In briefUsing Weisfeiler-Leman features to choose algorithms for constraint optimisation
SummaryThe paper proposes an automated feature extraction method that converts problem instances into graphs and applies Weisfeiler-Leman graph kernels to create structural representations. Its main contribution is a cut-based representation (WLc) intended to capture structural partitions. The authors evaluated it on MiniZinc Challenge instances from 2023 to 2025 using SVMs, Random Forests, and MLPs. Cut-based features outperformed fzn2feat with SVMs, while results with RFs and MLPs were closer.
In briefA persistent accuracy ceiling in automated verbal deception detection
SummaryThis paper systematically reviews 289 reports and 6,136 classification models from 25 years of research, and meta-analyzes 3,653 models across 97 datasets. Pooled accuracy was 74.4% (95% CI: 71.2%-77.4%), with substantial heterogeneity. Accuracy depended more on methodological quality, such as ground truth, data source, class balance and evaluation procedure, than on model complexity, and embeddings and large language models did not improve predictive performance. The authors suggest a ceiling of 70-75% for automated verbal deception detection.
Riccardo Loconte, Jonas Festor, Zane Fatjanova et al.
In briefCredal Machine Learning for Safer, Risk-Averse Predictions
SummaryThis paper addresses machine learning settings where worst-case losses must be guarded against, proposing to minimize conditional value-at-risk (CVaR) instead of average loss. Because the learner is uncertain about the true loss distribution, the authors represent this epistemic uncertainty with credal sets, which are sets of probability distributions. They develop an efficient learner that outputs credal sets and a decision rule that maps each set to one predictive distribution for CVaR minimization. The authors report that the approach reliably avoids catastrophic decisions in classification, under distribution shift, and in reinforcement learning, with little loss in expected performance.
Timo L\"ohr, Paul Hofman, Maximilian Muschalik et al.
In briefUniversal Textual Teaching for LLMs: a method that transfers a teacher's advantage to weaker models via a reusable natural-language primer, without updating parameters
SummaryThe paper introduces Universal Textual Teaching (UTT), a framework that finds where a student LLM falls short of a teacher through paired evaluations and turns those gaps into a reusable natural-language artifact called a Primer, without updating model parameters. Through iterative multi-role interaction among Student, Prompter, Teacher, and Synthesizer, the Primer is refined. On KernelBench, the Student's accuracy rose from 9.4% to 48.6%, and on Omni-MATH-2 math reasoning from 27.6% to 51.7%. The authors report that a Primer generalizes to Students that did not take part in its synthesis.
WhyA new parameter-free distillation paradigm via textual Primers, with comparisons on math and code benchmarks and cross-model generalization evidence.
In briefA geometric way to make SAC's critics more cautious, for robot locomotion learning
SummaryThe paper proposes GeZo-SAC, which has each critic predict generators that define a zonotope, using its probed width as a pessimistic offset and the disagreement between critics to control how they are combined. On four MuJoCo-v5 locomotion benchmarks, it achieves the highest mean return on Ant-v5 and Hopper-v5 and is competitive on the others. The deployed policy is an unmodified SAC actor, since the geometric components are used only during training.
Panagiotis Roditis, Panagiotis P. Filntisis, Petros Maragos
In briefContourVLA: A closed-loop contour policy for segmenting objects from language descriptions
SummaryThis paper presents ContourVLA, which treats generalized referring expression segmentation as a closed-loop process where editable contours act as policy states updated by geometric actions. It introduces Evolution-Aware Semantic Scheduling and DECT-GRPO to adapt perception and jointly optimize grounding and contour actions. On gRefCOCO val, testA, and testB, gIoU improves over the strongest evaluated baselines by 8.7, 2.8, and 2.7 points respectively.
In briefVINCIE-NExT: transferring image editing skills to video via in-context demonstrations
SummaryThis paper presents VINCIE-NExT, a framework that transfers image editing capability to video editing using in-context visual demonstrations, reducing the need for large paired video editing datasets. It breaks video editing into a chain of sub-tasks (Video, Image, Image, Video) and introduces a position encoding that links image demonstrations and video frames in a shared spatial coordinate system. Experiments on OpenVE-Bench report state-of-the-art performance across diverse editing categories, and ablations confirm the contribution of each component.
In briefFew-step image generation by iterating directly in data space
SummaryFlow matching generative models need many network evaluations to sample, and existing few-step methods depend on a manually chosen timestep schedule. This paper introduces data-space iteration, in which a shared generator refines its prediction directly in data space without flow discretization. The method integrates with distribution matching distillation with minimal changes. On class-conditional ImageNet 256x256, it outperforms standard discretization baselines and matches or improves on schedule-searched variants.
In briefSCORE: a spectral method for estimating correlations in multivariate Gaussians
SummaryThis paper proposes SCORE, a framework for neural network prediction with high-dimensional Gaussian targets that approximates the covariance matrix efficiently and stably. It splits learning into marginal distributions and a structured correlation matrix learned in spectral space, and trains with a closed-form Gaussian kernel score. The authors evaluate it on time-series forecasting, monocular depth estimation, and spatial weather prediction, reporting improved performance at lower computational cost.
Christopher B\"ulte, Emil Partow, Astha Gupta et al.
In briefDVLA-RL++: Vision-language few-shot learning with reinforcement-learning gating for cleaner class prototypes
SummaryThis paper addresses few-shot learning, where models recognize new categories from few labeled images, and notes that text-image agreement can be distorted by irrelevant context. The authors propose DVLA-RL++, which adds complementary semantic purification to separate intrinsic from nuisance descriptions and counterfactual reinforcement-learning gating to learn semantic fusion strengths. The method is reported to reach state-of-the-art accuracy on standard, fine-grained, and cross-domain benchmarks, averaging a 1.4% gain over DVLA-RL.
In briefDifferentiable systematic resampling enabling gradient learning in variational particle filters
SummaryParticle filters use a discrete resampling step that blocks gradient-based learning in variational sequential Monte Carlo. The paper proposes Differentiable Systematic Resampling (DSR), a temperature-controlled relaxation of systematic resampling that keeps its ordered, banded structure while allowing full gradient flow. DSR converges to exact systematic resampling as the temperature goes to zero, with a proven pointwise exponential convergence rate for the bias. Experiments on stochastic dynamical systems and handwriting data report comparable or better filtering and dynamics learning, at much lower cost than optimal-transport-based alternatives.
In briefReusing successful latent reasoning to improve vision-language-action robot control
SummaryThis paper presents FLOWMEM, a unified vision-language-action model that stores successful latent reasoning as reusable experience. It dynamically retrieves and recomposes compatible latent fragments as the context evolves, then refines the resulting route with current visual and proprioceptive evidence before generating actions. On RoboMME and LIBERO-Plus, FLOWMEM achieves 48.0% and 77.3% success, outperforming memory-free policies by 1.7 and 4.1 percentage points.
