In briefHierarchical Incremental Beta Splatting for Online Reconstruction: a new method for building 3D scenes during capture
SummaryThe paper presents Hierarchical Incremental Beta Splatting (HIBS), which applies beta kernels to online incremental 3D reconstruction. A shape parameter moves the kernel between a soft, Gaussian-like profile and a flat-topped, near-opaque surfel, serving as a geometric confidence control. The method adds confidence-aware primitive spawning, error-driven relocation, bounded-support fast rasterization, and a level-of-detail hierarchy built by shape-preserving merging. On Mip-NeRF 360 under the on-the-fly protocol, the authors report PSNR gains of up to 0.8 dB and optimization time reductions of up to 40%.
In briefInteractive Visual Gym: Testing Visual Assistant Models with Simulated Users
SummaryThis paper introduces Interactive Visual Gym, a framework that evaluates and trains visual interaction models in closed loop by jointly simulating executable task environments and human users. The authors build three simulators covering cooking, household manipulation, and mobile interfaces, and release IVG-Bench, a benchmark of 315 interaction scenarios. Tests of state-of-the-art vision-language models show limited performance, with the best model completing 53.4% of tasks within the time limit.
In briefA Generative Model of Contextual Integrity: Appropriate vs. Inappropriate Sharing
SummaryThis paper builds a generative model of contextual integrity to study how information is shared through simulated conversations. Two actors played by large language models exchange messages, with one holding private information and the other seeking it. The study examines how sharing and refusal vary with the situation, the holder's decision logic, disposition, and the underlying language model. Actors shared information with rightful recipients when needed, but also disclosed it to people not entitled to receive it.
In briefMultimodal-SEA: A Benchmark for AI Agent Security Against Multimodal Social Engineering
SummaryThis paper introduces Multimodal-SEA, an interactive benchmark of 608 expert-designed tasks based on real-world fraud cases, spanning 14 domains and 152 attack scenarios with multimodal content. Seven widely used multimodal models were evaluated in simulated environments, and all proved vulnerable, with attack success rates between 7.1% and 34.5%. Adding multimodal content raised attack success from 19.9% to 29.2%.
WhyThe authors introduce a 608-task interactive benchmark for multimodal social engineering attacks, reporting higher attack success across seven multimodal models.
In briefLeanContext: separating math resources from local context for Lean proof generation
SummaryThis paper introduces LeanContext, a benchmark of 129,826 Lean whole-proof generation problems for studying how retrieved premises and project context help proof generation. It separates selected complete preceding mathematical declarations (Mathematical Context) from the remaining local source commands (Environment Context). Selected declarations outperform comparable alternatives by a larger margin when training includes Mathematical Context. Prioritizing dependencies omitted from the final input improves Pass@32 by 1.40% for LLaDA and 1.70% for Llama on 2,000 targets.
In briefTraining image-based robot policies to learn 3D shape without 3D inputs
SummaryThis paper studies whether an image-based robotic manipulation policy can acquire 3D shape information during training while keeping an image-only interface at deployment. The authors propose shape-teacher alignment, which uses Point-MAE features from individual objects as alignment targets for corresponding image patches in the policy encoder. The deployed inputs and architecture are unchanged. On LIBERO, the method yields image representations more structured around object shape and parts, and substantially speeds convergence and improves success rates. For pretrained policies, convergence is accelerated 4.7x on the Object suite and 5.6x on the Spatial suite.
Madelon Bernardy, Mohammad Mahdi Derakhshani, David M Knigge et al.
In briefSpeeding up safe reinforcement learning by training many environments in parallel
SummaryThis paper studies how to scale safe reinforcement learning to massively parallel training, where many environments run with short rollout horizons. To handle the mismatch between short trajectories and episode-level safety constraints in constrained MDPs, the authors propose a phase-aggregated cost estimator and a constraint-tightening framework that accounts for finite-sample uncertainty. Experiments using an MJX-based implementation of Safety Gym show performance comparable to non-massively parallel methods with substantially reduced wall-clock training time.
In briefDART: A plug-in method that separates anomaly signals from domain-specific factors for generalist anomaly detection.
SummaryThis paper addresses Generalist Anomaly Detection, which aims to find anomalies in unseen domains using only a few reference samples. The authors propose DART, a plug-in framework that separates residuals into anomaly-relevant and domain-specific components through a Residual Disentanglement Module and Cross-Domain Meta-Training. Experiments on MVTec AD, VisA, and BraTS show consistent improvements over multiple base methods, with the largest gains under stronger domain shifts.
Muhammad Aqeel, Maham Nazir, Marco Cristani et al.
In briefPassing Yet Wrong: Silent Failures of LLM Code Repairs After API Migrations
SummaryThis paper studies how LLM-repaired programs can run without errors after an API or runtime migration yet still return wrong results, and how release checks on typical inputs let these programs ship. The authors ran a preregistered study of 73 tasks covering 60 documented changes in CPython, Node.js and SQLite, executing every candidate on the real runtimes against withheld contracts. An all-assertions gate on typical inputs released 0.410 of wrong repairs, which fell to 0.065 after adding one edge input written from the release note.
In briefSplatAnything: Gaussian Spacing Normalization for Feed-forward 3D Gaussian Splatting Across Arbitrary Scenes
SummaryThis paper identifies a limitation of existing feed-forward 3D Gaussian Splatting methods, where Gaussian attributes are predicted in a radial-distance-normalized space while the spacing between neighboring Gaussian centers varies across scene scales and camera setups. The authors propose SplatAnything with Gaussian Spacing Normalization (GSNorm), which normalizes a robust scene-level statistic of Gaussian spacing to provide a more consistent reference. A systematic benchmark covering diverse datasets, input sparsity levels, and extrapolated viewpoints is also established. Experiments report state-of-the-art performance, especially under large variations in scene scale and viewpoint.
In briefImproving brain-signal motor decoding using unpaired behavioral data
SummaryThe paper introduces BEEMO, a framework that trains neural motor decoders on arbitrary mixtures of paired neural-behavioral recordings and unpaired behavioral data. A behavior decoder learns movement dynamics from behavioral trajectories, and neural activity is added through cross-attention when available. On intracortical motor datasets from nonhuman primates, adding unpaired behavioral data improves decoding, especially when transferring to new tasks. Behavior-only pretraining performs comparably to models pretrained on much larger neural datasets.
Divyansha Lachi, Vinam Arora, Shivashriganesh P. Mahato et al.
In briefSimulated Perspectives in Multimodal Models: Representation vs. Behaviour
SummaryThis paper examines how assigned personas change the internal representations of Qwen3-VL-8B-Instruct when judging how interesting images are. Across 100 images and 53 prompting conditions, the authors recorded behavioural ratings and layer-wise hidden states. Early layers mainly encode image identity, while persona-specific structures appear in later layers. Prompt framing alters internal representations more than persona identity, whereas persona content has a greater effect on judgements.
Mathis Immertreu, Fitim Abdullahu, Thomas Kinfe et al.
In briefEstimating Tropical Cyclone Wind Fields and Intensity from Geostationary Satellite Images
SummaryHigh-resolution surface wind speeds from synthetic aperture radar (SAR) help model storm structure and intensity but are sparse and not available in near-real time. This paper presents simple neural networks that predict two-dimensional wind fields and storm-scale intensity metrics from geostationary satellite imagery. Ablation studies show that jointly predicting the 2D fields and scalar metrics from the model's latent space improves retrieval of the maximum wind and its radius, especially during rapid intensification.
Simon Donike, Pritthijit Nath, Cristina Radin et al.
In briefSplitting Data by Series for Cross-Validation in Global Forecasting
SummaryThis paper studies cross-validation for time series forecasting with global deep learning models. The authors formalize series-wise cross-validation, which partitions data across series rather than only over time, and compare K-fold, bootstrapping, and Monte Carlo variants on 11 datasets with 10 neural network architectures. Series-wise cross-validation leads to better downstream forecasting accuracy by favoring models that generalize across time and series, while time-wise cross-validation has a slight edge in estimating out-of-sample error.
In briefGeometric Relevance Is Not Sufficient: Evaluating Focused Supervision in Diffusion Docking
SummaryThis paper studies post-training for DiffDock, a diffusion-based molecular docking model, under a fixed training budget. It compares broad continuation with torsional supervision focused on Cartesian consequence or conditional target consistency, and with a mixture of both. Evaluation separates fixed score fidelity, generated-pose quality, best-of-10 sampling capacity, and confidence-selected Top-1 docking success. At under 5 Angstrom, broad continuation achieved the highest mean Top-1 success among trained policies, and focused or mixed supervision showed no consistent advantage.
Michael Brown, Maryam Rahimimovassagh, Mehdi Yazdani-Jahromi et al.
In briefValue-oriented policy distillation for interpretable wind farm power control
SummaryThis paper addresses wind farm power control under implicit balancing in Europe, where a wind farm adjusts its power setpoint using information from the Transmission System Operator to help restore system balance and reduce its own imbalance costs. The authors propose a policy distillation framework using explainable boosting machines, a variant of generalized additive models, to obtain a transparent control policy from reinforcement learning. They introduce a value-oriented perspective and expert knowledge into model design. Using real data from the Belgian grid and an offshore wind farm, the controller reduces imbalance costs by 82% and outperforms a price threshold controller and a decision tree distillation baseline.
Max Bruninx, Timothy Verstraeten, Jan Decuyper et al.
In briefFewStepRL: Using reinforcement learning to fine-tune fast, few-step diffusion models
SummaryThis paper presents FewStepRL, a framework that adapts reinforcement learning objectives designed for many-step diffusion models to few-step diffusion models. The method works in three phases: transformation, decomposition, and adaptation, which express each objective as a sum of self-regression toward a reference and reward augmentation. The authors implement few-step versions of FlowGRPO, AWM, DiffusionNFT, DiffusionOPSD, and ReFL in one modular codebase and compare them experimentally.
In briefTB-MathModel: A Verifiable Terminal Benchmark for Agentic Mathematical Modeling
SummaryThis paper introduces TB-MathModel, a benchmark of 18 complete contest-style mathematical modeling problems from CUMCM and MCM (2023–2025). Tasks are delivered in a containerized terminal, and agents return a single JSON answer scored against a reproduced outstanding-award paper on the numeric core of the final question. The benchmark uses direction-aware one-sided scoring and two-layer feasibility verification, combining programmatic checks with Lean 4 obligations, and scores answers against a panel of human outstanding papers.
In briefBeyond Quadratic Attention: Separable Gaussian Operators for ODE Vision Transformers
SummaryThis paper introduces SSoG, a spatial operator that replaces dense self-attention in Vision Transformers with a learned mixture of axis-factorized Gaussian kernels, reducing per-layer cost below quadratic in token count. With fixed kernels, SSoG is globally Lipschitz continuous, which guarantees a unique solution for the continuous-depth variant. On CIFAR-100 at 224×224, it improves Top-1 accuracy by 2.2 to 4.1 points while using 8.8% to 16.6% fewer parameters. At inputs up to 1024×1024, the continuous-depth architecture reaches up to 154% higher throughput and over 60% lower peak memory than dense attention.
Carlos Boned Riera, David Romero Sanchez, Oriol Ramos Terrades
In briefHow Users Scrutinize LLM Answers: A Real-World Study of Evidence, Context, and Challenges
SummaryThis paper analyzes 489,283 assistant-user response pairs from ShareChat and WildChat to study how features of LLM answers relate to whether users accept, check, or challenge them. Broad attribution language appears in 31.9% of qualifying responses, while explicit URLs appear in only 1.33%, revealing a gap between claimed grounding and inspectable support. The authors conclude that supporting verification requires attention to what evidence can be inspected, what question the user is resolving, and the interaction history.
In briefCan LLMs Reliably Spot Multi-Dimensional Hallucinations? A New Benchmark
SummaryThis paper introduces HaluCheck, a bilingual benchmark with over 42k instances covering factual hallucinations in text and faithfulness hallucinations in agent trajectories. It provides fine-grained, multi-level annotations for detection, localization, and correction. Experiments on fourteen LLMs show basic sample-level detection ability but weak fine-grained localization, correction, and agent hallucination detection. Thinking, web search, and supervised fine-tuning each give partial gains, but none fully solves the problem.
In brief“Blind Compliance”: Coercive Control in AI Jailbreaking
SummaryThis paper argues that current AI jailbreaking research focuses mainly on whether a model's output is harmful, overlooking the relational interaction between user and AI. Drawing on interpersonal violence scholarship, the authors propose a framework of coercive control categories and examine how these dynamics are embedded in modern AI training paradigms. The paper reframes AI safety and alignment as a surface where power and control dynamics are enacted, not merely a technical security problem.
In briefAgent-R²: Scaling Agents for Realistic Office Work Using Real-Source Data
SummaryThis paper introduces Agent-R², a framework that synthesizes training data for AI agents from real occupational seeds, workflow evidence in corpora, and authentic web-sourced files. It distills 4,000 task instances from 10,000 web-corpus candidates, each pairing real multi-format attachments with fine-grained, deliverable-level rubrics. The resulting Agent-R²-4B model scores 51.0% on GDPval and 14.8% on OfficeQA, surpassing Qwen3-235B-A22B-Thinking-2507 at 48.0% and 13.8%.
In briefProQuant: Progressive Quantization-aware Training for Edge-Deployable Multimodal Models
SummaryThis paper introduces ProQuant, a weight-activation quantization-aware training framework for low-bit multimodal large language models. It uses self-relabeling to align the finetuning data with the pretraining distribution and progressive block activation with a vision-sensitive factor to prioritize recovering multimodal understanding. Experiments on models from 2B to 8B parameters show it outperforms existing methods, with the W4A8 Qwen3-VL-2B-Instruct gaining 6.6% average accuracy and nearly matching full-precision performance with about a 1% drop.
WhyThe authors propose a progressive QAT framework for low-bit MLLMs that decouples multimodal understanding from language modeling. They report 6.6% average accuracy gains for W4A8 Qwen3-VL-2B across mu
In briefDo sEMG Foundation Models Help Generalise Across Subjects?
SummaryThis paper systematically evaluates three surface electromyography (sEMG) foundation models on three gesture classification datasets. It tests both intra-session and inter-subject splits and uses three fine-tuning strategies to isolate the effect of pretraining. The authors find that pretraining improves performance when training and test samples come from the same recording session, but this benefit largely disappears under inter-subject evaluation. In datasets with few subjects, pretraining gives little or no benefit, and classical machine learning baselines match or outperform the foundation models.
In briefSMART: A multi-agent method that reasons over implicit entity states to verify complex claims
SummaryThis paper proposes SMART, a multi-agent framework that captures the implicit states of entities in a problem to obtain knowledge needed for reasoning. It uses several large language models and a structured role-based prompting strategy to curate diverse, complementary, and solution-relevant premises. According to the paper, knowledge-guided reasoning with these real-world verified premises improves validation performance by up to 20% on complex scientific claims across three datasets. The authors also state that grounding premises in real-world knowledge is necessary and yields gains of up to 5%.
In briefLabel-efficient adaptation of geospatial foundation models for mapping building damage using positive-unlabeled learning
SummaryThis paper proposes PUMA, a framework that adapts geospatial foundation models to building damage mapping using annotations for only a small number of damaged buildings. It combines non-contrastive weakly supervised representation refinement with latent-state-based label disambiguation to reduce the impact of noisy supervision. Experiments on disaster benchmark datasets from around the world show its effectiveness and robustness, and a real-world application to the 2025 Los Angeles wildfires is also reported.
In briefMedHiVE: a hierarchical mixture-of-experts model linking 3D medical scans and reports
SummaryThis paper introduces MedHiVE, a framework that jointly models 3D medical scans and their reports. It combines a hierarchical vision-language Soft Mixture-of-Experts with contrastive learning to capture both fine-grained anatomy and global scan context. Pre-trained on over 25,000 CT scans with anatomical supervision, it outperforms 11 state-of-the-art methods across 6 datasets on segmentation, classification, and detection.
Gizem Mert, Szymon P{\l}otka, Pedro R. A. S. Bassi et al.
In briefBUMP: A brief accuracy peak when interpolating neural networks
SummaryThe paper reports that when linearly interpolating between permutation-aligned neural networks, test accuracy briefly rises near the endpoints before falling toward the midpoint, a effect the authors call BUMP. It appears across datasets, architectures and training regimes, and persists under perturbations such as early checkpoints and extreme pruning. Building on this, the authors propose weight-prediction mixed ensembling, which combines weight mixing with prediction-space mixing and outperforms standard logit ensembling in test accuracy with an improved calibration trade-off.
Melinda Fl\'ora Kiss, M\'arton Muntag, Adri\'an Csisz\'arik et al.
In briefKnowing When to Teach: Learning from Teachable Prefixes in Multilingual Reasoning
SummaryThis paper studies On-Policy Self-Distillation (OPSD), a method where a model learns from its own reasoning guided by reference solutions. It identifies "Teacher Drift," where the teacher follows the student's accumulated errors instead of correcting them, and finds this happens earlier in lower-resource languages. The authors propose Teachable On-Policy Self-Distillation (TOPSD), which uses teacher-student likelihood to locate where drift begins, selects a prefix for supervision, and distills only prefixes the teacher can verifiably correct. Experiments across ten languages and three model scales show TOPSD consistently outperforms conventional OPSD, especially in low- and medium-resource languages.
In briefBabyTheorist: A Benchmark for Learning Theories of the World from Observation Alone
SummaryThis paper introduces BabyTheorist, a benchmark for testing whether machines can learn reusable theories purely from observations and apply them to new situations. It generates continual streams of before-after visual observations from hidden compositional programs, without program labels or task identities during training. Seven representative baselines perform well below ceiling, and the gap widens as the curriculum becomes more complex.
In briefPCBRouteAgent: How Close LLM Agents Are to Becoming PCB Routing Engineers
SummaryThis paper presents PCBRouteAgent, a multimodal agent that automates printed circuit board routing. A global planner uses board images and electrical and geometric data to set connection topology, routing order, corridors, and layer transitions, while a detailed executor carries these out with KiCad's native routing engine and performs bounded local recovery. Evaluated on 247 synthetic and real-world boards from four PCBWorld-Bench splits, it achieves clean-pass rates of 100%, 94.9%, 100%, and 40%. On three real-world subsets, it outperforms the best baseline by 13.9, 36, and 10 percentage points.
In briefRegret bounds for a broker's profit in bilateral trade: learning the best posted prices
SummaryThis paper studies a broker who mediates trade between a seller and a buyer with private valuations, posting a price to each in every round and observing only whether the trade occurred. Within a regret minimization framework, it gives a polynomial-time algorithm for adversarial valuations with regret of order T to the 2/3 power. For i.i.d. valuations satisfying the monotone-hazard-rate property, it achieves regret of order square root of T, and for regular distributions it shows a minimax regret rate of order T to the 3/5 power.
WhyThe authors give regret bounds for profit maximization in bilateral trade, including a near-tight T^(2/3) adversarial rate and a separation between MHR and regular distributions.
Johannes Br\"ustle, Simone Di Gregorio, Federico Fusco et al.
In briefRouteAlign: Matching Task Needs to Model Profiles for Choosing LLM Services
SummaryRouteAlign is a method for choosing large language model services that uses task requirements to guide both how model performance is measured and how services are selected. Task requirements describe the capability demand and difficulty of a task, so they can be matched with corresponding performance measurements, and a shared scorer combines this evidence. A generation-cost preference then controls the choice among eligible services. Evaluations across diverse tasks and heterogeneous services report higher task quality than comparison decision backends, with lower generation cost than their strongest tested setting.
In briefChoosing patch length for attention-based forecasting, with a theory explaining why it matters
SummaryThis paper studies how to choose patch length in attention-based time series forecasters, where it has mostly been selected by trial and error. For linear state-space models, the authors derive an exact factorization of the forecasting coefficient matrix that separates forecast dynamics from filtering dynamics. Combined with estimation bounds, they propose a patch-selection criterion that balances representation and estimation errors under temporal dependence, and they state conditions under which balanced patching is justified.
In briefMaking Convolutional Networks Less Reliant on Background Context Using a New Training Loss
SummaryThis paper proposes a training framework to make convolutional neural networks more robust when objects appear in unfamiliar backgrounds at test time. The authors first introduce ContrastiveCAMs, a feature attribution method that gives more faithful attribution maps than existing approaches. They then build on it with Context-Regularized Cross-Entropy (CR-CE), a loss that suppresses the influence of contextual regions. The method is evaluated on Waterbirds, Spawrious, ImageNet/ImageNet-BG and Hard-ImageNet, with consistent gains in context-shift robustness reported.
In briefLa-JEPA: Compact Robot Control Policies Using Latent World Modeling
SummaryThe paper introduces La-JEPA, a compact world-action model whose policy transformer is trained from scratch in the feature space of a pretrained V-JEPA 2.1 encoder. It predicts future features alongside actions during training and generates only actions at inference. Among policies without embodied pretraining, it reports state-of-the-art results on RoboTwin 2.0 and LIBERO-Plus with far fewer parameters and faster inference than Fast-WAM.
Mateo Cl\'emente, George Z. Xue, Amin Abyaneh et al.
In briefInterpretable MRI-clinical data fusion for predicting stroke recovery after acute ischemic stroke
SummaryThis study examines whether day-1 multimodal MRI adds prognostic value for three-month functional outcome in acute ischemic stroke. Sixty-six patients' ADC, DWI, and FLAIR scans and clinical data were analyzed using pretrained ResNet features fused with clinical variables and evaluated with cross-validated linear support vector machines. FLAIR-clinical fusion performed best, followed by ADC-clinical fusion, suggesting compact imaging representations provide complementary prognostic information beyond clinical variables.
Sina Raeisadigh, Henning M\"uller, Abderrahmane Hedjoudje
In briefA reusable generative foundation model for multimodal tokamak tasks
SummaryTokamak experiments record waveforms, profiles, fields, and scalars used to reconstruct plasma, forecast its evolution, and estimate disruption risk. The authors extend the TokaMark benchmark to 20 MAST tasks in six groups and introduce GenToka, a generative model that jointly generates schema-selected targets using a shared conditional diffusion Transformer with signal-specific codecs. On MAST measurements, GenToka lowers macro-average normalized RMSE relative to TokaMind by 4.7% on Groups 1-4 and 17.7% on Group 5, and raises Group 6 macro-average window-level F1 by 8.4%.
In briefUsing satellite data to detect weather fronts: what reanalysis data adds
SummaryThis paper studies machine learning methods that detect weather fronts, the boundaries between air masses with different thermodynamic properties, pixel by pixel from geostationary infrared satellite observations. The authors compare a satellite-only model with one that also uses 28 ERA5 reanalysis fields, tested over the North Atlantic and Europe. The satellite-only model reaches a 150 km tolerance CSI of 0.402, rising to 0.575 with ERA5. Low-level winds are the hardest to predict from satellite data, and removing winds or moisture causes the largest drop in detection performance.
Ioannis Dravilas, Miruna Stoicescu, Roope Tervo et al.
In briefMCUFlowNet: Hardware-Aware Architecture Search for Optical Flow on Microcontrollers
SummaryThis paper designs dense optical flow networks that run on microcontrollers with neural accelerators, aiming to help small drones avoid obstacles. The authors propose PACE-NAS, a search framework that fixes the backbone and peak SRAM first, then searches operators and depths to trade accuracy against latency. The fastest resulting model, MCUFlowNet-S, runs 1.76x faster than a retrained EdgeFlowNet with 1.6% higher EPE, and reaches the goal without collision in 51.5% of simulated forest flights versus 41.8%.
In briefCatching silent faults where reinforcement learning environments meet the trainer
SummaryThe paper studies silent faults that occur when a robot reinforcement learning environment runs in a separate process, machine, or language from the trainer. The authors injected faults into the data crossing this boundary on Pendulum, HalfCheetah, Hopper and a 7-joint arm (Pusher). Learning-curve thresholds and comparisons with fault-free seeds caught very few of them. A step ledger kept on both sides of the boundary and compared step by step caught and located the faults.
In briefCross-Layer Prediction Method for Unsupervised Industrial Defect Detection
SummaryThis paper proposes a single-class anomaly detector for industrial images that is trained only on normal images from one product category. It treats the layers of a frozen vision transformer as a sequence of transitions, learns token-wise hidden-state prediction between blocks, and uses adjacent-layer cosine similarity to identify fault lines that set source and target roles. Image AUROC reaches 99.75 on MVTec-AD, 99.42 on VisA and 94.27 on Real-IAD.
In briefLinking image encoders with large language models to describe facial expressions in words
SummaryThis paper introduces a new task of describing facial expressions in natural language, capturing nuances beyond emotion categories or facial action units (AUs). The authors propose a three-stage training framework that trains a vision-to-language model using synthesized image-text pairs and BLIP-2 pre-training techniques. Captions are synthesized with GPT-3.5 from existing emotion or AU annotations to overcome the lack of training data. Experiments show the method generates captions describing facial actions, emotions, and their relationship, and that the approach also improves visual features for AU detection and emotion classification.
In briefHow Far Can Untrained Neural Networks Go in Anomaly Detection? From Tables to Pixels
SummaryThis paper proposes PRISM, which combines fixed random nonlinear features with Mahalanobis scoring against normal references for tabular anomaly detection and visual localization. On 47 ADBench tabular datasets, PRISM-T achieves the highest mean AUROC (86.4%). On MVTec AD and VisA across 27 visual categories, PRISM-V reaches mean image-level and pixel-level AUROCs of 90.4% and 96.4%. Both models work without neural weight optimization or pretrained representations.
In briefWho Can Evaluate AI Independently? A Framework for Analysing Evaluation Infrastructure
SummaryThis paper proposes PAAI (Power, Access, Adequacy, Independence), a framework for examining who controls the infrastructure behind AI evaluations, on what terms it is available, and how much autonomy it allows. The authors apply it to twelve African AI evaluation projects from 2022 to 2025 and to the roles of member institutes in three joint testing exercises of the international AI Safety Institutes network. They conclude that participating in an evaluation does not by itself confer evaluative authority, and that credible evaluation requires control over priorities, methods, publication and continuity.
Sumaya Nur Adan, Joanna Wiaterek, Nourhan Abdelrhim et al.
In briefBeyond K/V Caches: Transferring Model State Between Hybrid Linear-Attention LLMs
SummaryThis paper addresses the cost of switching between language models during long conversations, where the new model must re-process the accumulated context. Hybrid linear-attention models keep K/V caches along with recurrent states and convolution histories, which interact and complicate transfer between models. The authors propose Hybrid State Transfer (HST), which handles all three state types, uses structured recurrent maps, and calibrates adapters against the frozen target model's continuation outputs. On five benchmarks, transfers from Qwen3.5 4B and 9B to 27B retain 89.88% and 94.11% of task performance on average, and handoff speedups on synthetic 16K inputs reach about 21.66x and 21.91x.
