In briefEditJudge-Bench: Testing AI Judges of Image Edits with Synthetic Ground Truth
SummaryThis paper introduces EditJudge-Bench, a benchmark of 1,382 synthetic before-and-after image pairs rendered in Blender, with edits based on nine categories of real user requests and exact ground truth from known 3D scene parameters. It evaluates vision-language models as binary judges of whether an editing instruction was correctly carried out, across 25 judge configurations. The results show substantial reliability gaps, such as the VIEScore template reaching only 69.0% AUROC with GPT-4o, and spatial edits like rotation and shape proving difficult. Judge rankings correlate strongly with two real-image human-preference benchmarks.
In briefCan System 1 AI Decision Models Be Trusted? A Test of Their Confidence and Explanations
SummaryThis paper examines whether System 1 decision models, which return typed probability distributions in a single forward pass instead of generating text, can be trusted. The authors build an open, deterministic audit benchmark across calibration under perturbations, cross-question interference, and explanation faithfulness, testing five open-weight models from 0.8B to 27B parameters against generative baselines. Findings include that reported confidence is a closed-form dispersion measure rather than empirical accuracy, that bundling questions causes interference in smaller models, and that larger System 1 models are faster and better calibrated under syntactic shifts.
In briefUniHam: A Large-Scale SOC-Complete Hamiltonian Dataset and Benchmark for Materials
SummaryThis paper presents UniHam, a large-scale dataset and benchmark suite of over 100,000 DFT-computed complex-valued Hermitian Hamiltonians that include spin–orbit coupling. It covers 72 elements and diverse crystal geometries and symmetries. The authors benchmark two representative models under a standardized protocol, using metrics for reconstruction accuracy, out-of-distribution generalization, and downstream property prediction. The results show the benchmark can distinguish model capabilities and highlight the need for spectral-level evaluation.
In briefMARS: A Video Forgery Detector Combining Multiscale Feature Differences and Appearance Features
SummaryThis paper introduces MARS, a two-path video detector that pairs fixed spatial and temporal difference-of-Gaussians features with adapted appearance representations, with a clip-dependent gate combining their predictions. Under a shared protocol across seven datasets, MARS achieves the highest average AUROC of 91.43%, outperforming seven compared detectors. Trained only on facial manipulations, it reaches 93.99% AUROC on GenVideo and 75.50% on DF26.
WhyThe authors combine multiscale feature differences with adapted appearance features in a two-path detector, reporting results across seven datasets and transfer to generated videos.
In briefHow grouping translated text before speech synthesis changes timing comparisons between simultaneous translators
SummaryThe study compares two simultaneous speech translation systems on ten English-German talks and examines whether grouping translation updates before speech synthesis affects their measured completion times. The effect of grouping depends on the synthesizer used, and in some setups it reverses which translator appears faster. The authors conclude that timing comparisons must specify how text is fed into synthesis, not only which synthesizer is used.
In briefReady-to-use tabular foundation models as surrogates for materials discovery optimization
SummaryBayesian optimization is widely used in active learning for materials science, typically with Gaussian processes as surrogate models that require task-specific kernel choices. This paper evaluates two tabular foundation models, TabPFN and TabICL, as out-of-the-box surrogates on two standard materials science benchmarks, comparing them against baselines including Gaussian processes, random forests, and CatBoost. The authors find that these models are highly competitive across both benchmarks without task-specific training for active learning or Gaussian process kernel selection. The results motivate further development of such models as surrogates for active learning in materials discovery.
Klemens Fl\"oge, Simon Bing, Leo Grinsztajn et al.
In briefDetecting AI-written text in real student essays
SummaryThis paper studies how well AI-text detectors work on authentic student writing, using the PISA 2025 dataset of 15-year-olds in Germany writing English essays with and without a chatbot. Keystroke, clipboard and chat logs were used to trace each character to its origin. Two detectors separate AI from human text reliably at the whole-text level, but the feature-based one loses accuracy after a simple student-style rewriting prompt, and the embedding-based one performs worse on single sentences.
Nils-Jonathan Schaller, Thorben Jansen, Olaf K\"oller et al.
In briefAuDiR: Diagnosing Audio Reasoning via Rule and Percept Interventions
SummaryThis paper introduces AuDiR, a benchmark that diagnoses audio reasoning by controlling inputs: it progressively supplies a rule and textual acoustic observations, and also removes the audio. The benchmark has 201 items across 18 families from 9 sources. In a preliminary evaluation, accuracy rose sharply when acoustic observations were supplied as text, though the authors note limits on stronger failure-attribution claims.
In briefPerturb, Aggregate, Reconstruct: Multi-view anomaly detection via consensus and Fourier attention
SummaryThis paper proposes PAR, a reconstruction-based framework for detecting anomalies across multiple views in industrial inspection. It perturbs features and uses a consensus aggregation module to create view-consistent pseudo-anomalies, and assigns a dedicated decoder to each view with Fourier attention. Experiments on the Real-IAD dataset report that PAR outperforms state-of-the-art methods at sample, image, and pixel levels.
In briefUsing convolutional neural networks to pick sparse linear solvers
SummaryThis paper studies using convolutional neural networks to automatically choose a solver for sparse linear systems. The network takes spy plot images of matrix nonzero structure as input and is trained to predict log timing ratios between candidate solvers. It achieves strong prediction accuracy on held-out SuiteSparse matrices and on real matrices from high-performance simulation codes.
Artemis Pados, Alan Edelman, Emmanuel Lujan et al.
In briefPrediction with foundation models under a per-query cost budget
SummaryThis paper studies how to improve task predictions by querying several foundation models while keeping the average query cost per input within a budget. The first method queries the same models for every input, and the second varies the queried models across inputs, using Q-learning to choose which model to query next and when to stop. The authors give theoretical results on prediction accuracy and cost, and report that both methods outperform existing approaches in simulations and a multimodal laptop price application.
In briefPhrase-level speculative decoding for a frozen language model
SummarySpeculative decoding speeds up text generation by verifying more candidate tokens per forward pass. This paper changes the candidate unit from single tokens to one- to three-word phrases, predicted by a small MLP head that reads the frozen target model's hidden states. Candidates are mapped back to target tokens and verified in one pass, so output matches greedy decoding. On one A5000 GPU at batch size 1, the method reaches speedups of 2.391x on GSM8K, 2.444x on MATH-500 and 1.544x on MT-Bench.
In briefCausal discovery via Bayesian symbolic regression
SummaryThe paper treats causal discovery as a Bayesian symbolic regression problem, sampling closed-form expressions from a posterior over symbolic models for each target's candidate parents. Each candidate's posterior inclusion probability, the fraction of sampled expressions containing it, serves as the causal score. Evaluated on coupled nonlinear bivariate dynamics, the CauseMe benchmark suite, and a four-variable biogeoscience structural causal model, the method is competitive with and often outperforms specialized algorithms while yielding interpretable closed-form mechanisms.
Roger Guimer\`a, Kai-Hendrik Cohrs, Sergio Pedreira et al.
In briefV3DSAM: a dense SLAM system that verifies loop closures and tracks object instances consistently across views
SummaryV3DSAM is an open-vocabulary semantic dense SLAM system that combines geometry estimates from VGGT-SLAM 2.0 with semantic observations from SAM3. It verifies loop closures at both pixel and object levels using reprojection and 3D spatial correspondence, without accessing backbone-internal features. Object instances are defined by class-constrained 3D connectivity in a global coordinate system and receive global IDs without inter-frame identity propagation. In TUM RGB-D tests the verifier rejected all four false loop candidates, and ScanNet long-sequence experiments showed improved instance identity consistency.
In briefValuing Symbolic Graph Rules by What Neural Networks Fail to Learn
SummaryThis paper proposes a way to measure the value of symbolic rules, such as motif or ring counts, added to graph neural networks by the label information the trained network's representation has not captured. Rule design is treated as rule discovery: candidates are scored on out-of-fold residuals against calibrated permutation nulls with family-wise error control. On the ZINC dataset, the method recovers the hidden ring term of the target in three independent runs and outperforms an expert cycle-count rule.
In briefLongSCOP: Semantically Consistent Long Video Outpainting
SummaryThis paper addresses long video outpainting, which extends video content beyond its frame boundaries, an area where prior work has mostly focused on short clips. The proposed LongSCOP method combines self-forcing training, latent calibration, and future references to stabilize long-horizon generation, and uses a VLM-based agent to select informative reference frames that keep subjects, scenes, and backgrounds consistent. The authors also build a 62K-clip training dataset and introduce LVO-Bench, an expert-curated benchmark for long-horizon stability and semantic consistency. Experiments show improved generation stability and semantic coherence over existing approaches.
In briefMoSTE: A mixture-of-experts method for detecting deepfake videos
SummaryThis paper proposes MoSTE, a framework for detecting deepfakes in videos. It splits adaptation into parallel spatial and temporal expert groups and routes each token to the most relevant experts through unsupervised routing. Tests across cross-dataset and cross-manipulation benchmarks report a new state of the art in average generalization.
WhyThe authors report a spatio-temporal expert mixture with unsupervised token routing adapting CLIP to deepfake video detection, with state-of-the-art average generalization across cross-dataset benchma
Charbel Yahchouchi, Nabyl Quignon, Noemi Roggero et al.
In briefOccupancy models for biased landslide records: when separating occurrence from recording beats neutralisation
SummaryLandslide susceptibility models are usually trained on recorded sighting locations rather than true occurrences. This paper adapts occupancy-detection models from ecology, treating several inventories as imperfect observers of one event to separate occurrence from recording. It compares this approach with the standard neutralisation method using near-complete inventories from three earthquakes with simulated reporting bias. The occupancy model reduces road bias in most cases, but its advantage depends on the structure of the true landslide pattern.
In briefImproving Circuit Learning Through Parameter Sharing: A Recursive Method for Probabilistic Circuits
SummaryProbabilistic circuits are generative models that guarantee tractable probabilistic inference, but they tend to be large. The paper finds that models assuming independence can counterintuitively achieve better likelihoods than those that do not. It introduces recursive parameter sharing (RePA), which improves efficiency without assuming independence, and reports large reductions in time and memory use with competitive likelihoods. The approach is further generalized to non-patch-structured data (NoPA), showing similar efficiency gains on seven tabular datasets.
Charles Bricout, Gennaro Gala, Viacheslav Borovitskiy et al.
In briefSIMD or Not: Comparing RVV and CGRA SoCs for AI Non-Linear Functions
SummaryThe authors build two systems-on-chip, one with an RVV vector processor and one with an elastic CGRA, both paired with a matrix multiplication unit and a single-core CPU. They propose mappings of common feed-forward networks onto each design and compare them. The CGRA achieves up to 7.12x fewer clock cycles than RVV on the activation stage of simple point-wise functions such as SiLU. RVV reaches up to 9.8x over the CGRA on complex functions such as softmax.
In briefCan advanced automated research agents reproduce human research decisions in AI?
SummaryThis paper examines how closely research decisions and outputs from Auto Researchers, which are AI agent systems that carry out research tasks autonomously, match those of human AI researchers. The authors introduce AutoHRS, a stage-wise evaluation framework for measuring how faithfully these agents reconstruct human research artifacts. Using 152 ICLR oral papers, they generate 456 research trajectories and find that current Auto Researchers often fail to recover key research decisions and their resulting artifacts across multiple stages.
In briefUCB-min: a bandit method that adapts to unequal reward variance
SummaryThis paper addresses multi-armed bandit problems where each arm's reward variance is unknown and varies widely. It proposes UCB-min, which for each arm picks the tighter of a variance-independent and a variance-aware exploration bonus, without prior variance knowledge or reward normalization. The method is tested on synthetic data, an educational tutoring task, and the MovieLens and Steam group-recommendation benchmarks, where it consistently tracks the best base algorithm across variance regimes.
In briefdNaViT: A study on whether discrete codes can serve as general-purpose visual representations
SummaryThis paper presents dNaViT, a discrete native-resolution vision transformer that produces discrete visual primitives analogous to language tokens. It builds on a frozen pretrained encoder, using lightweight residual modulation and a dual-branch objective, and uses grouped product quantization to expand discrete capacity without lengthening spatial sequences. The authors report that dNaViT matches or exceeds task-specific baselines across 12 visual understanding benchmarks, 4 generation benchmarks, and an image editing benchmark.
In briefPhysics-informed neural networks with few labelled data compared to classical surrogates
SummaryThis paper examines whether physics-informed neural networks (PINNs) can make accurate predictions from few labelled observations by using governing equations and initial and boundary conditions. The authors propose hierarchical loss weighting (H-PINN) and a gradient-direction-aware extension (HO-PINN). They compare these with a vanilla PINN and classical regressors on a heat equation and three two-phase-flow benchmarks. At 20 and 50 labels, at least one PINN variant achieves the highest mean primary-target R² in most comparisons, at substantially higher fitting cost than the classical methods tested.
In briefTwo-phase fast simulation-based inference using differentiable simulators
SummaryThis paper introduces Rapid Simulation Based Inference (RSBI), a diffusion-based variational method for likelihood-free Bayesian inference. It first builds an initial surrogate likelihood from a distance measure computed with a differentiable simulator, yielding an efficient one-shot proposal distribution. A Schrödinger Bridge diffusion sampler then transports this proposal to the true posterior, corrected by a neural likelihood ratio. The method is evaluated on standard SBI benchmarks and on gravitational wave ring-down posterior estimation.
Uros Zivanovic, Rahul Srinivasan, Roberto Trotta et al.
In briefPath-level alignment for hierarchical multi-label text classification
SummaryThis paper addresses hierarchical multi-label text classification, where models must capture both structural dependencies among labels and fine-grained label semantics at different hierarchy levels. It uses complete root-to-node label paths as a shared unit and adds two auxiliary training objectives while keeping a standard multi-label classifier at inference. PathSimNCE aligns document representations with hierarchical paths using hierarchy-aware contrastive targets, and a complementary semantic objective uses LLM-generated, document-conditioned descriptions of positive paths. Experiments on three benchmarks show consistent improvements, with structural alignment especially helpful in deeper hierarchies and semantic alignment giving extra gains for fine-grained labels.
In briefOptimal Sample Rates for Adaptive Differentially Private k-PCA
SummaryThis paper studies differentially private stochastic PCA, which seeks a k-dimensional subspace capturing the leading variance directions from i.i.d. random matrices while protecting each sample's privacy. The authors present the first algorithm achieving optimal sample complexity in both the dimension d and the target dimension k, while keeping the adaptive noise and arbitrary-k properties of prior work. Unlike deflation-based methods, it updates the full d×k subspace jointly, which simplifies the algorithm, reduces hyperparameters, and improves computational efficiency.
In briefUnprompted Coordination via Shared Memory Among LLM Agents: A Replication and Extension
SummaryThis paper examines whether large language-model agents coordinate when they only share a key-value store and no prompt mentions other agents. Four agents each received an arithmetic task, with two holding complementary halves of one computation, and the study was replicated by two independently built implementations. Joint success occurred only through the shared store, and coordination depended on whether an agent wrote its intermediate result and checked the store again.
Pekka Abrahamsson, Muhammad Waseem, Esa Karjalainen et al.
In briefPhysOPSD: Using Code as Privileged Information for Physically Plausible Video Generation
SummaryThis paper addresses the weakness of modern video generators in producing physically plausible motion and complex interactions. It proposes PhysOPSD, which uses executable code as privileged physical context, and PhysOPSD-Engine, which converts real-world physical scenarios into verifiable simulation programs. The method combines dual-level physical privilege with controlled physical variations under shared rollout noise. The authors report that with only 3K samples it outperforms SFT and RL baselines while preserving general video generation quality.
Harold Haodong Chen, Rui Peng, Lingteng Qiu et al.
In briefA Method Linking Task Planning and Execution for Language-Directed Bimanual Robot Manipulation
SummaryThis paper presents a zero-shot vision-language framework that lets a bimanual robot plan tasks while jointly considering task goals and physical feasibility. Planning uses a structured scene representation combining semantic object relations with execution constraints, and organizes actions into parallel two-arm stages with pre- and post-conditions. A hierarchical validator checks these conditions under perceptual uncertainty and uses stage-level execution history to enable local recovery or targeted replanning. Experiments on the CENTAURO wheeled-legged humanoid across nine language-conditioned manipulation scenarios show higher task success than a baseline that does not account for execution.
Emanuele Cuzzocrea, Giuseppina Iannotti, Luca Rossini et al.
In briefBDBBQ: A Bengali benchmark for measuring social bias in AI language models, built with human and AI collaboration
SummaryThe paper introduces BDBBQ, a bilingual English and Bengali extension of BanglaBBQ with 61,312 question-answering instances across 12 social categories. It validates the dataset through structural checks, 220 statistical tests, and expert annotation by four annotators, reporting strong agreement. Evaluating eight large language models on 490,496 instances, the authors find that accuracy and stereotype-aligned bias are distinct dimensions of model behavior.
Provat Saha, Ashika Habib Khan, Tahseen Tanha et al.
In briefPTC-IoU: A CPU-friendly method for estimating detection quality in edge video analytics
SummaryEdge video analytics systems need to estimate detection quality at runtime without ground-truth labels, to flag frames that should be sent to a more powerful cloud model. The paper proposes Predictive Temporal Consistency IoU (PTC-IoU), which compacts detection box histories so that scoring runs deterministically on the CPU instead of the GPU. On the Nvidia Jetson Orin Nano, per-frame estimator latency falls from 10.3 ms for a neural approach to 2.2 ms, pipelined throughput rises by 13.6%, and estimator energy use drops by up to 95%. The authors report that final detection quality remains competitive across multiple datasets.
In briefAutoBuddy: Testing Always-On AI Agents That Find and Do Work Unprompted in a Simulated Office
SummaryThe paper introduces AutoBuddy-Bench, a simulator that models an organisation's employees, projects, business systems and event history as a temporal graph and renders it as realistic workspace documents. Ground-truth tasks are derived from the events that create obligations, and their outcomes are computed from the world state. A reference agent, AutoBuddy, is evaluated with different backbone LLMs, showing weaker models fail mainly at discovering tasks, intermediate models find tasks but often do not complete them, and the strongest models fail mainly during execution, especially business-system operations. Discovery precision is only weakly related to execution quality.
Shiqi Chen, Daniel Madrigal Diaz, Wei-Ning Chen et al.
In briefBenchmarking how personality tendencies shape large language model behavior beyond writing style.
SummaryThe paper argues that large language model outputs are shaped by underlying behavioral tendencies resembling human personality, in addition to explicit prompts and instructions. It introduces SPACE, a benchmark combining a controlled evaluation protocol with datasets for personality verification, capability assessment, and hierarchical situational tasks, to separate personality effects from capability changes and instruction effects. Experiments show that prompting can change generation style, but internal personality tendencies persist and are not easily overridden by opposing instructions, and they measurably affect task performance.
In briefSystematic review of AI clinical prediction models across neurosurgery and its subspecialties
SummaryThis systematic review and meta-analysis pools results from 1063 studies using AI to predict clinical outcomes in neurosurgery. Across 451 studies reporting AUC, pooled discrimination was 0.86, with performance highest for diagnostic tasks and lowest for post-operative tasks. Most studies relied on internal validation only, and 89 percent were rated at high risk of bias.
Prajwal Ghimire, Aswin Chari, Jonathan Shapey et al.
In briefAutoVLA-D: a driving VLA model that learns trajectory-geometry conditions from reasoning
SummaryThis paper addresses a weakness in autoregressive vision-language-action planning, where continuous trajectories as final outputs are poorly tied to driving instructions. It introduces TrajCond, a discrete trajectory-geometric condition learned through a two-stage training framework before continuous trajectory prediction, with the method named AutoVLA-D. Experiments on NAVSIM navtest, nuScenes, and NeuroNCAP show that discrete trajectories are weaker as direct outputs but effective as TrajCond, consistently improving continuous trajectory prediction on closed-loop and open-loop benchmarks.
In briefRebuilding hourly traffic on every Paris street from a sparse sensor network using a graph neural network.
SummaryAir quality models need hourly traffic data on every road, but permanent counters cover only a small share of the network. The researchers built a graph neural network that estimates hourly flows on all 32,163 Paris intra-muros road segments from the 1,573 segments (4.9%) equipped with counters. Eight-fold spatial cross-validation gave R² of 0.89 for interpolation at withheld counters and 0.74 for extrapolation to districts without counters.
Xavier Jurado (ICube), C{\'e}dric Wemmert (ICube, ICube-SDC) et al.
In briefAudioCodex: Streaming Autoencoders for Live Audio
SummaryThis paper introduces AudioCodex, a fully causal autoencoder for live generative audio that compresses stereo audio into compact continuous latent sequences at about 10 Hz. It combines a Mean Flow magnitude autoencoder with latent hypersphere projection, bounded-noise augmentation, representation alignment, and adversarial phase recovery. Ablation studies show it outperforms state-of-the-art methods in the generation-reconstruction trade-off while remaining strictly causal in encoding and decoding.
Marco Pasini, Javier Nistal, Cyran Aouameur et al.
In briefTANSAR: Editable, Traceable Memory for Correlated Keys with Rule-Checked Multi-Step Knowledge Editing
SummaryThe paper analyses correlation matrix memories, a simple associative storage scheme, when the stored keys are correlated, as they are with semantic encoders. It derives capacity bounds, characterises the side effects of deleting an association, and gives a bound on attribution. Building on this, it proposes TANSAR, which retrieves along a directed acyclic association graph, prunes invalid paths with logical rules, and injects verified results into a frozen language model. Experiments report that multi-hop editing contradictions fall from 25.1% to 0.5%, with edit transactions taking 13.6 ms on CPU and 2.37 ms on GPU.
Abdulhady Abas Abdulla, Jiyan Salim Mahmud, Fatemeh Daneshfar et al.
In briefTRIAD: A Multimodal Retrieval Benchmark for Complex, Evolving Requests
SummaryThis paper introduces TRIAD, a benchmark for testing whether multimodal retrieval systems can satisfy complex requests that span abstraction levels, combine multiple constraints, and change across dialogue turns. It has three tracks (abstraction, constraints, state) with six subsets of 1,000 queries each, scored with multi-positive mean average precision. Evaluating 20 configurations of 17 systems, the authors find that no system leads on all subsets and that multi-turn retrieval usually yields lower mAP.
In briefDForkTree: Faster Language Model Decoding with Fork-Aware Draft Trees
SummaryThis paper presents DForkTree, a training-free method that speeds up speculative decoding for language models. It adds n-gram statistics from the target model's outputs to the drafter's scores as pointwise mutual information, weights them more after a fork, and builds the optimal draft tree on the GPU. On Qwen3.5-4B and 9B, DForkTree is the fastest tree method on every benchmark, up to 41.5% faster than DFlash and 15.1% faster than DDTree.
In briefQCC-WM: A visual world model linking macro-level state changes to image generation
SummaryThis paper proposes QCC-WM, a visual world model based on Quotient-based Causal Construction for action-conditioned prediction. It groups local visual features into spatially corresponding macro-nodes using translational and rotational actions, and recursively predicts node content and motion under a shared reference measurement. On the US-VLA dataset, it achieves an ROI-MSE of 0.01861 for four-step prediction and reduces LPIPS by 48.59%, 41.68%, and 30.23% over the best baseline for four-, eight-, and sixteen-step predictions, respectively.
In briefUniOMA: A unified optimal-transport method for aligning multimodal robot sensor data
SummaryThis paper proposes UniOMA, a multimodal alignment method that adds a Gromov-Wasserstein distance-based regularization to existing contrastive losses. It uses a dynamically learned GW barycenter to reduce structural differences between modalities, targeting force and tactile signals with varying sample densities. The authors report improved structural alignment and downstream performance over state-of-the-art contrastive baselines on five robotic benchmarks, and note the regularization can be integrated into most InfoNCE-like objectives.
Xinrui Zu, Mark Hoogendoorn, Kevin Sebastian Luck et al.
In briefReVMap: Building Global Vector Maps by Fusing Local Maps with Connectivity Awareness
SummaryReVMap is a framework that builds global vectorized high-definition maps from local predictions, aiming to reduce the cost of manual labeling. It uses point-wise localization uncertainty to bring global map priors into local prediction and revises the global map through connectivity-guided, uncertainty-aware fusion. The two steps are iterated to improve both local and global map consistency. Experiments on nuScenes and Argoverse 2 report state-of-the-art results for local map prediction and global map construction.
In briefTMT: Adapting CLIP to infrared images by transferring modality knowledge guided by text.
SummaryThe paper notes that CLIP aligns images and text less reliably on infrared imagery, since its pre-training is dominated by visible-spectrum data. It proposes TMT, which adapts CLIP using only paired infrared and visible images, with text embedding differences guiding the modality transfer. The method adds Generative-Perceptual Co-Supervision that combines CLIP semantic alignment with diffusion noise-prediction feedback. On four source-disjoint infrared classification benchmarks, TMT reaches the highest mean zero-shot accuracy among compared methods, at 48.92%.
In briefEvidence or Marginal? How Neurons Combine Beliefs with Inputs
SummaryThe paper compares two output formulas for a neuron whose input is missing or known only through a belief q: p_CD = E_q[N]/E_q[D] and p_EU = E_q[N/D]. It interprets p_CD as Bayesian updating with the belief as virtual evidence and p_EU as installing the belief as a new marginal, following Jeffrey's rule. The authors prove this post hoc reading and illustrate it for up to eight inputs. In a pre-registered synthetic study, 5 of 15 hypotheses were supported.
In briefRethinking radar 3D perception: learning from raw radar spectra
SummaryThe paper argues that raw radar spectra can serve as an effective representation for 3D perception without handcrafted reductions, provided the learning objective separates structure from noise. It proposes self-supervised masked modeling with Leakage-aware Spectral MAE, using a corruption strategy inspired by spectral leakage. Treating Doppler as a primary spatial dimension substantially outperforms naive channel stacking, and the approach outperforms state-of-the-art radar detectors and point-cloud-based self-supervised methods.
In briefAMSchem: Automatically drawing schematics for analog and mixed-signal circuits using verifiable AI agents
SummaryThis paper presents AMSchem, a training-free framework that uses verifier-guided AI agents to turn analog and mixed-signal circuit netlists into readable schematics. It separates circuit-aware visual planning from executable schematic geometry and checks device, pin-net, and geometric correctness with deterministic validators. The authors also build AMSchem-Bench, a benchmark of 64 circuits across 12 types.
In briefCan machine learning models make real predictions, or do they only fit existing patterns?
SummaryThe paper asks whether the philosophical distinction between prediction and accommodation, based on whether evidence was used in construction, still applies when theories are replaced by machine learning models. The authors argue it does, but the boundary becomes blurry, since an output's status depends on how its input patterns relate to those used in training. They suggest most successful ML outputs have a strong accommodation character, and trace implications for confirmation, novelty, and scientific realism.
In briefTensor network models for learning nonlinear system behavior
SummaryThis paper introduces Tensor Network Kernel Machines, a model class that uses low-rank tensor networks to identify nonlinear systems efficiently. It explains the two defining properties of this model class for readers new to the topic. The paper also discusses how such models are optimized and regularized, along with recent developments and starting points for further study.
In briefHow many exposures a world model needs to learn unseen game rules at test time
SummaryThis paper asks how many observed occurrences a world model needs to predict new game mechanics in an unfamiliar environment. The authors introduce MechEval, which labels 69 recurring mechanics across 20 ARC-AGI-3 games, and build WARC, a 35M-parameter vision transformer pretrained on other games. On held-out games, WARC predicts four in five mechanics with at least 80% accuracy on the cells they change, and half reach that accuracy within about 80 exposures.
In briefReducing SM-level load imbalance in 3D Gaussian Splatting rendering on GPUs
SummaryThis paper examines how 3D Gaussian Splatting (3DGS) rendering on GPUs suffers from load imbalance across streaming multiprocessors (SMs), because some screen tiles carry far more work than others. The authors propose a software-defined scheduling method that uses a persistent kernel to process heavy tiles first and splits outlier tiles across multiple CTAs. The approach needs no retraining or specialized hardware and speeds up alpha blending by up to 2.93x on average per scene in a large urban scene.
In briefA modular framework for designing and comparing multi-agent reinforcement learning systems
SummaryThis paper addresses the fragmentation of multi-agent reinforcement learning (MARL) research by proposing a modular perspective for modeling, analyzing, and implementing MARL systems. It decomposes MARL systems into eight functional modules, including agents, environments, learning, and communication, and reviews representative approaches for each. The authors present MARLKIT, a framework with explicit, interchangeable modules, and experiments show it supports prototyping, comparing, reusing, and combining modules.
Tiziano Maisonhaute, Jean-Christophe Souli\'e, Fabien Michel
In briefMNN-Pocket-0.3B Technical Report: A Compact Language Model for On-Device Use
SummaryThis report presents MNN-Pocket-0.3B, a 300-million-parameter dense language model supporting context lengths up to 32K tokens, designed for efficient deployment on smartphones and smart glasses. Its architecture, training data, training pipeline, and inference system were co-designed, and it has an INT4 footprint of 173 MB with decoding speeds up to 466 tokens/s on iPad Pro M5. The authors report strong results in Chinese and English knowledge and reasoning, mathematics, coding, and instruction following. Open weights and MNN deployment packages are released.
In briefOptimal Speedup of Heavy-Ball Momentum over Gradient Descent on Smooth Strongly Convex Functions
SummaryThis paper studies how much faster the heavy-ball method can be than gradient descent on smooth strongly convex functions. The authors show that the best worst-case rate constant c(κ) tends to 8 in every dimension two or more, giving a rate of order 1-8/κ. The lower bound holds for every parameter choice, and the matching upper bound comes from an explicit tuning and a Lyapunov function verified in exact arithmetic. Optimally tuned heavy-ball is therefore asymptotically about four times faster than gradient descent on this class.
WhyThe authors derive a tight rate constant for heavy-ball on smooth strongly convex functions, showing c(κ)→8, with an explicit tuning and a Lyapunov proof.
In briefPACE-ECG: A patient-aware ECG pretraining method that learns transferable heart-signal representations.
SummaryThe paper presents PACE-ECG, an ECG pretraining framework that contrasts temporally neighbouring recordings from the same patient and adds clinically guided multitask supervision. The frozen representations were tested for predicting post-myocardial-infarction all-cause mortality and incident heart failure on MIMIC-IV and two external HEEDB cohorts. The model outperformed DeepECG-SSL in-domain and externally while using a 41-times smaller encoder and less pretraining data. The authors conclude that pretraining-objective design strongly affects transfer, and that in-domain performance alone does not establish clinical generalisability.
Riccardo Cavarra, Lupo Lovatelli, Shaheim Ogbomo-Harmitt et al.
In briefHumanSightBench: A Benchmark Testing Human-Level Visual Perception in Multimodal AI Models
SummaryThis paper presents HumanSightBench, a benchmark for testing how well multimodal large language models perceive visual content as people do. It contains 2,876 short-answer visual questions across 7 perceptual categories and 21 subcategories, filtered through five inclusion gates. The best pass@3 score among eight evaluated models is 65.6%, and none exceeds 27.5% on Counting. The authors argue these results reveal capability gaps that overall scores obscure.
In briefA formal framework for agent harnesses whose tools and policies can change during execution
SummaryThis paper introduces Reconfigurable Action Theories (RAT), a formalism in which actions can change the action theory itself, so that agent harnesses can alter their tools, providers, and policies while running. It gives conditions for a unique next configuration, preservation of existing behavior when modules are added, and order-independent module loading, and it analyzes the computational complexity of reasoning tasks. Experiments compare finite RAT models with the official DeepSeek Harness, Pi, and OpenHands implementations.
In briefPhysics-informed neural networks for finding damage in plates using ultrasonic guided waves
SummaryThis paper proposes a frequency-domain physics-informed sinusoidal representation network (PI-SIREN) that embeds the Mindlin-Reissner plate theory's governing equations into training to reconstruct ultrasonic wavefields from sparse measurements without temporal discretization. The model is trained in three stages: Adam, residual-adaptive refinement, and L-BFGS. In experiments, the approach detected and localized a milled notch from sparse measurements with a localization error of 1% of the plate diagonal, while data-driven benchmarks and standard MLPs failed to identify the damage.
In briefRecoGEM: A Serving System for Large-Scale Generative Recommendation
SummaryThis paper presents RecoGEM, a production serving system for generative recommendation, where models generate item identifiers step by step rather than scoring candidates. This makes online serving harder, so RecoGEM separates a dedicated runtime from the execution engine and optimizes for short token sequences and large candidate sets. The system has been deployed for over nine months, serving hundreds of millions of users. On OneModel production workloads, it reports over 1,000 QPS per instance, P99 latency below 35 ms, and 2.88 times the throughput of representative baselines.
In briefBe My Eye: a method for robots to see beyond their own viewpoint
SummaryThis paper proposes a task-oriented active context sharing framework for multi-robot collaboration. Instead of passively sending raw observations, robots compress task-relevant information into compact representations under limited communication budgets. Training has two stages: first learning the compressed representations, then integrating them into a pretrained Vision-Language-Action (VLA) policy for bidirectional exchange. Controlled experiments show substantial gains under constrained communication, and three real-world dual-robot tasks validate the approach.
In briefTreating Depth as a Source of Candidate Solutions: Budgeted Allocation for Recurrent Test-Time Scaling in LLMs
SummaryThis paper studies recurrent latent reasoning in large language models, where test-time computation is scaled by increasing recurrent depth. The authors find that deeper recurrence can lose correct solutions found at shallower depths, and that the sets of problems solved at different depths are largely non-nested, so depth also determines which candidate solutions are reachable. They propose Depth Portfolio Allocation (DPA), which allocates a shared inference budget across depth-specific candidate sources, with fixed and feedback-guided variants. Experiments on GSM8K, SVAMP, MultiArith, and MBPP report that mixed-depth portfolios increase oracle coverage and feedback-guided allocation matches or improves uniform allocation.
In briefHow deep speech models' internal representations match brain responses to speech in noise
SummaryThis paper examines how the internal representations of deep neural networks align with the human brain's process of moving from acoustic to linguistic speech processing. The authors use neural encoding models and representational similarity analysis to compare brain responses to acoustic and linguistic features with embeddings from a pre-trained automatic speech recognition model. Shallow layers align more with acoustic responses and deeper layers with linguistic ones, and fine-tuning on audiovisual noisy data strengthens this alignment.
In briefVLA-MOPD: Merging Robot Skill Experts via On-Policy Distillation
SummaryThis paper introduces VLA-MOPD, a protocol for consolidating several task-specific flow-matching vision-language-action robot policies into one shared student policy. The student collects robot rollouts and, at each visited observation and intermediate ODE state, queries the instruction-routed frozen teacher, learning actions through a masked local matching loss. Reported results include higher success on LIBERO, RoboTwin 2.0, and a Piper robot after consolidation.
In briefTraining a model to reason from context clues when forecasting time series
SummaryThis paper introduces CastReasoner, a model that forecasts time series by reasoning over contextual features. It uses a structured chain of thought to extract predictive evidence, infer the future dynamics that evidence suggests, and check these against the forecast prior. The model is trained in two stages, supervised fine-tuning and reinforcement learning, and the authors report superior forecasting performance across multi-domain benchmarks.
In briefModelling the socio-economic effects of the pandemic on a metropolitan area
SummaryThis paper extends the BTH-SEIRS model, a mathematical model previously proposed for predicting infections, to study how reduced mobility affects an urban area. The authors focus on the metropolitan area around Helsinki, Finland, using mobile phone mobility data, railway public transport data, and accommodation occupancy rates. The model is shown to reproduce real infection data and mobility-related economic activity trends.
Jan E. Snellman, Rafael A. Barrio, Cecilia I. Ventura et al.
In briefInterpretable time-to-event prediction via log-odds additive distributional regression
SummaryThis paper proposes ODDS-DR, a distributional regression method for time-to-event prediction in which output parameters are additive in log-odds, a measure common in clinical risk modeling. The design aims to make the predicted distribution parameters interpretable and to support time-dependent feature attribution. The authors validate the approach on two real-world clinical cohorts. They report that it matches state-of-the-art predictive performance while turning distributional outputs into interpretable risk markers.
In briefHierarchical multi-channel prediction of rover tip-overs using separate temporal models
SummaryThis paper proposes a hierarchical framework for predicting tip-overs of planetary rovers, using separate temporal models for different signals. A Temporal Convolutional Network learns gradual instability from inertial measurements, while a GRU-based dynamics model detects unexpected deviations through prediction residuals. The two signals are fused at the decision level to issue graded safety actions such as early warnings and emergency braking. Experiments on inertial measurement data from heterogeneous terrains report 97.58% event prediction coverage and a 1.82 s median lead time.
In briefA framework for testing AI models on crime scene 3D reconstruction
SummaryThis paper proposes a structured framework for evaluating egocentric foundation models in forensic settings and applies it to EFM3D, a model for 3D scene understanding. Four simulated indoor crime scenes were recorded with Meta Project Aria Gen 1 and assessed for room reconstruction, semantic understanding, computational performance, and forensic suitability. The authors conclude that EFM3D is best suited as a complementary tool for contextual documentation and scene review, not as a replacement for high-fidelity forensic capture.
Davidson Chisom Onwubiko, Shahana Bano, and Carlos Francisco Moreno-Garc\'ia
In briefRAIL: Teacher policies trained to be reproducible by students
SummaryThis paper introduces Representation-Aligned Imitation Learning (RAIL), a method for training teacher policies that a student can later imitate from raw sensory inputs. A teacher with privileged state information can create an imitation gap the student cannot close. RAIL uses contrastive learning to learn a latent representation shared by teacher and student observations, and trains the teacher within that space. Across multiple environments, the authors report it outperforms strong baselines and direct reinforcement learning from raw observations, without reward modification or student fine-tuning, and enables zero-shot transfer to new tasks.
Meraj Mammadov, Pedro Zuidberg Dos Martires, Johannes Andreas Stork
In briefA method for automatically discovering symbolic local learning rules
SummaryThis paper presents a general method for discovering symbolic local learning rules that could replace backpropagation in principle, applicable to any deep learning architecture. The method automatically finds layer-specific local loss functions, and it was tested on classification, reconstruction, and self-supervised learning tasks. The discovered rules create non-trivial interactions between layers and yield more interpretable results than MLP-based alternatives.
In briefCADMA: A capacity-aware method for checking how well generative models cover real data
SummaryThis paper introduces CADMA, a framework for evaluating how well generative models cover the real data distribution. It matches local capacity between real and generated samples to separate geometry-based coverage from effective coverage. Experiments show that CADMA gives more reliable coverage estimates and more detailed signals for diversity failures such as mode dropping and mode collapse than existing metrics.
In briefBalancing pseudo-label classes over time to stabilize online test-time adaptation
SummaryThis paper addresses instability in continual test-time adaptation, where models trained on unlabeled streaming data can drift or collapse as pseudo-labels concentrate on only a few classes. The authors propose SCALAR, a plug-and-play module that buffers recent predictions and uses the Sinkhorn algorithm to balance their class distribution toward a target. The balanced soft targets are then used for adaptation, with the model architecture unchanged and only about 10 MB of extra peak GPU memory on ImageNet-C. SCALAR can be combined with existing entropy- or diversity-based stabilization methods.
In briefPruning versus merging: compressing mixture-of-experts LLMs for multilingual production tasks
SummaryThis paper compares two one-shot expert compression methods, REAP (pruning) and REAM (merging), on five multilingual production tasks from a web-hosting company. It tests two 30B-parameter models with 25% and 50% of experts removed. Pruning performed better, winning 16 of 20 significant comparisons against REAM's 2. Most quality loss came from output-format failures such as invalid JSON, broken mobile layouts and runaway generation.
In briefECOS: environment-calibrated self-teaching method for language agents
SummaryThe paper proposes ECOS, an on-policy self-distillation method in which a language agent acts as its own teacher using skills written from its past trajectories. It uses environmental outcomes to estimate command-level advantages, which determine how much to trust the teacher's guidance and which self-written skills to reinforce. On ALFWorld and WebShop, 3B and 7B agents reach strong success rates without skills at inference, and on unseen ALFWorld tasks they outperform GRPO under the same protocol.
In briefPractice accuracy can hide the cost of worked-solution training data
SummaryThis paper studies a 2B-parameter reasoning model that learns two invented quantitative rules, comparing self-generated training targets made from a rule sheet alone with targets that also include a program-written worked solution. Both give identical forward-use accuracy on unseen values, but sheet-only targets score substantially higher on composition, which requires reusing the same rule across two segments. The worked-solution models also generate about 2.6 times as many tokens per composition success.
In briefPARIS: A multi-agent AI framework for automating systematic literature reviews
SummaryThe paper presents PARIS, an open-source framework in which a conductor agent coordinates 15 specialized large language model agents to carry out a systematic literature review from scoping a research question to exporting a manuscript draft. It introduces QualityScore, a 0-100 index that grades corpus quality before synthesis, and includes three human checkpoints. Two case studies show end-to-end feasibility and reveal failure modes such as incomplete retrieval and drifting counts.
Christophe Cruz (ICB, UBE), Hussam Ghanem (ICB et al.
In briefEmpirical Test of Quantum Kernel Stability for Network Intrusion Detection, Checking Whether Hardware Noise Is the Main Cause of Weak Performance
SummaryThis paper evaluates Pegasos-QSVC on the NSL-KDD and ToN-IoT datasets across 56 quantum experiments, comparing one noiseless exact statevector backend with three noise-simulated superconducting backends. XGBoost clearly outperforms the quantum model, which also varies much more across random seeds and produced several degenerate classifiers, two of them on the noiseless backend. The authors argue that average accuracy can hide unacceptable behavior, so worst-case results, seed-level variability and class-specific metrics should be reported.
Sreejonty Chakraborty, Adrita Rahman Tory, Muhammad Aminur Rahaman et al.
In briefA geometric invariant that predicts attention head specialization from model weights alone
SummaryThe paper derives a scalar invariant, ρ, that measures how strongly each attention head can couple to rotary positional encodings, allowing heads to be analyzed without forward passes. Across four bias-free RoPE models ranging from 135M to 7B parameters, ρ correlates with induction heads at Spearman values between -0.55 and -0.74. As a preliminary test, applying YaRN only to the top 50% of positional heads degraded perplexity less than applying it uniformly.
In briefGraph-based reinforcement learning for planning sequential meetings of patrol USVs
SummaryThis paper addresses a command unmanned surface vehicle that sequentially rendezvous with multiple patrolling USVs, aiming to minimize mission completion time and total tardiness against priority-dependent due dates. The authors combine grid-based path planning, a surrogate model for travel-time prediction, and a graph neural network policy trained with reinforcement learning over a feasible transition graph. Evaluation with grid-based path planning reports average objective reductions of 10.5%, 3.5%, and 1.5% over greedy, MLP-RL, and 2-step lookahead greedy baselines, respectively, at about 1/35 of the latter's computation time.
In briefA world-model method that learns to act by valuing predicted next states
SummaryThis paper proposes WM-PPO, which uses a world model to predict the next observation for each action the policy samples and has the critic value those predicted successors, providing auxiliary signals for policy and value learning under sparse terminal rewards. Success rates rise from 63.5% to 87.5% on ALFWorld and from 39.3% to 52.9% on ScienceWorld compared with PPO.
In briefTrajLift: Modeling Memory Recall Over Time with Latency-Modulated Heat Diffusion
SummaryThis paper presents TrajLift, a spectral framework that models verbal memory recall as timed trajectories on a hierarchical graph, with potential relevance to Alzheimer's disease assessment. Its core mechanism is latency-modulated heat diffusion on a semantic hierarchy, which converts each recalled word into a representation that captures both multi-scale structure and continuous retrieval timing. On synthetic benchmarks and real-world clinical data, TrajLift outperforms diverse baselines by jointly modeling timing and semantic structure, with larger gains when retrieval timing is more informative.
In briefDeception Archive: A Standardized Dataset Repository for Verbal Deception Detection Research
SummaryThis paper presents the Deception Archive, a curated repository of publicly available verbal deception datasets compiled through a large-scale systematic review and standardized for consistent use. The repository currently contains 41 datasets with 387,995 truthful or deceptive statements. By offering centralized access to standardized data and metadata, it aims to support dataset discovery, cross-dataset validation, and reproducible evaluation of computational models.
Caterina Borgese, Riccardo Loconte, Lucca Pfr\"under et al.
In briefUsing Vision-Language Models for Federated Unsupervised Visual Recognition
SummaryThis paper introduces FedUR, the first federated learning framework for unsupervised visual recognition, using the zero-shot capabilities of pre-trained vision-language models such as CLIP. It combines a class-aware pseudo-labeling strategy with class weights and negative pseudo-labels to handle data heterogeneity. It also aggregates general visual prompts on the server while keeping personalized text prompts on clients. Experiments report consistent gains over baselines under both label-level and feature-level distribution shifts.
In briefNearFormer: Linear-size sparse attention using randomized near neighbors for graph transformers
SummaryThis paper presents NearFormer, a graph transformer that limits its global attention to a randomized near-neighbor graph instead of all node pairs. Each node connects to a small number of near neighbors chosen from a broader neighborhood, so the attention size grows linearly with the graph. Experiments on long-range graph benchmarks and synthetic structural tasks show a favorable trade-off between accuracy and scalability compared with local message passing and dense or sparse attention.
In briefBlind Variation, Selective Retention: Improving Creative Idea Generation in AI Scientists
SummaryThis paper proposes BVSR, a discovery framework inspired by the cognitive model of Blind Variation and Selective Retention. It generates research proposals under deliberately limited task information to broaden exploration, then selects candidates using outcomes of previously verified proposals on the same task. Across five proposer models and 12 scientific tasks, the method increases originality and non-obviousness of proposals and yields stronger discoveries when integrated into existing AI-scientist workflows.
In briefTS-GEAR: A Generative Ad Retrieval Framework for Taobao Shangou
SummaryThis paper addresses on-demand delivery platforms, where merchants outside a user's delivery range are invalid candidates, by proposing TS-GEAR, a unified generative retrieval framework. It combines Geo-Semantic IDs that make reachability a prediction target, Query-Geo Target Attention over historical behaviors, and GMPO, a policy optimization method using deduplicated group rewards. Deployed in Taobao Shangou's search advertising pipeline with over 150 million daily queries, it outperforms prior methods on HitRate and geographic validity, with 3.8% more ad impressions and 4.2% more ad revenue.
WhyThe authors embed geographic reachability into generative semantic IDs and add set-level deduplication, reporting HitRate gains and ad revenue lift in Taobao Shangou deployment.
In briefReplication Study Finds No Gain from Thermal Comfort Cohort Models
SummaryThis paper replicates a study of cohort comfort models, which group building occupants by demographics, psychometric traits, or similar thermal preferences to reduce the individual data needed by personal comfort models. The authors assessed the original study's methods and replicated it on two datasets, DORN and PersonAI. Under F1-micro, a baseline that always predicts the majority class matched the cohort models, and under F1-macro, all cohort models scored below a random baseline based on the preference distribution. The authors conclude that the original claim that cohort models improve thermal preference prediction is not supported.
Thomas Hirsch, Katja Ceplak, Veronika Windisch et al.
In briefEmbodied First Exam: A Benchmark for Coding Agents Controlling Robots
SummaryThis paper introduces the Embodied First Exam, a unified evaluation framework testing whether coding agents can reliably control diverse robots under physical-world constraints. It includes 1,156 tasks across 35 suites on 47 robots, where agents may use direct control, code-based control, or policy training. In the Standard Setting that mimics the real world, the strongest of eight agents achieves a 34% success rate on 456 core tasks.
In briefGeoMind: A Benchmark Testing Spatial Understanding in Robot Manipulation
SummaryThis paper introduces GeoMind, a benchmark for evaluating spatial understanding through tabletop manipulation tasks, comprising 54 task variants centered on spatial reasoning and spatial memory. It also presents a scalable pipeline that automatically generates and annotates task instances with verified targets and executable interactions. Evaluations across diverse embodied agents show consistent weaknesses, particularly in tracking object identities after spatial changes and converting relational or geometric constraints into executable actions.
In briefTransactional Data Change for Real-Time Analytics in Lakehouses
SummaryThis paper proposes Retina, a data synchronization framework for propagating data changes to lakehouses in real time while preserving ACID guarantees. It aims to keep query performance and resource use efficient despite lakehouses' immutable file storage. The authors implemented and open-sourced Retina. Their evaluations report data freshness at the 10-millisecond level and linearly scalable throughput above 3.2 million rows per second, outperforming existing lakehouse systems and lake-compatible HTAP databases.
In briefFraudBench: A Legal Evaluation of AI Deception
SummaryThis paper introduces FraudBench, a benchmark testing whether large language models recommend, assist with, or directly take actions that would constitute fraud under U.S. law. It uses 159 scenarios based on real U.S. court cases, validated by legal experts, tested across advising, drafting, and acting tasks at increasing autonomy. The evaluation covers 20 LLMs under 25 conditions and includes a preregistered human baseline.
WhyThe authors build a fraud benchmark from real U.S. court cases across three autonomy levels, reporting 20 LLMs, over 300,000 generations, and a preregistered human baseline.
Kevin Wei, Sumaya Nur Adan, Stephan Llerena et al.
In briefTraining a code locator using a fixer model's likelihood scores
SummaryThis paper proposes Fixer-Guided Localization (FGL), in which a code localizer suggests several candidate locations and a frozen fixer model scores each by how likely it is to produce a verified patch. The localizer is then trained on these scores with weighted supervised fine-tuning, without running tests or using reinforcement learning. The full pipeline is built on Gemma 4 E4B using 1,629 test-verified trajectories.
Khalil Hennara, Muhammad Hreden, Zeina Aldallal et al.
In briefSame physics, different gradients: a survey of differentiable physics simulation
SummaryThis survey reviews differentiable physics simulation, where derivatives are propagated through a simulator to enable gradient-based optimization. It separates the object a returned derivative refers to (the finite computation, the converged equation, or the continuous dynamics) from how the backward pass is organized, and classifies methods and engines accordingly. The paper also measures four mechanisms by which useful gradients are lost on one testbed and reviews evidence that zeroth-order methods can outperform simulator gradients in those regimes.
In briefBeyond Flat Walks: A Hierarchical Approach to Autoregressive Graph Generation
SummaryAutoregressive graph generators usually linearize graphs with flat node-edge walks, which can hide higher-order structure in large, complex graphs. The paper introduces a hierarchical abstraction and coarse-to-fine generation pipeline that separates coarsening, tokenization, and decoding, and studies community coarsening, hierarchical tokenization, and order randomization. Across five molecular and five non-molecular benchmarks, hierarchical serialization gives consistent distributional and structural gains on all molecular datasets, while non-molecular gains depend on how well the supplied hierarchy matches the target distribution's structure.
In briefBalancing feature learning and mode collapse in dynamical-systems generative models
SummaryThis paper studies the learning dynamics of generative models based on dynamical systems, such as normalising flows and diffusion models. The authors use tools from dynamical systems and operator theory to express the gradient flow of a general class of stochastic differential equations as a composition of linear operators, linking collapse to low-rank biases. In variational inference and flow-matching tasks, they show this bias appears through two mechanisms, pointing to a trade-off between feature learning and collapse.
Julian Brandon, Bruno Loureiro, N Alex Cayco-Gajic et al.
In briefEgoLabel: Zero-shot first-person video auto-labeling with cross-model consistency checks
SummaryThis paper presents EgoLabel, a zero-shot pipeline that combines existing models to automatically label egocentric (first-person) data for depth, camera pose, objects, and hand pose without target-domain labeled training. It adds explicit cross-modal registration and consistency diagnostics. Across 20 sequences, per-hand depth anchoring cuts the mean of per-sequence median hand-depth residuals by 23% versus a single global scale. Hand PA-MPJPE is 14.53±1.29 mm, while absolute MPJPE is 119.1±31.3 mm.
In briefDisasterVL: A Dataset and Model for Evidence-Based Visual Reasoning About Disaster Scenes
SummaryThis paper introduces DisasterVL, a dataset and benchmark for reasoning about disaster imagery with answers grounded in visible evidence. It contains about 1.2 million images from 142 countries and 25 million structured annotations, including captions, incident and location assessments, spatial relationships, and visual question answering. On the test benchmark, the 9B DisasterVLM model outperforms the evaluated general-purpose baselines on captioning, open-ended VQA, incident reasoning, and region-level grounding. The authors plan to publicly release the dataset, trained models, and code.
In briefCan Time-Series Foundation Models Forecast Index-Futures Returns? A Pre-registered Test
SummaryThe paper tests whether frozen time-series foundation models forecast intraday returns on Nasdaq-100 and S&P 500 E-mini futures better than a no-change forecast. In a blind, pre-registered evaluation, none of 160 model, input and horizon combinations beat the no-change baseline. The apparent forecasting skill is attributed to standard metrics and a simple mean-reversion rule rather than genuine predictive power.
In briefVisible Tails: Adapting to Unknown Moments in Heavy-Tailed Bandits
SummaryThis paper studies heavy-tailed multi-armed bandits in which the learner does not know the reward moment parameters. It characterizes the optimal-arm distributions that allow adaptation to unknown moments at constant-factor cost, through a condition called variance-core visibility. A quantile-trimmed UCB index that knows the visibility constants is shown to attain the central-moment benchmark on this class.
WhyThe authors characterize variance-core visibility, a condition enabling adaptation to unknown moments, and propose a quantile-trimmed UCB index. They report matching the central-moment benchmark and a
