In briefMultiInpaintCOCO: A Multilingual Benchmark for Fine-Grained Visual Understanding
SummaryThis paper extends the English InpaintCOCO dataset, which focuses on fine-grained visual details, to 13 languages from five language families, producing MultiInpaintCOCO with 953 minimal-pair examples. The authors test several state-of-the-art multilingual multimodal retrieval models, including CLIP-like encoders and large language models. Results show performance drops substantially for lower-resource languages, and most models prefer English captions over correct non-English ones, with the degree of bias varying across architectures.
Andrei-Alexandru Manea, Jind\v{r}ich Libovick\'y, Philipp J. R\"osch
In briefLLMs rely on specialist tools to generalize to unseen MS/MS structure elucidation
SummaryThis paper tests whether a frontier language model can propose molecular structures directly from tandem mass spectrometry (MS/MS) spectra without candidate lists or reference libraries. The authors built a 100-analyte benchmark from MassSpecGym, split into in-distribution compounds that are well documented in the literature and out-of-distribution compounds that are essentially undocumented. Without tools, the model reached 52% Top-1 accuracy in-distribution but only 2% out-of-distribution; access to de novo structure generators raised out-of-distribution accuracy to as much as 38%, and that agent outperformed the standalone de novo models.
Gabriel Asher, Cailum M. K. Stienstra, Tomo Oga et al.
In briefCalibrated Multiaccuracy: Sample Complexity of Deterministic and Randomized Predictors
SummaryThis paper studies calibrated multiaccuracy, where a predictor is globally calibrated and multiaccurate with respect to a bounded test family G, meaning its residuals are uncorrelated with every test function. The authors determine its sample complexity, which differs from that of multicalibration, and give an upper bound achieved by a randomized predictor. For deterministic predictors they establish a tight rate, revealing a genuine sample-complexity separation between the two.
WhyThe authors establish tight sample complexity for calibrated multiaccuracy, separating it from multicalibration and showing a deterministic versus randomized predictor gap.
In briefAn axiomatic framework for reinforcement learning in systems where only some agents can be controlled
SummaryThis paper studies reinforcement learning in open multi-agent systems, where only some agents are controllable, such as in smart grids. It derives value function learning from the Shapley axioms (Efficiency, Symmetry, and Linearity) within the n-agent ad hoc teamwork (NAHT) setting, linking it to cooperative game theory and the Shapley value for Dec-POMDPs. Two generic algorithmic templates, Shapley Machines and Banzhaf Machines, are introduced and can be combined with standard reinforcement learning methods. The authors report improved learning efficiency and generalisation to unseen agent types on NAHT benchmark tasks.
In briefSQLoom: Set-Contrastive Learning for Text-to-SQL Syntax, Semantics, and Logic
SummaryThis paper presents SQLoom, a constraint-aware fine-tuning method for large language models that converts natural language questions into SQL queries. It supervises outputs at three levels (syntactic validity, semantic alignment with the question, and logical consistency across clauses) using a set-contrastive objective that favors correct output sets over minimally different incorrect ones. On Spider, the approach raises Qwen3-1.7B's execution accuracy on the development set from 62.9% zero-shot to 88.7%, and reaches 73.4% on BIRD development.
In briefLearning rewards that transfer across environments by actively choosing which dynamics to study
SummaryInverse reinforcement learning seeks to recover expert rewards that remain useful when the environment's dynamics change. The paper proposes a two-phase method that first cycles through available dynamics to get a reliable reward estimate, then actively selects dynamics to reduce uncertainty about the expert reward. The authors provide sample complexity guarantees on transfer that hold for any dynamics, with a rate independent of the number of available dynamics. Experiments show the method substantially reduces the number of expert demonstrations needed to learn a transferable reward.
Victor Villin, Till Freihaut, Andreas Schlaginhaufen et al.
In briefENGINE: A Zeroth-Order Method for Fine-Tuning Large Language Models
SummaryThe paper proposes ENGINE, a zeroth-order (ZO) method for memory-efficient fine-tuning of large language models. It alternates between a low-dimensional subspace and the full high-dimensional space, adapting variational multiscale theory from computational mechanics to dynamic optimization. On standard LLM benchmarks, the authors report higher accuracy and faster convergence than existing ZO methods, with up to a 2.61x speed-up in forward passes and no additional peak memory.
In briefTop-k Retrieval Leaves Memory Value Undetermined
SummaryThe paper asks which memory values an agent can infer from its retrieval log alone. It proves that under top-k retrieval, no memory value is determined however many episodes are logged. The proposed LEDGER method estimates only the determined Shapley values, with confidence intervals, and raises the fraction of determined memories on pooled HotpotQA from 0.001 to 0.797.
WhyThe authors prove top-k retrieval logs leave memory values undetermined and propose LEDGER, reporting a large rise in determined memories on HotpotQA.
In briefMesh-agnostic refinement of 3D human body shape latent encodings
SummaryThis paper proposes a framework for refining reconstructions from parametric human body models such as SMPL. It uses the geometric difference between an initial decoded surface and a target mesh, without needing vertex correspondences. The framework has three methods: Latent-VaRefine corrects the latent code using varifold-distance gradients, Vertex-VaRefine predicts vertex displacements in a canonical pose with a DiffusionNet backbone, and ElaOpt optimizes surface deformation under elastic regularization and normal-direction constraints.
Chaima Zouaouid, Jonathan Cerqueirab, Sylvain Arguillere et al.
In briefReVA-HPE: Relative 6DoF Head Pose Estimation Using Synthetic Training and Virtual Anchors
SummaryThis paper studies head pose estimation from a single face image by estimating the relative geometric change between a query image and an anchor image, extended to full 6DoF with translation. It reconstructs a personalized avatar from one unlabeled enrollment image and renders virtual anchors with known poses, using multi-anchor consensus for rotation and the real enrollment image for translation. The estimator is trained only on synthetic face renderings. On real benchmarks, the authors report state-of-the-art rotation accuracy on BIWI and competitive metric translation.
Vasiliki Vasileiou, Panagiotis P. Filntisis, Petros Maragos et al.
In briefSpecificity Bias and the Wikipedia Paradox: Explainability Analysis of AI Text Detectors Under Domain Shift
SummaryThis paper examines how well Transformer-based AI text detectors (DistilBERT, RoBERTa) hold up under domain shift and how they make decisions. The authors test them on a custom dataset of hyper-local news generated by open-source LLMs and analyze them with SHAP-based explainable AI. They report several biases and argue that purely statistical detection methods have systematic limits in highly structured human writing.
In briefProper scoring rules that respect graph structure for finding a single node
SummaryThis paper studies node selection, where the goal is to identify one unknown target node in a graph and the estimator outputs a probability distribution over nodes. The authors introduce two strictly proper, graph-aware scoring rules: the parameter-free Resistance score and the Diffusion score with adjustable coarse-graining resolution. Experiments on epidemic source detection and ImageNet classification show that these scores detect estimation-quality differences that classical metrics miss.
Lorenz Hilfiker, Nathan Cincera Brack, Martin Sterchi
In briefA human-verified AI workflow for labeling environmental sound recordings
SummaryThis paper describes a human-AI workflow used to annotate the Kenya Forest Sounds Dataset (KFSD), a working collection of 7,531 recordings across 21 categories. The fine-tuned Audio Spectrogram Transformer reached 89.36% accuracy for binary detection of person talking, below the predefined 95% threshold for autonomous labeling, so its predictions were used as suggestions and checked by a human. The verifier reviewed all 2,337 machine-labeled files and found no false positives, taking about 4.2 seconds per file versus 35 seconds for manual annotation. The hybrid workflow saved about 20.0 hours, or 41%, on the remaining queue.
In briefDeepWaveRL: Reinforcement Learning for Seismic Subsurface Velocity Estimation
SummaryThis paper proposes DeepWaveRL, a framework that applies reinforcement learning to Full Waveform Inversion (FWI), a geophysics method for estimating subsurface velocity from seismic measurements. The authors recast FWI as policy learning, where a policy network predicts velocity maps as actions and a forward operator is used only for simulation. Learning relies on a single global scalar reward based on data misfit, so no gradients from the forward operator are needed. Experiments show competitive performance against supervised and self-supervised baselines.
In briefThe Cost of Mismatch: Noise Amplification in Zeroth-Order RL
SummaryThis paper studies whether zeroth-order (ZO) optimization can reliably train large language models with reinforcement learning from verifiable rewards (RLVR). The authors show that gradient estimation reliability depends on whether the two finite-difference evaluations use matched data: mismatched on-policy RL amplifies noise as the perturbation scale shrinks, while the matched-data BOLT method cancels this noise. Building on this, they propose ZO-BOLT, which consistently outperforms the pretrained baseline across three model variants and two mathematical reasoning benchmarks while keeping memory use close to inference levels.
In briefRethinking how fairness is measured when groups and classes are more than two.
SummaryThis paper observes that existing fairness metrics usually assume sensitive features and output labels are binary, and that non-binary data is forced into binary form using a one-versus-the-rest approach. The authors discuss the limitations of current metrics and propose mathematically sound, generalized formulations for multi-group and multi-class settings. They also show that binarization strategies distort empirical fairness assessments, revealing the limits of binary fairness metrics in multi-class contexts.
Matteo Grazioso, Fabio Naccarella, Chiara Gallese et al.
In briefPoisson Super Learning for Survival and Competing Risks on Discretized Time
SummaryThis paper introduces the Poisson Super Learner, a scalable ensemble method for survival and competing risks prediction with right-censored data. It uses the link between piecewise-constant hazard models and the Poisson likelihood, training candidate hazard learners under a common Poisson loss and combining them with a Poisson meta learner. The ensemble estimates survival probabilities and cause-specific absolute risks, and is illustrated by predicting cardiovascular disease risk in type 1 diabetes patients using Steno Diabetes Center data from 2005 to 2022, supported by simulations. The methods are implemented in the R package poissonsuperlearner.
Gabriele Pittarello, Thomas Gerds, Pernille Falberg R{\o}nn et al.
