In briefKGATE: A Modular Training Environment for Knowledge Graph Embedding Models
SummaryKnowledge graph embedding models map the entities and relations of a knowledge graph into a low-dimensional latent space for tasks such as classification or link prediction. The paper presents KGATE, a modular Python library built on PyTorch Geometric and TorchKGE, where users assemble encoders, decoders, losses, negative samplers, and metrics as building blocks. It includes a preprocessing procedure that controls data leakage and a built-in training pipeline. Benchmarks against six existing KGE libraries show training time comparable to the fastest ones while offering a broader feature set.
Benjamin Loire, Galadriel Bri\`ere, C\'elia Brahimi et al.
In briefVERDICT: A Plug-in Framework Adding Traceable Evidence and Calibrated Confidence to Document Field Extraction
SummaryThe paper presents VERDICT, a plug-in framework that attaches to an existing schema-guided field extractor without modifying it. A small vision-language model localizes the evidence for each extracted value once per document and calibrates both value and evidence confidence. With fine-tuning, the framework gives usable precision-coverage trade-offs at a strict 0.80 threshold on in-distribution and out-of-distribution datasets.
Harikrishnan P M, Goutham Vignesh, Siva Prasad et al.
In briefEntity-Aware OCR Evaluation for Historical Document Transcription
SummaryThis paper examines OCR errors at the entity level, such as names, dates, and identifiers, as a complement to standard metrics like Character Error Rate. The authors use 175 documents from the Berrutti Archive with 5,209 annotated entity mentions to compare sequence-recognition OCR systems, OCR-specialized vision-language models, and general-purpose vision-language models. They find that system rankings depend on the evaluation criterion: Qwen 3.6 achieves the highest entity coverage across tolerances, while some systems recover substantially when small character errors are allowed.
Camilo Mari\~no (UDELAR, CB), Diego Belzarena (UDELAR et al.
In briefInverting Multi-Vector Visual Document Indices: Recovering Pages From Stored Vectors
SummaryThis paper examines whether the index stored by multi-vector visual document retrievers can leak the original pages. The authors frame inversion as conditional document image generation and show on the ViDoRe v3 benchmark that pages rebuilt from raw indices recover 47% of words and rank their source page first 98.4% of the time. Two simple protections, token pooling and shuffling, cut word recall to about 8%, but a model that restores the order of a shuffled index raises first-place source ranking to 93.5%.
WhyProposes inverting multi-vector visual document indices to reconstruct pages, with substantial experiments showing strong recall and ranking, a notable security impact.
In briefHessian eigenvalues in deep learning: where symmetry comes from and how it breaks
SummaryThe paper studies Hessian spectra of trained deep learning models, which show a large cluster of near-zero eigenvalues and a few isolated outliers. The authors argue these patterns arise when the original setting is viewed as a departure from a nearby, hidden, highly symmetric reference configuration. That reference has Hessian invariances beyond weight symmetries, allowing precise spectral description. Returning to the original configuration breaks this symmetry, producing the observed hierarchy of clusters and outliers.
In briefNeuralZip: Reusable Setup for Fast Lossless Compression of Model Weights
SummaryThis paper introduces NeuralZip, a lossless compression method for model weights that groups chunks with similar exponent distributions, shares Huffman codes, and packs recurring exponent tuples. A setup step can be done once and transferred to other compatible architectures while still achieving similar compression. After setup, compression is 1.81 to 21.33 times faster than baselines and reconstructs weights bit-for-bit exactly. GPU experiments reduce active memory usage by up to 27.5% while reproducing logits exactly.
Mart\'in Bravo, Samuel Horv\'ath, Gonzalo Navarro et al.
In briefRollVerify: A reinforcement learning method that balances speed and accuracy for long rollouts
SummaryThis paper addresses two problems in reinforcement learning for large language models: long-tail rollouts that leave GPUs idle, and stale off-policy samples introduced by asynchronous or partial-rollout training. The proposed framework, RollVerify, uses an off-policy shift metric (OPS) to measure deviation in partial trajectories and verifies samples at sequence and token levels before training, truncating invalid suffixes. Experiments on mathematical and tool-assisted mathematical reasoning report accuracy comparable to on-policy training with lower training cost, with preliminary code-generation results also included.
In briefFedSSMCoOp: Lightweight Federated Prompt Learning with SSM Encoders for Few-shot Image Classification
SummaryThis paper proposes FedSSMCoOp, a federated few-shot image classification framework that preserves data privacy while enabling multimodal learning. It uses SSM-based Vision Mamba and Cross Mamba blocks and optimizes only soft-prompt and communication-prompt updates, so no external large language model is needed for feature alignment. Evaluated on biomedical image datasets, the framework reports stable performance relative to baselines and is on average 1.96 times lighter.
Ankita Das, Ambarish Parthasarathy, Sumohana S. Channappayya et al.
In briefSafety-first ranking of candidate driving paths using BEV perception outputs
SummaryThe paper presents a modular pipeline that takes BEVFusion object detections and map segmentations from nuScenes data and ranks manually specified candidate trajectories by safety criteria. Rankings prioritize avoiding collisions and walkway violations, then endpoint distance to the target, then path efficiency. In a study with seven BEV traffic scenarios and 21 survey participants, 124 of 146 pairwise votes matched the top-ranked trajectory, an agreement of 84.9%. The authors conclude that simple, transparent BEV-based safety heuristics can approximate human preferences in constrained trajectory choice tasks.
Leon Pohl, Martin Frauendorfer, Corinna Schmitt et al.
In briefLearning joint weather forecasts from station data alone, without weather models
SummaryThis paper introduces CLARA, a model that learns joint Gaussian predictive distributions of five surface weather variables using only station observations, without numerical weather prediction or reanalysis data. The model has about 28,000 parameters and can be trained and run on a CPU. Across 96 stations and six multi-year folds, it achieves lower energy scores than a learned comparator and statistical baselines, and learned correlations between variables improve joint likelihood.
In briefNL2Hull: A Language-Driven Constrained Framework for Ship Hull Design Decisions
SummaryThis paper presents NL2Hull, a framework that treats ship-form editing as a typed discrete decision problem and links language requests to hull geometry through a constrained free-form deformation engine based on NURBS. The authors build a 134,558-record dataset and a 5,000-record benchmark with 43,496 typed questions. Their model Chip reports 95.90% question accuracy and 99.32% FFD exact match.
In briefLove for Believable AI: Engineering Artificial Partiality and Relationship Persistence in Chatbots (paper, cs.HC/cs.CL)
SummaryThe paper studies a behavioral mechanism that lets conversational agents show per-user partiality, built as a persistence layer over one open-weight base model. Partiality combines caring, a finite allocation of interaction surplus above a guaranteed service floor, and particularity, a per-user state estimate built from relationship history. Evaluation uses constructed multi-session relationships, with no human-subjects data and no claims about attachment, trust, or machine interiority. The authors present the mechanism as a candidate for a later perception study, not as evidence that users perceive love.
In briefQCATS: Query-Aware Transformer Slicing for Faster Predictive Queries in Databases
SummaryThis paper proposes QCATS, a framework that makes Transformer model inference inside databases faster by using query-level information. Instead of routing individual tokens or tuples, it uses query predicates and metadata statistics to pre-select context-aligned feed-forward network slices before execution. Experiments on four predictive-query workloads with BERT-base and Qwen-0.6B report up to 4.42x latency reduction with prediction accuracy comparable to dense baselines.
In briefReview and experiment on human feedback reinforcement learning for safer human-robot teamwork
SummaryThis paper reviews 20 peer-reviewed studies (2020–2025) on reinforcement learning from human feedback in human-robot collaboration, selected via PRISMA guidelines. It also reports a virtual reality experiment comparing system-initiated and user-initiated feedback on robot proximity behaviour. The user-initiated feedback better captured perceived safety, indicating that feedback timing affects feedback quality.
Alexandra Coroiu, Andrea Vogt, Viktor Werbilo et al.
In briefOutperformance Inverse Optimization: Learning Objectives That Beat Observed Decisions
SummaryInverse optimization infers objective-function weights that make observed decisions appear optimal. This paper proposes outperformance inverse optimization, which seeks weights whose optimal solutions outperform the observed action in every component at each state. It introduces a loss function computable with forward-problem oracles alone, making it applicable to mixed-integer linear programs, along with gradient-based and DC optimization algorithms. Experiments on synthetic and real data show improved prediction of outperforming solutions over existing methods.
In briefA dissipative model of learning: exact parameter recovery under designed excitation, but degenerate on observational data
SummaryThe paper models human learning as a nonlinear dissipative ODE system with a transfer matrix, per-concept forgetting rates, and a saturating practice gain. For two concepts it proves a constructive structural identifiability theorem and provides an L-stable implicit scheme and a faster batched solver. Under the stated excitation conditions, synthetic recovery is exact, but on large observational benchmarks an apparently strong recovery fails four controls. The authors attribute this to degeneration consistent with the theory when designed excitation is absent.
In briefSparsifying Randomness, Not Capacity: Making Some Bayesian Network Weights Fixed
SummaryThis paper studies which parameters of a Bayesian neural network should be stochastic. It applies deep weight factorization to the prior scales, the standard deviations of parameter priors, and fits the functional prior to a Gaussian process using a maximum mean discrepancy objective. Parameters whose prior scale falls below a cutoff become deterministic. On UCI benchmarks, the method performs on par with a fully stochastic network while keeping about half its parameters deterministic.
In briefLayerwise error attribution for fast, robust mixed-precision quantization
SummaryThis paper proposes a mixed-precision post-training quantization method that assigns bit-widths layer by layer under a global memory budget, using a small calibration set. The authors derive a layerwise probabilistic analysis that separates propagated quantization error from each layer's local perturbation, and use this local term to build a separable score for a simple allocation algorithm that needs no external solver. On DRUNet denoising at an average of 4 bits per weight, the method matches or improves on baselines with clean calibration data and is more robust to corrupted calibration data, with bit allocation 28x to 2,570x faster than the studied baselines.
Samy Houache (IMB, UB), Yann Traonmilin (IMB et al.
In briefThinking before painting: recursive latent reasoning that helps diffusion models solve visual puzzles
SummaryThis paper proposes Painter-Thinker (PaTh), where a small recursive network reasons over learned tokens and guides a frozen diffusion model through ControlNet adapters. It is trained only with the standard reconstruction loss, without symbolic targets, a solver, or a verifier. The authors report improved results on Sudoku, mazes, Queens, and CLEVR scenes compared with prior methods, using about 10M parameters.
WhyIntroduces recursive latent reasoning into pixel-space diffusion without symbolic supervision, substantially beating prior methods on Sudoku with fewer parameters and showing growing advantage.
Pawe{\l} Skier\'s, Ma{\l}gorzata Grzanka, Wojciech Masarczyk et al.
In briefSANet: A selective attention network for detecting faint infrared targets
SummaryThis paper proposes SANet, a network for detecting small, dim infrared targets against complex backgrounds while limiting false alarms. It combines a dual-path module using standard and pinwheel-shaped convolutions with a selective attention fusion module that adaptively merges multi-scale features. On three public benchmarks, its IoU exceeds the second-best method by 1.93, 4.32, and 2.21 percentage points.
In briefApplying binary neural networks to fast event-camera processing
SummaryThis paper adapts and evaluates modern deep binary neural networks (BNNs), which compress weights and activations to one bit, on event-camera data. It introduces the Polar-wise Binary Event Volume (PBEV), a binary representation that lets BNNs process event data directly. The authors also show that cross-modal pretraining from RGB data improves BNN classification accuracy on neuromorphic datasets. The best BNN reaches 90.58% accuracy on N-Caltech101 with 7.5 times fewer operations than full-precision counterparts.
In briefLiveMACE: Process-Aware Evaluation of LLM Agent Abilities in Changing Markets
SummaryThe paper introduces LiveMACEBench, a benchmark that uses live financial markets to test persistent LLM agents running continuously. Five frontier LLMs are evaluated under Tool Use, Persistent Memory, Rule Following, and Multi-Agent Collaboration setups, using both realized returns and diagnostics drawn from full decision traces. Over 30 days of live evaluation, the authors find that returns often diverge from capability-specific measures, and similar outcomes can stem from different mechanism use. They argue that outcomes alone cannot reliably measure agent capability.
In briefAI-assisted seismic processing and geological interpretation workflow for implicit geological modeling
SummaryThis paper presents an AI-assisted workflow for seismic interpretation to support implicit geological modeling, covering onshore data from shallow to deep depths of about 300 to 3500 m. It uses self-supervised and semi-supervised contrastive learning CNNs for noise reduction and interpolation, and interprets horizons and faults with minimal human-generated training data. As a first demonstration, the top of the Dutch Maassluis Formation was interpreted in the Leeuwarden and Waalwijk 3D seismic cubes, and the authors conclude such workflows are mature enough for integration into geological modeling and decision-making.
Stefan Carpentier, Jan Diederik van Wees, Eva de Boever et al.
In briefMulti-agent reinforcement learning scheduler for deadline-aware TSN in vehicular edge networks
SummaryThis paper proposes a multi-agent reinforcement learning approach for scheduling time-sensitive networking (TSN) queues in vehicular edge computing. Each TSN queue gets an agent that learns service order and time-slot duration to reduce deadline misses under speed-dependent latency requirements, using multi-agent proximal policy optimization (MAPPO). According to the paper, compared with centralized single-agent methods, it reduces service latency by up to 66.2% and improves reliability by up to 271.8%.
Bernardo A. C. Pereira, Marcos Carvalho, Fatih Temiz et al.
In briefMUNITE: A Unified Latent Inference Framework for Any-to-Any Multimodal Generation
SummaryMUNITE is a latent-variable framework for generating across any combination of modalities, treating encoding and latent generation as one inference problem under varying amounts of observed input. It uses a single conditional flow model and extends conditional flow matching with self-distillation so it can learn from incomplete training examples. Across PolyMNIST-D-Q, FFHQ64, and image-text-audio tasks, it reports competitive or better generation quality and source-target alignment, with the highest joint-generation coherence in the one-to-many and unconditional image-text-audio comparisons.
In briefAgent4RE: A Self-Refining Multi-agent System for End-to-End Requirements Engineering and Benchmarking
SummaryThis paper presents Agent4RE, a multi-agent system for software requirements engineering that coordinates specialized agents and uses two iterative improvement loops. The authors also build RE-E2E, a dataset from human-written requirement specifications, to evaluate the full workflow from elicitation to generation. Across eight LLMs, all three Agent4RE variants outperform a prompting baseline by an average of 8% on text-based metrics, and the two enhanced versions receive the highest LLM-as-a-judge and human ratings.
In briefGlobal Average Precision: a differentiable ranking loss for representation learning
SummaryThe paper introduces Global Average Precision (gAP), which ranks all query-candidate pairs in one list to compute a single AP, so that similarities are compared across queries. It proposes gSAP, a differentiable surrogate that takes the same inputs as existing losses and can replace them. Reported results show gains in supervised metric learning, cross-modal alignment, and self-supervised pretraining, including more positive pairs retrieved at the same precision.
WhyProposes a differentiable global-AP surrogate loss with cross-query comparability, validated across supervised, cross-modal, and self-supervised experiments.
Bill Psomas, Mohammad Mahdi, Michalis Thomas et al.
In briefSample efficiency of Kolmogorov-Arnold Networks for reinforcement learning, showing gains with fewer training samples
SummaryThis paper studies how efficiently Kolmogorov-Arnold Networks (KANs) learn from data in reinforcement learning. The authors run computational experiments on the Feynman dataset and the Gymnasium RL benchmark. They report similar performance with 40% fewer samples and relative improvements up to 50% during training. The gains remain robust under varying levels of reward noise.
In briefDeepTopoClustering: unsupervised grouping of 4D terrain-change data into surface process types
SummaryThe paper introduces DeepTopoClustering (DTC), an unsupervised framework that derives a hierarchical taxonomy of surface processes from 4D point clouds captured by permanent laser scanning. Each object-based surface change is converted into a GeoMorphogram, which a convolutional autoencoder embeds before hierarchical deep clustering. On sandy beach datasets, the taxonomy with eight major process types reaches F1 = 0.78 and match accuracy = 0.92 against expert annotations, outperforming dimensionality reduction and flat clustering.
Jiapan Wang, Daan Hulskemper, Mathilde Letard et al.
In briefTraining Advisors for LLM Agents from Task Outcomes
SummaryThe paper introduces Caddie, a method that trains a critic model to give LLM agents natural-language analysis and advice during multi-step tasks. Instead of step-level labels or reference critiques, the critic learns from whether the agent ultimately succeeds after receiving its feedback, optimized with reinforcement learning while the base model stays frozen. A Qwen3-4B critic trained on multi-hop question answering raises MuSiQue success rates by over 25 percentage points and also improves out-of-domain benchmarks such as τ³ and DeepDive without extra training.
Sergei Polezhaev, Barys Liskavets, Ori Press et al.
In briefSelf-evolving tool agents with a frozen historical anchor (AnchorLoop)
SummaryThis paper studies self-evolving tool-integrated agents that learn without external task or answer supervision. It proposes AnchorLoop, which adds a frozen copy of the previous iteration's Executor as a historical reference for both the Executor's advantage computation and the Curriculum Agent's agreement-based reward. Across 13 reasoning benchmarks, it reports gains of 2.5% on mathematical reasoning and 2.8% on general reasoning over Agent0.
In briefDeltaSplat: Iterative Refinement of Gaussians for Pose-Free 3D Scene Reconstruction
SummaryThis paper presents DeltaSplat, a lightweight module that refines 3D Gaussian Splatting reconstructions made from unposed images in a single network pass. It iteratively renders the current Gaussians and predicts per-Gaussian updates from the residuals, using Plücker rays and rendered depth as geometric guidance. On the DL3DV dataset, it reaches 26.64 dB PSNR in the pose-free setting, improving its backbone by 1.75 dB while adding about 2.2% parameters.
Chanung Park, Seunghyeon Song, Joo Chan Lee et al.
In briefHard yet reducible: a controlled method for choosing useful synthetic image degradations
SummaryThe paper proposes the controlled Reducible Degradation Gap (cRDG), a score estimating how much training on a synthetic degradation improves generalization under a fixed training budget. It compares two budget-matched probes from a common checkpoint that differ only in whether a synthetic degradation or a clean augmentation is used, measuring gain on held-out degraded images relative to the clean control. The resulting curation method, Curation of Reducible Bands, selects synthetic data without changing the predictor and is reported to improve semantic segmentation and salient object detection while preserving clean performance.
In briefData-efficient partial inverse design using active learning on automotive glass run channels
SummaryThe paper proposes CoNN-AL, which adds stream-based active learning to a Cooperative Neural Network with Denoising Autoencoder, using Monte Carlo dropout to estimate uncertainty and select which candidate designs to label. Tested on a dataset of over 900,000 simulated automotive glass run channel designs, it reaches R-squared of 0.967 to 0.982 with only 20,000 labels, about 2.3% of the pool. At harder missing-variable levels it needs 30 to 40% fewer labels than random sampling to reach R-squared of at least 0.95. The dataset is publicly released.
In briefRotation-consistent neural surrogates for 3D fluid dynamics using exact SO(3)-equivariant kernels
SummaryThe paper studies how neural surrogates for three-dimensional Navier–Stokes dynamics change their predictions when the same physical state is rotated, and proposes the Invariant-Conditioned Isotropic Kernel Neural Operator (IKNO). IKNO builds local interactions from rotation-invariant scalars and rotation-covariant vector directions, so predicted velocity changes rotate exactly with the input. It is consistent to numerical precision, matches the accuracy of a rotation-aware Tensor Field Network with 5.6 times fewer parameters, and outperforms a parameter-matched graph simulator.
In briefFaithful attribution method optimizing insertion and deletion curve area to rank feature importance
SummaryThis paper addresses attribution methods in explainable AI, which produce heat maps showing the relative importance of input features. It derives an objective from the faithfulness measure based on the area under insertion and deletion curves, linking insertion curves to top-k feature selection and approximating its gradient via randomization. The authors combine this loss with the neural explanation mask framework to create Ra-NEM, which works with any differentiable model. Experiments report accurate, robust and efficient attributions with higher faithfulness and good performance on other XAI metrics.
WhyTurns insertion/deletion faithfulness into a differentiable objective with randomized gradient approximation, integrated into NEM and validated against baselines on several metrics.
Bj{\o}rn Leth M{\o}ller, Bulat Ibragimov, Christian Igel
In briefMinimizing cumulative envy when goods arrive one by one
SummaryThis paper studies fair division of indivisible goods that arrive in sequence and must be allocated irrevocably, with valuations known in advance. It introduces cumulative maximum envy, the sum over rounds of the largest pairwise envy, which measures both how large and how long envy lasts. Minimizing this objective for a fixed arrival order is strongly NP-complete and has no constant-factor approximation unless P = NP. The authors also give pseudopolynomial, polynomial-time, and FPTAS results in special cases, and a greedy 3/2-approximation for two agents when the arrival order can be chosen.
In briefORCA: Finding Compositional Failures in Text-to-Image Diffusion Models
SummaryText-to-image diffusion models often fail on compositional prompts, such as binding attributes to the wrong objects or losing object counts. The paper argues the cause is misalignment between text and visual representations rather than missing information. It proposes ORCA, which adds an auxiliary loss that aligns a diffusion transformer's latents with a low-rank target from a frozen visual encoder. Across three backbones, ORCA improves FID and GenEval over vanilla and REPA baselines with no inference-time cost.
Arshia Hemmat, Amirhossein Vahidi, Amitis Shidani et al.
In briefFully Interpretable Minimal Transformers: Geometry to Algorithm
SummaryThis paper presents a framework for building and interpreting minimal transformers by limiting embedding and head dimensions to two, enabling full two-dimensional visualization of internal representations. The model is trained on a task that outputs the most recent even digit whenever a '+' appears in a digit sequence, and each computation step is visually traced. The authors argue that the learned geometry of points and decision boundaries can be read as a step-by-step algorithm.
In briefWhy universal machine-learning force fields err in multicomponent materials
SummaryThe paper builds a benchmark of 7,599 multicomponent atomic configurations and compares eleven pretrained universal machine-learning force fields against density functional theory for energies, forces and stresses. It also extends the assessment to elastic, vibrational and adsorption-related properties. Force errors are linked to differences in training-reference coverage and local geometric heterogeneity, and compression-side errors are 1.81 to 1.95 times extension-side errors. Spherical-harmonic representations gave lower test force errors for most elements, which informs force-field selection and training-data sampling.
In briefA Deafening Silence: Where Catastrophic Forgetting Concentrates in LLMs
SummaryThis paper studies where and why catastrophic forgetting occurs when large language models undergo continual pre-training or fine-tuning without replay data. Across experiments spanning 160M to 12B parameters and four model families, the authors find forgetting concentrates in the output embeddings of tokens rarely seen in the new corpus. They propose raising Adam's epsilon only for the output projection during training, which removes 39.4% to 67.9% of forgetting across seven stable configurations.
WhyShows forgetting localizes to output embeddings of corpus-rare tokens and proposes a one-line Adam epsilon fix, validated across 160M–12B parameters and four model families.
In briefSkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles
SummaryThe paper proposes SkillForge, an agentic reinforcement learning method that manages an LLM agent's skill library through trial, active, stable, and retired states based on fitness. It reports the highest aggregate success rate across several interactive agent benchmarks, with up to 7.8% relative improvement over the strongest baseline, while keeping the library compact. The authors also introduce SkillFurnace, a dataset of over 5k annotated records.
In briefMOTIF: Detecting deepfakes of a specific person beyond 3DMM coefficients
SummaryVideo deepfakes targeting a specific person are the most harmful, and detectors can be built from that person's genuine footage. The paper dissects the 3D Morphable Model (3DMM) coefficients these detectors use, finding the shape block alone recovers almost all accuracy, while temporal changes add a real but bounded contribution. The authors show the dense surface that detectors discard carries identity information the coefficients lack, and assemble this into MOTIF, a visual detector trained only on real videos. MOTIF outperforms two state-of-the-art person-specific detectors across all datasets, manipulations, and quality levels in the benchmark.
WhyDissects 3DMM coefficient groups under fixed settings and shows dense surfaces carry identity information the coefficients lack, with consistent gains.
Giovanni Affatato, Sara Mandelli, Paolo Bestagini et al.
In briefDual-QK: Sharp Queries and Flat Keys for Prunable 2-bit KV Caches, a method to shrink long-context memory use
SummaryThis paper addresses the high storage and read costs of the key-value cache in long-input, long-generation language models. It proposes Dual-QK, which pairs non-orthogonal query and key transforms so that keys suit 2-bit quantization while query energy stays concentrated enough for channel pruning. The authors report better accuracy than OSCAR on most tasks at 40% query-channel sparsity, and 6.8x KV-cache compression at 128K context.
In briefWhy multi-agent AI systems can become too similar, and the risks that follow
SummaryThis paper studies homogenization, where agents in multi-agent systems converge to similar behaviors through their interactions. The authors measure it with three metrics: conformity to the majority, polarization toward extremes, and growing inertia to change. They evaluate it in code generation, hiring, and scientific peer review, finding risks such as amplified correlated errors and uneven evaluation standards. Simple diversity methods, such as sampling randomness and mixed models, did not reduce these risks.
In briefUltraText Bench: A bilingual benchmark for testing how well image generators draw dense text
SummaryThis paper introduces UltraText Bench, a bilingual English-Chinese benchmark for evaluating image generators on dense visual text. It has 432 prompts across 24 scene categories and three difficulty levels, each specifying exact strings for four to twelve text regions. The authors use the Q-Judger vision-language model to score images on text fidelity, text clarity, spatial quality, and scene quality. Results across 24 model configurations show that different models have different strengths, and performance drops noticeably as workload difficulty increases.
In briefLightweight 3D Gaussian head avatar generation that runs on edge devices
SummaryThis paper proposes an efficient generator for unconditional 3D Gaussian head synthesis, using a parameter-efficient synthesis block and depth-wise separable convolutions while keeping style-based conditioning. Without model compression or quantisation, it reports 94% fewer FLOPs, 70% fewer parameters, and 81% smaller model size than the baseline, with competitive generation quality. The authors also demonstrate CPU inference and browser-based execution on mobile devices using ONNX Runtime.
Umar Farooq, Jean-Yves Guillemaut, Adrian Hilton et al.
In briefBackdoor attacks on acoustic foundation models using physical sound triggers
SummaryThis paper proposes the FAB attack, a backdoor method targeting acoustic foundation models (AFMs) under practical, minimal-access conditions. The backdoors preserve normal performance, survive fine-tuning, and degrade multiple downstream tasks when triggered by inconspicuous physical sounds such as a background siren. The authors evaluate the attack on two AFMs, nine tasks, and four triggers, and report that extensive end-to-end fine-tuning can mitigate it at high cost. They call for stronger defenses.
WhyProposes a backdoor attack on acoustic foundation models that survives fine-tuning; evaluated across two models, nine tasks, multiple triggers, and defenses.
In briefConcentration, Not Uncertainty: Why Targeted Synthetic Data Fails for Camouflaged Object Detection
SummaryThis paper studies whether uncertainty-guided allocation of synthetic training data outperforms random allocation for camouflaged object detection under a fixed generation budget. Across 103 training runs, uncertainty-based targeting did not beat random allocation, and control experiments attribute the observed effects to budget concentration rather than where uncertainty lies. The authors also report that 50 of the 76 images in the CHAMELEON benchmark are duplicates of training data.
In briefBoT-GRPO: Faster reasoning training for language models using token-level rewards without a critic
SummaryThe paper proposes BoT-GRPO, an extension of GRPO that aggregates token-level rewards across rollouts using inverse-length weighting, without needing a value network. On React code generation it reaches 80% compile rate up to 1.9 times faster than GRPO, and on AIME math reasoning it improves Pass@k by up to 8.1% in half the steps. The authors also report that stable, bounded reward signals matter more than rich ones for learning speed.|
Yingxiang Yang, Weihang Xiao, Zhunxuan Wang et al.
In briefDisParQ: Self-supervised part concepts that make vision models easier to interpret
SummaryDisParQ learns discrete, spatially grounded concepts from a frozen, vision-only self-supervised backbone, without class labels or language supervision. Each image patch is assigned to one concept from a learned prototype dictionary, and continuous residuals are quantized into discrete attributes. A decoder reconstructs the backbone's representation from these concepts and attributes, showing the discrete form preserves its information. The authors report 83.2% top-1 on ImageNet linear probing, close to the DINOv2 teacher, plus higher concept consistency than language-aligned models and cross-category part-based retrieval.
Adam Pardyl, Siddhartha Gairola, Sukrut Rao et al.
In briefBeyond Masks and Trajectories: Flow-Guided Latent Action Injection for Stable Surgical Video Generation
SummaryThis paper proposes FLAIR, a reference-free framework that generates surgical videos from text prompts alone, without masks, trajectories, or other auxiliary conditions. It learns action priors from optical flow in real surgical videos and injects predicted latent action representations into a frozen base model to improve action consistency. The authors also introduce SurgActionClip-30K, a large action-centric surgical dataset with structured captions, and SurgMetrics, surgery-specific evaluation metrics.
In briefMIRROR: Moving LLM personalization from copying style to learning user preferences
SummaryThis paper introduces MIRROR, a self-distillation framework that shifts LLM personalization from imitating reference wording toward internalizing user preferences. It aligns the model's next-token distributions on its own generations with a reference-conditioned version of itself, and adds MIRROR-F, a plug-in that selectively supervises informative reference tokens. The authors report leading personalization performance and better text quality across three benchmarks, with less catastrophic forgetting than SFT-based baselines on unseen tasks.
In briefWeakly supervised labeling of fine-grained EEG components to reduce muscle artifacts: a proof-of-concept study
SummaryEEG signals are often contaminated by electromyographic (EMG) artifacts that overlap with neural activity, and reliable component-level ground truth is scarce. The authors propose combining a frequency-aware high-dimensional representation, which splits separated components into frequency-resolved intra-components, with Multi-Instance Learning that learns artifact-likelihood scores from epoch-level labels alone. On held-out subjects, the framework reduces artifact-related spectral deviations most clearly for jaw tension, with moderate effects for raising eyebrows and limited effects for frowning.
In briefCounterfactual Route Optimization for Gaussian Head Avatar Modeling
SummaryThis paper addresses the challenge of jointly optimizing geometry, appearance, and cross-view consistency when building Gaussian head avatars. The proposed framework compares short-horizon outcomes of alternative update routes from the same training state, using a unified utility to decide which optimization direction is more beneficial at that point. A lightweight controller learns this preference from offline evidence and applies bounded modulation to the training objectives. Experiments on the NeRSemble dataset reportedly outperform existing methods, with clearer local facial structures and finer details.
In briefJudging in Latent Space: Compressing Reward Model Evaluations to Cut Inference Cost
SummaryThe paper introduces LatentGRM, a framework that compresses rubric-guided evaluation into compact continuous latent trajectories, allowing pairwise judgments without generating textual assessments. Under matched data and backbones, it achieves competitive aggregate preference accuracy compared with explicit SFT judges at 4B and 8B scales. For LatentGRM-8B, evaluation trajectories are compressed 8.9–9.2x and judge inference time is reduced 6.1–7.0x at vote@5.
In briefUltraWorld: Learning ultrasound world models from untracked clinical videos using an acoustic sampling map
SummaryThis paper presents UltraWorld, a method that learns interactive ultrasound world models, which predict how probe movements change the observed image, from routine clinical videos without real action labels. It uses self-distillation with synthetic action-video pairs generated from anatomical masks, and introduces an Acoustic Sampling Map that encodes probe pose and imaging settings. In nine simulated closed-loop planning episodes, the authors report 29% and 38% reductions in final distance to goal and orientation error, respectively, compared with visual servoing.
In briefCIRSeg: A liver segmentation method for contrast-enhanced MRI that handles intensity variation and adapts to new scanners without source data.
SummaryThis paper presents CIRSeg, a coarse-to-fine liver segmentation framework for contrast-enhanced MRI built on nnU-Netv2. It uses 3D CutMix and stochastic intensity transfer (Nyul or histogram matching) to handle varying MRI intensities, separates low-resolution localization from full-resolution boundary refinement, and applies source-free test-time adaptation. Largest connected component filtering removes isolated false positives; on the CARE 2026 test set it reports Dice of 97.13% and 97.93% and HD95 of 20.18 mm and 11.30 mm on in-domain and unseen-domain subsets.
In briefAdaPS-LiNGAM: Adaptive Predecessor Selection for Small-Sample Causal Discovery
SummaryThis paper studies causal discovery when samples are few relative to variables, a problem affecting the LiNGAM framework. It shows that DirectLiNGAM's repeated residualization becomes degenerate when variables exceed samples, before the full causal order is found. The proposed AdaPS-LiNGAM rebuilds each residual from the original data using an adaptively chosen sparse set of earlier variables, and uses the same selection for edge pruning. Synthetic experiments indicate accurate structure recovery with small samples and a more gradual decline in performance as sample size decreases.
In briefUsing object relations to help diffusion models solve spatial reasoning tasks
SummaryDiffusion models generate images well but often fail to reliably satisfy structured spatial reasoning constraints. The paper proposes a framework that uses unsupervised object discovery and abstractions of object relations to guide generation during training and inference. The authors also introduce a generative spatial reasoning benchmark with four puzzle-inspired datasets. The reported results show improved reasoning on complex tasks and robust generalization to out-of-distribution settings.
In briefReproducible LLM Inference Benchmarking: A Protocol for Stable Regression Testing
SummaryThis paper proposes a Sequential Isolation Methodology, a controlled protocol for benchmarking LLM inference that reduces variance between repeated runs. The authors tested three open-source models on an NVIDIA A100 80GB GPU with vLLM 0.9.1, across six context sizes and eight concurrency levels, with five repetitions each. The final protocol lowered average coefficient of variation from 15.2% to 2.2%, and the measurements showed a marked latency change between 200 and 500 concurrent users on the tested stack.
Arnold Olympio, Juan Manuel Servera Bondroit, Wael Abdelmalek et al.
In briefSeparating logic from persona to protect edge LLM agents from context pollution
SummaryThe paper studies how the logical reasoning of an edge language-model agent degrades when its context window is filled with long, misleading, persona-heavy history. It proposes AO-DA, a Decoupling Architecture that splits logical inference from persona expression into two inference paths on one INT4 base model with hot-swappable LoRA adapters. In experiments on an Apple M2 laptop, the decoupled logic path was unaffected by pollution, while the mixed single-pass approach degraded steadily as pollution increased.
Masaaki Nakatsu (AO, Inc. / OrbLabs AG), Reno Wang (AO et al.
In briefPathways for AI from Weather to Climate: A Review and Two Minimum Requirements
SummaryThis paper reviews recent progress in deep learning for weather forecasting, where autoregressive models now rival dynamical models from nowcasting to subseasonal-to-seasonal timescales. It examines the challenges of extending these methods to climate projections under altered forcings. The authors propose two minimum requirements for AI climate modeling: explicit inclusion of external forcing agents, and stress-testing robustness in out-of-distribution regimes such as extremes and counterfactual scenarios.
In briefBolzano: Automated Open-Problem Solving Built on Expert-Guided Proof Search
SummaryThis paper presents Bolzano, an open-source multi-agent system with parallel prover agents, a verifier agent, and a human-readable research state. Manual use on expert-selected problems produced 8 results whose proofs domain experts checked. Run without problem-specific human guidance on about 3,800 open problems from four sets of papers, it solved about 200. In one experiment on STOC 2026 accepted papers, it answered four questions raised in those papers, as confirmed by the authors.
Adri\'an Z\'ame\v{c}n\'ik, Mat\v{e}j Kripner, Martin Kouteck\'y et al.
In briefHow Random Fluctuations in Neural Network Training Affect Measured Quantities
SummaryThis paper studies how fluctuations in the parameter distribution of wide shallow neural networks trained by stochastic gradient descent carry over to finite-dimensional nonlinear observables. Working in the weighted Sobolev space where the limiting fluctuation process is built, the authors apply a functional Delta method and obtain a central limit theorem with an explicit covariance formula. They also give a condition, under a constant rank assumption, for deciding whether chosen observations locally determine a target quantity.
In briefBeyond Policy Support: Interaction-Constrained Offline RL for Autonomous Driving
SummaryThis paper proposes Interaction-Constrained Drive Policy (ICDP), an offline reinforcement learning method for autonomous driving that controls distribution shift at the level of ego and surrounding-agent interactions. It uses contrastive density-ratio estimation to isolate interaction compatibility without explicit joint-density modeling or rollouts. The authors report that closed-loop evaluations on nuPlan, Interplan and real-world truck experiments show ICDP suppresses high-value but interaction-unsupported trajectories and improves performance in interaction-critical scenarios.
Mahmoud Selim, Cristina Cipriani, Karl Henrik Johansson
In briefPARC-Loc: Locating Positions in 3D City Maps from Text Using Partial Matching of Object Layouts
SummaryThe paper proposes PARC-Loc, a method that estimates a position in a city-scale 3D point cloud map from text descriptions of nearby objects. It uses Partial Assignment with Relational Consistency to model object compatibility and spatial relations, allowing unmatched elements while favoring layouts that match the query. The approach selects submaps more reliably and expands context with relevant objects from adjacent submaps. On KITTI360Pose, Top-1 recall at 5 m rises from 0.50 to 0.67, a 34% relative gain over the strongest baseline.
In briefSmaller transformers can learn to perform clustering easily
SummaryThis paper studies how transformers can carry out Lloyd's algorithm for k-means clustering. The authors present a smaller transformer that is equally expressive, using embedding size d + ⌈log₂ k⌉ instead of the previously required d + k. They train these models on a distribution of clustering tasks and analyze, theoretically and empirically, factors affecting convergence and in-distribution generalization. They also examine when the learned clustering algorithms succeed or fail.
Charlotte Park, Kenneth L. Clarkson, Lior Horesh et al.
In briefUnrolled Flow Models for Reasoning: Using More Integration Steps to Improve Reasoning
SummaryThe paper proves that a flow model parameterized by a two-layer Transformer can solve graph reachability, with the required integration steps growing with the target's distance from the root. It finds that standard flow language models often fail to benefit from extra steps because their objectives supervise each time point independently. The authors train through the model's own latent rollout, retract the latent state onto a sphere to stabilize dynamics, and report gains on ProsQA, Sudoku, and Maze.
WhyProposes rollout training for flow models, proves graph reachability capability, and validates on ProsQA, Sudoku, and Maze.
In briefEntroPrefill: Entropy-guided context pruning for faster retrieval-augmented generation, with theoretical guarantees
SummaryThe paper proposes EntroPrefill, which prunes context mid-prefill in transformer models so deeper layers process fewer tokens, using Rényi-based constraints on discarded attention mass. It derives a computable bound on feasible token removal, a finite-sample observer guarantee valid under adaptive layer selection, and conditional perturbation bounds. The work is purely theoretical and does not report measured speedups or task accuracy, leaving experimental validation to future work.
In briefShaer: controllable Arabic classical poetry generation using meter and meaning
SummaryShaer is a framework that generates Classical Arabic poems controlled by natural-language descriptions, meter subforms, and target line counts. The authors built a corpus of 116,032 poems from Ashaar and fine-tuned Yehia-7B with QLoRA. The model reports 95.17% base-meter accuracy, 91.75% poem-level meter-subform accuracy, and 83.40% exact count accuracy. Analysis of 3,481 test generations found no exact copies from the training corpus.
In briefWatermarking Diffusion-Generated Images: A Framework and Case Studies
SummaryThis paper classifies watermarking methods for diffusion-generated images along two axes: insertion mechanism and primary signal-bearing representation. It formalizes a representative z_T-Fourier pipeline for verification and identification. Three protocol-bounded case studies examine frequency integrity, persistence under editing, and single-shot VAE-latent phase modulation. The authors do not claim a universal ranking but offer a framework for protocol-aware comparison.
In briefAI Plankton Image Classification for Detecting Temporal Occurrence Events Across Datasets
SummaryThis study tests whether AI classifications of phytoplankton images, aggregated over time, can reproduce expert-derived occurrence patterns. Models were trained on MedPlanktonSet and evaluated on an independent IFCB dataset from SMHI, with EfficientNet-B0 selected for temporal analysis. After aggregation, AI-derived occurrence rates closely matched expert signals (Pearson r = 0.9696), suggesting the approach can provide monitoring-relevant temporal indicators.
In briefSoftSEEPS: a differentiable version of a precipitation score to improve ML rainfall forecasts
SummaryThis paper introduces SoftSEEPS, a differentiable approximation of the SEEPS score, so a machine learning model can be trained to forecast precipitation directly. The method is tested on the IMERG dataset at 0.1 degree resolution by training a precipitation decoder on the latent space of a pre-trained low-resolution forecasting model. The authors report that combining SoftSEEPS and RMSE in one joint objective is possible, with only marginal trade-offs in either metric.
In briefFlow-of-Thought: a visual reasoning framework using intermediate sketches for spatial tasks
SummaryThis paper introduces Flow-of-Thought (FoT), a framework that generates visual sketches as intermediate reasoning steps, mimicking human mental imagery. The method trains trajectory flow fields and compares competing generative hypotheses on frozen dynamics. FoT reaches 100.0% accuracy on Tetris and 99.0% on colored shapes, and on BLINK Multi-view under frozen transfer it scores 72.2% versus 63.9% for an endpoint-only control.
In briefSoccerNet-FoulRet: Finding Soccer Foul Videos with Similar Refereeing Precedents
SummaryThis paper introduces SoccerNet-FoulRet, the first benchmark for retrieving soccer foul videos that referees would consider relevant precedents, regardless of camera angle, teams, or appearance. It is built from the SoccerNet-MVFoul dataset and evaluated on 693 human-verified queries. The strongest zero-shot model reaches under 5% HitRate@10 on human-verified precedents, and category-supervised fine-tuning transfers only modestly to precedent retrieval.
In briefHealthy AI Skepticism: A Data Visualization Research Agenda
SummaryThis paper argues that data visualization research on AI has mostly aimed to build trust through visual explanations, but human trust in AI ranges from critical skepticism to over-reliance. The authors contend that a skeptical stance toward both AI results and AI use is healthy. They discuss the rise of over-reliance on AI, its risks and opportunities, and the role of data visualization in situations where people are not motivated to think critically.
G. Elisabeta Marai, Marc Baaden, Michael Behrisch et al.
In briefA strength-monotonic law for when domain alignment helps frozen bioacoustic embeddings
SummaryThe paper studies when distribution alignment (MMD) helps a frozen foundation-model embedding generalize across acoustic domains in mosquito-species classification. The authors report that stronger encoders rely more on the MMD term for unseen-domain generalization and are more harmed by domain-rebalanced sampling. Based on this, they propose a reduced recipe using a frozen Perch 2.0 embedding, a lightweight probe, cross-entropy, one MMD term, and input augmentation, which performs within seed noise of the full composite.
In briefBridge Routing Heads: Internal Circuits for Multilingual Multi-hop Reasoning in LLMs
SummaryThis paper asks whether multilingual large language models use a shared internal circuit when answering multi-hop reasoning questions in different languages. Using a three-stage pipeline, the authors identify Bridge Routing Heads (BRH) in two models and find that language-specific head sets are nearly mutually exclusive across five languages. Ablating general BRH increases two-hop negative log-likelihood, and amplifying these heads rescues up to 51.7% of cross-lingual failures without training.
In briefA study of how key kernel properties relate when kernels are unbounded
SummaryThis paper examines the expressivity of unbounded kernels used in machine learning. Kernel methods rely on the reproducing kernel Hilbert space of each kernel, and concepts such as characteristic, L_p-universal, and integrally strictly positive definite describe that expressivity. These concepts are well understood for bounded kernels, but their relations for unbounded kernels have received little study. The authors establish these relations under mild assumptions.
Jose Cribeiro-Ramallo, Florian Kalinke, Zolt\'an Szab\'o
In briefBeyond Group Splits: Specimen-Level Validation and Visual Attribution for Fruit Shelf-Life Prediction
SummaryThe paper estimates remaining shelf life of avocados from images using a frozen ImageNet-pretrained backbone with a lightweight regression head. It shows that splitting data by observation yields much more optimistic error (RMSE 2.37 days) than specimen-disjoint cross-validation (RMSE 3.12 days) on 426 fruits. Under specimen-disjoint evaluation, MobileNetV3-Small matches ResNet-18 accuracy with higher throughput, and Grad-CAM shows differences in spatial attribution between the two models.
Rovhona Mudau, Jean Frederic Isingizwe Nturambirwe, Clement Nthambazale Nyirenda
In briefExplaining Why Query Expansion Methods Perform Differently in Search
SummaryThis paper examines why query expansion, a long-used search technique, works better for some queries than others. It explains performance through two views: an Ideal Expanded Query that maximizes BM25 retrieval effectiveness, and a separability measure using Cohen's d to assess how well relevant and non-relevant documents are distinguished. Experiments on several TREC collections find that expanded queries closer to the ideal tend to perform better, and separability offers a complementary explanation.
Sourav Saha, Aditya Dutta, Soumajit Pramanik et al.
In briefMeshCarve: Generating detailed 3D meshes in compact latent spaces with flow matching (paper)
SummaryThis paper presents MeshCarve, a method that generates detailed 3D artisan meshes entirely in compact latent spaces. It generates vertex positions and edge connections separately, using VertexVAE and EdgeVAE anchored on discrete vertices to shorten token sequences. The authors report it outperforms state-of-the-art autoregressive and flow matching methods on Objaverse and generalizes to Toys4K.
In briefDynStream: Online 4D Gaussian reconstruction of dynamic scenes from unposed video streams
SummaryThis paper presents DynStream, a framework that reconstructs dynamic 4D scenes online from long streams of video without known camera poses. It rebuilds the scene within local time windows, then aligns and fuses them into one globally consistent scene without per-scene optimization. Experiments report high-fidelity online reconstruction and rendering across diverse indoor and outdoor dynamic scenes, achieving state-of-the-art results.
In briefYUBI-STAG: Automatically Adding Contact and Semantic Labels to Robot Videos to Align Vision-Language-Action Models
SummaryThe paper presents YUBI-STAG, a framework that automatically annotates robot manipulation demonstrations with fine-grained details such as object identities, per-gripper actions, bimanual coordination, and spatial relations. It also introduces YUBI-VLM, a distilled model that recovers these annotations directly from raw, unsegmented video in fewer inference calls. The authors report that post-training Vision-Language-Action policies on these annotations improves bimanual performance and instruction following.
Masatoshi Tateno, Takehiko Ohkawa, Yueh-Hua Wu et al.
In briefSTLQ: Spatio-Topological Linkage Quantization for Approximate Nearest Neighbor Search
SummaryThis paper proposes STLQ, a shallow multi-codebook quantization method for compressing vectors and enabling approximate nearest neighbor search. It organizes each IVF cluster as a directed forest, first quantizing vectors independently, then re-encoding residuals relative to reconstructed parent vectors to reuse local information. The authors report higher recall than evaluated shallow methods at comparable bit budgets, with higher offline construction cost and competitive recall-QPS tradeoffs against deep methods.
In briefEC-EarthFlow: Fast probabilistic emulation of daily climate model simulations
SummaryThe paper introduces EC-EarthFlow, a generative flow matching model that emulates the physical climate model EC-Earth3. It is trained on EC-Earth3 transient simulations under SSP2-4.5 (1950-2166) to predict the next day's temperature field from previous days and annual mean temperature, rolled out autoregressively. The authors report that using only this variable, the model reproduces daily variability, spatial patterns, annual cycle and long-term trend at much lower computational cost, remains stable over long inference periods, and learns the physical relationships simulated by EC-Earth3.
Kirien Whan, Nikolaj T. M\"ucke, Karin van der Wiel
In briefBoundary-aware reinforcement learning for hypercube state spaces using deterministic policy gradients
SummaryThis paper proposes a continuous-time deterministic policy gradient framework for reinforcement learning where states follow a controlled reflected stochastic differential equation on a hypercube. It links the value function to the Neumann Bellman equation, derives a deterministic policy gradient formula, and proves a martingale characterization theorem. Based on this, the authors present a continuous-time deep deterministic policy gradient algorithm that enforces the Neumann boundary condition through soft penalties or hard architectural constraints. Experiments on reservoir control show that boundary-aware methods substantially reduce Neumann boundary residuals and improve learning stability.
In briefSpikingVLA: Asynchronous Spiking Vision-Language-Action Models for Low-Latency Efficient Robot Inference
SummaryThis paper presents SpikingVLA, a framework that converts pretrained artificial neural networks into spiking neural networks for energy-efficient Vision-Language-Action inference. It introduces a Dendritic Integrate-and-Fire neuron to reduce the number of timesteps needed for accurate conversion, and an asynchronous execution mechanism that overlaps computation across VLA components. Experiments report that, compared with existing spiking VLA methods, SpikingVLA improves SR and SPL by 11.9% and 12.6%, and reduces first-action latency by 11.2 times.
In briefPredicting Foundation Model OOD Robustness from Pretrained Weight Spectra
SummaryThe paper proposes predicting whether a foundation model will generalize out-of-distribution using only its pretrained weights, without target data. The authors prove the OOD accuracy gap is bounded by how concentrated source representations are, and use a weight-based spectral statistic as a proxy whose direction depends on architecture and pretraining strategy. Across 116 models, the chosen statistic ranks 92% of model pairs within each cell, and predictions held for held-out model families including EEG, genomic and protein.
In briefInterior-Guided Boundary Repair for Better Multi-Region Style Transfer
SummaryThis paper addresses boundary artifacts that appear when separately stylized image regions are composited in region-based neural style transfer. It proposes IGBR, a lightweight, model-agnostic method that repairs boundary pixels using interior-guided propagation and applies distance-based blending limited to object-background boundaries to prevent style leakage between objects. The authors also introduce metrics that evaluate boundary consistency, gradient artifacts, inter-object leakage, and interior preservation without annotated stylized images, and report that IGBR outperforms prior blending techniques on these measures.
In briefLatent Watermarks Under Generative Editing: Benchmark on Detection Survival
SummaryThis paper benchmarks eight watermark methods against five editors, four generative backbones, four editing strengths, and five semantic categories. It finds that ordinary prompt-based editing alone can cause latent watermark detection to fail, without explicitly targeting the watermark. The authors report that clean standardized score separation (d') organizes survival tiers, while continuous score separation reveals differences that aggregate detection rates hide.
In briefAuditing LLM Program Search for Bermudan Options with Primal-Dual Certificates
SummaryThis paper studies LLM-driven search for Python policies that decide when to exercise Bermudan options. A trusted evaluator supplies lower and dual upper bounds, which the authors use to audit 24 generation attempts, a feedback-representation ablation, and a holdout set frozen before evaluation. No selected program consistently beats the Ridge LSMC baseline by more than Monte Carlo noise. On one-dimensional puts, most of the estimated gap traces to a loose dual bound, so the audit confirms validity but not progress.
In briefWhy Synthetic Hard Negatives Work in Vision-Language Pretraining
SummaryThis paper studies how synthetic hard negatives created in representation space can be used in vision-language pretraining, which is harder to transfer than in unimodal learning. The authors identify two failure modes in cross-modal and intra-modal synthesis and report logit-scale saturation with a learnable temperature, which fixing the temperature improves. They propose SNAP, which generates intra-modal hard negatives that never involve the positive, requiring no external generative models and adding under 10% training overhead. Tested on CLIP and FLIP across several architectures and datasets, SNAP consistently improves zero-shot retrieval, zero-shot classification, and linear probe results.
Nikos Giakoumoglou, Paschalis Giakoumoglou, Andreas Floros et al.
In briefSAVU-Bench: A Real-World Benchmark for Spatial Audio-Visual Understanding
SummaryThe paper introduces SAVU-Bench, a real-world benchmark testing how well models understand spatial relationships between audio and video across three capability levels and seven tasks. It also presents SAVU-Diag, a diagnostic set that breaks reasoning questions into grounding and alignment subtasks. Evaluating 12 models, the authors find visual spatial grounding is relatively mature while audio-involved spatial perception is the main bottleneck. They further propose SAVU-EA, a training-free baseline that improves spatial grounding and joint matching, though high-level spatial reasoning remains difficult.
In briefRecurrent Self-Improvement: Dynamic Cross-Loop On-Policy Distillation for Looped Language Models
SummaryThis paper studies Looped Language Models, which reuse shared parameters across recurrent computation steps to scale reasoning efficiently. It proposes LoopOPD, which uses a frozen terminal loop as a teacher for an intermediate loop, and D-LoopOPD, which refreshes that teacher as parameters update. Experiments on Ouro-Thinking models report better mathematical reasoning, and models trained only on math data also improve on general reasoning and code generation benchmarks.
In briefDo Better Visual Representations Always Improve End-to-End Self-Driving?
SummaryThis paper studies when visual representations from foundation models improve end-to-end autonomous driving. It introduces ViRA, a planner-agnostic alignment framework that leaves planner architecture and inference cost unchanged. The authors find that these representations generally help, that the choice of alignment target matters, and that auxiliary perception supervision reduces that sensitivity. Guided by these results, they present ViRA-Diffusion, which reaches 92.3 EPDMS on NAVSIM v2 navtest without auxiliary perception supervision.
In briefTwo Error Types: Hidden Trade-offs in Automated Judges When Evaluating Uncertainty Quantifiers
SummaryThe paper examines how uncertainty quantification (UQ) methods for LLM outputs are evaluated. A meta-analysis shows automated judgement is the norm, yet it is rarely validated against human judgement. Experiments in question answering with 4 LLMs and 7 popular UQ methods find that a judge's performance only coarsely predicts how its errors affect UQ evaluation reliability, and that judgement errors most often misrepresent the most informative UQ methods.
In briefFew-shot learning for personalised automated pain assessment
SummaryThis paper applies Few-Shot Learning, a meta-learning sub-area, to personalise automated pain assessment across individuals. It reframes the move from population-level to subject-level evaluation as a task-domain shift and tests on the BioVid, SenseEmotion, and PainMonit datasets. Using a few samples for k-shot conditioning raises accuracy, and the authors suggest support-conditioned adaptation helps under inter-subject variability.
In briefWorldBench: Benchmarking LLM-Generated Three.js Voxel Worlds
SummaryThis paper introduces WorldBench, a benchmark and judge for evaluating open-ended 3D worlds that large language models write as Three.js code. The judge runs each world, controls its clock, orbits it, and uses a navigator agent to frame each biome while also reading the source code, checking visual claims against measured pixels. In a mutation test that removes features, the method cuts credit for removed features compared with code-only judging. Five frontier models are evaluated.
