessment
SummaryThe paper presents PatEx-Gym, a framework for developing and evaluating agents that perform patent examination, including 1K examination tasks, retrieval corpora of over 1M patents and 52K scientific papers, and a runtime supporting iterative search and Office Action generation. It also introduces PASS, an annotation-free method for synthesizing prior-art search trajectories, which produces PatEx-SynData used to fine-tune a 4B-parameter PatEx-Agent. The authors report that final decision accuracy alone does not fully capture examination quality, and that the compact 4B model performs competitively against substantially larger general-purpose agents.
In briefDualPace: Decoupled scheduling and GPU-CPU cooperative decoding for LLM serving
SummaryThis paper presents DualPace, a runtime for serving large language models that treats the host CPU as an independently scheduled decoding engine. The GPU handles prefills and GPU-resident decodes, while an AMX-enabled CPU engine batches and advances CPU-resident decodes, with the two engines decoupled in scheduling but cooperating in execution. According to the paper, in decode-oriented scenarios it improves throughput by 6.9–43.2% over NEO and 11.5–90.6% over vLLM while keeping similar latency.
In briefFedSEAM: federated learning of language models that learn new tasks without forgetting old ones
SummaryFedSEAM is a framework for training a language model on a sequence of tasks across several clients without sharing their data, while keeping knowledge of earlier tasks. It organizes task knowledge into sparse experts, uses a subspace-based compatibility test to decide which expert parts to reuse, and merges client updates on the server only along agreed directions. The authors evaluate it on the TRACE benchmark with Qwen3-4B and Llama-3.2-3B.
Fatema Siddika, Abhijit Chunduru, Zilinghan Li et al.
In briefFive Models, One Shared Space: Aligning Language Model Representations Without Merging Weights
SummaryThe paper examines whether five independently trained, frozen language models with different architectures (four transformers and one state-space model) can be aligned using only their activations, without merging weights. The authors find statistically robust correspondence in how the models represent the same text, and build a shared space with Generalized Procrustes Analysis that outperforms pairwise linear maps at the final layer. A router that selects among the five models did not beat always using the strongest model, which the authors report as a negative result.
WhyAligns five heterogeneous models via activations alone without weight merging, validated by permutation tests and ablations; routing results are negative, limiting impact.
Oladimeji Anthonio, Dimeji Olawuyi, Oloruntoba Ajayi et al.
In briefRBGA: Residual-Budget Guide Arbitration for Safe Offline-to-Online RL in Dexterous Manipulation
SummaryThis paper addresses safe offline-to-online reinforcement learning for dexterous robot manipulation. It proposes RBGA, which tracks the remaining episode safety budget and uses it to choose step by step between offline guide actions and the evolving online policy. Experiments reported in the paper indicate that RBGA supports online policy improvement while maintaining strong safety performance.
In briefSmall Typos, Big Shifts: Testing How Well LLMs Handle French Input Variations
SummaryThis paper builds a controlled French benchmark of 12,600 input instances across seven task types, testing common input variations such as omitted apostrophes, missing diacritics, and keyboard typos. The authors evaluate six models from two LLM families and find that accuracy drops and probability-quality metrics (Brier score and NLL) worsen as perturbation intensity increases. Keyboard typos cause the largest losses, and combining calibration with input restoration improves prediction reliability beyond either method alone.
In briefExplicit Friction Contact Modeling for Neural Garment Simulation
SummaryGraph neural networks simulate garments efficiently, but body-garment friction is not set explicitly at inference, and retraining with a friction coefficient barely changes the response. The paper proposes a Modular Contact Operator that attaches to a frozen predictor, applying a closed-form Coulomb friction correction without retraining or iterative solving. On HOOD and ContourCraft, sticking rises and slip falls monotonically with the friction coefficient, matching a reference solver at about one-eighth of its runtime.
In briefWhether LLM tutoring interactions add predictive value for student grades
SummaryThis paper tests whether semester-level features from a large language model tutor improve assignment-score prediction over constant baselines. It uses 16,851 student-assistant exchanges from two offerings of an undergraduate AI course, matched to 175 student-semester records. Gradient Boosting's mean absolute error (3.023) is slightly lower than the median baseline's (3.138), but the confidence interval includes zero. The study does not establish a learning effect.
Zeyu Dong, Mikael Kwame Kuada, Asa Holley Buchakjian-Tweedy et al.
In briefReSight: A Self-Improving Memory System for Robots That Grows Tools as Tasks Require
SummaryReSight is a memory system for embodied agents such as robots that stores keyframes and their spatiotemporal links, so past observations can be reprocessed for new needs. An execution agent and a coding agent share this memory, and the Explore-Distill-Mount (EDM) method turns exploration results into reusable memory builders, query tools, and skills. Experiments on three public benchmarks report improved position error and success rates, and tests on real robot patrol memories show the added capabilities can be reused.</summary>
In briefCan One Hear Geometry? Invariant Spectral Descriptors for Text-to-B-rep Retrieval
SummaryThis paper presents a framework for retrieving 3D solid models (B-rep) from natural language text, using group actions and spectral geometric invariants so that descriptors stay unchanged under translation, rotation, reflection, uniform scaling, and contour recording conventions. For sketch-extrusion solids, it builds a fixed 280-dimensional descriptor combining cross-sectional regions, contour boundaries, and extrusion structure. On text-to-3D solid retrieval, the method reaches 88.6% Recall@5, 13.6 percentage points above the previous state of the art.
In briefError-Subspace Calibration for Improving Merged Multi-Task Models
SummaryThis paper proposes ESCal, a calibration method that improves models created by merging several task-specific expert models. It identifies low-dimensional error subspaces from weight differences between the experts and the merged model, then learns linear bridge maps at block boundaries to correct features. Experiments on vision and natural language understanding benchmarks report improved merged-model performance using a small calibration set per task.
In briefClimate downscaling with a likelihood-based functional flow-matching model
SummaryThis paper introduces CD-FFM, a functional flow-matching model that treats high-resolution climate fields as continuous random fields queryable at any location and resolution. Applied to downscaling ERA5 daily temperature to 1 km over Fennoscandia, it reports lower continuous ranked probability scores than regression and a larger neural operator across three test years. The authors also state the model transfers to precipitation and makes large kilometre-scale ensembles more affordable.
Zhi-Song Liu, Helmi Toropainen, Ammar Kheder et al.
In briefEnergy-efficient lung tumor segmentation using a modified CBAM-ENet skip model
SummaryThis paper proposes a lightweight lung tumor segmentation model based on ENet, with UNet-style skip connections and CBAM attention modules added. It was trained on TCIA data from the Medical Segmentation Decathlon Challenge and evaluated on LIDC-IDRI, reaching a DSC of 0.90812 with 430,052 parameters and 0.549 GMAC. The authors report it is up to 10 times faster, requires 79 times less GMAC than UNet, and has 22 times fewer parameters than SegNet.
Sharan Thummagunti, Aayushi Padia, Abhiram Koppuravuri et al.
In briefMarineLife-16K: A Motion-Aware Marine Video-Text Dataset with Rich Attributes
SummaryThis paper introduces MarineLife-16K, a video-text dataset of 2,000 clips annotated with 13 attributes, including object properties, behaviors, environment, and camera motion. The clips span varied lighting, times of day, and seven countries. The authors generate 16,080 video-question pairs from these annotations and benchmark proprietary and open-source video-LLMs on question answering and captioning. Results show proprietary models often perform better, VLMs handle object motion well but struggle with complex camera motion and static cameras.
Quang-Trung Truong, Thanh-Nghia Nguyen, Phu-Phong Vo et al.
In briefA Unified Analysis of Importance-Corrected Momentum Actor-Critic on Markovian Trajectories
SummaryThe paper presents a generic actor-critic method combining a ridge-regularized LSTD critic with importance-corrected momentum actors. The critic needs no stepsize, mixing-time knowledge, or rank condition. The authors give what they describe as the first finite-time bounds for such methods along a single Markovian trajectory, with normalized and unnormalized steps. They specialize the bounds to Muon, normalized SGD with momentum, SGD with momentum, and AdaNorm, and compare them in the nuclear norm.:
In briefLLMs in the Cloud: A Cross-Layer Review of Architecture, Orchestration, and Inference
SummaryThis paper is a systematic review, following the PRISMA 2020 framework, of the physical and system bottlenecks in serving large language models in cloud datacenters. It unifies research on hardware virtualization, cluster scheduling, and runtime inference optimization, and examines cloud-native designs such as serverless deployment, fault-tolerant checkpointing, and trusted execution environments. It also analyzes trade-offs among throughput, latency, cost, and confidentiality and outlines future research directions..
Jannatun Noor, Athar Noor Mohammad Rafee, Md. Sadman Haque et al.
In briefQwen-Image-PE: Teaching AI to Add Visual Design to Image Prompts
SummaryThis paper introduces Qwen-Image-PE, a prompt enhancement model that turns vague user requests into detailed visual descriptions for text-to-image generation. It uses two training stages: supervised fine-tuning with Caption-Based Reverse Distillation, and reinforcement learning with a Groupwise Design Reward that refines lighting, composition, layout, and detail. On Qwen-Image-Bench, the 397B model improves over original requests by 7.81 points, and the 9B model outperforms all evaluated baselines.
In brief"The Future Is Finer": A paper on forecasting fine-grained futures from coarse archived time-series data
SummaryThis paper addresses multivariate time series on web platforms that are often archived at coarse granularity, introducing Multivariate Time Series Super-Resolution Forecasting (MTS-SRF) to predict fine-grained futures from coarse histories. The authors propose TimeMagnifier, which combines Fourier Anchor, De-noise Only, and banded DTW Alignment components. On a ten-dataset benchmark, it outperforms 15 baselines on every dataset and metric, reducing MSE by 12.3% on average.
In briefLegoSR: A unified framework for self-improving agent-guided scientific discovery
SummaryEquation discovery solvers differ in data interfaces, equation representations and evaluation criteria, which makes their results hard to compare. LegoSR is a unified framework for algebraic symbolic regression and differential equation discovery that integrates seven solvers through a common representation and evaluation process, and supports searches constrained by partially known equations. Its agent, LegoSR Agent, selects solvers and adapts searches using solver knowledge, intermediate results and lessons from earlier tasks. On noisy ordinary and partial differential equation benchmarks, it recovers nearly as many equations as the union of seven expert-tuned solvers.
