In briefOne for All, All for One: Coordinated Multi-Agent Diffusion Steering via Stochastic Optimal Control
SummaryThis paper introduces CMDS, a framework that reuses independently trained, frozen diffusion models as generative components and learns only a control that coordinates their reverse processes. Coordination is formulated as a stochastic optimal control problem that balances an assembly-level reward against deviation from pretrained dynamics. The learned control can be reused across task instances, and experiments show it can recover a target distribution, satisfy different spatial constraints, and recover individual sources from degraded mixtures. The method is evaluated on multi-agent maze navigation, articulated robot planning, and text-conditioned human motion.
Riccardo Barbano, Vincent Pauline, Runchang Li et al.
In briefDetecting Several Visual Bugs at Once in Game Videos with Deep Learning
SummaryThis paper proposes ResNet-BiLSTM, a deep learning model that detects multiple visual bugs in the same video game frame, and compares it with video classification models such as Inflated 3D ConvNet and 3D ResNet. The model reaches an F1 score of 85.78% on the benchmark dataset, and the authors report that modelling temporal dependencies helps video-based bug detection. They also introduce a new dataset of 77,969 video clips with about 1.2 million frames, where each frame may contain combinations of 5 bug classes.
In briefCNet: A Complex-Valued Deep Learning Framework with Wirtinger Autodifferentiation
SummaryCNet is a C++/CUDA framework for complex-valued neural networks trained by gradient descent using explicit Wirtinger derivatives. The paper presents three studies: a fully complex causal language model, radio-modulation classification using a learnable cyclostationary operator, and an analysis of the Fourier phase problem in coherent-diffraction imaging. It introduces Born-rule attention, which replaces softmax with squared-magnitude quantum-measurement scores, and reports results matching softmax self-attention on character-level language modeling.
In briefMINDSET: An energy-based memory method that updates long-term knowledge for conversational agents
SummaryThis paper introduces MINDSET, a memory controller for long conversational agents. It stores conversations as immutable episodes and organizes them into versioned schemas through minimum-energy state transitions. On 850 evaluation questions, it achieved the highest observed LoCoMo answer F1 and improved retrieval ranking over LightMem.
Sujato Dutta, Sreekruthy Tummala, Shashank Vanga et al.
In briefLarge language models are vulnerable to stray details in clinical notes and reasoning
SummaryThis paper tests how sensitive large language models are to information from a patient visit that is not clinically relevant, in both note-writing and clinical reasoning. Across 576 patient-clinician dialogues, frontier models inserted small talk into 35% of notes, and in 3.7% of those notes they misattributed it or used it clinically. In 57 mock consultations, background speech from another encounter leaked into 48.2% of transcripts, contaminating 5.3% of downstream notes from four open-weight models. The authors propose a dual encoding hypothesis and recommend testing resistance to incidental information before clinical use.
Krithik Vishwanath, Brandon Ye, Anton Alyakin et al.
In briefPhysics-guided surrogate model for predicting supercavity shape behind cone-shaped cavitators
SummaryThis paper proposes a physics-guided multi-fidelity surrogate model (PG-EHK) to predict the maximum length, maximum radius, and profile of natural supercavities behind conical cavitators. It fuses several analytical cavity formulas as low-fidelity trend functions, corrected with sparse high-fidelity CFD data. The surrogate uses fewer than 20 high-fidelity evaluations. Cross-validation on five hold-out CFD cases shows a 61% reduction in prediction error versus the best single analytical approach.
In briefRandom Fourier-feature Gaussian process attention: linear-time attention with calibrated uncertainty
SummaryThe paper proposes RFF-GPA, a plug-and-play attention module that treats attention as a Gaussian process with a stationary kernel approximated by random Fourier features. This low-rank approximation makes computing the posterior mean and variance scale linearly with sequence length. Experiments on several real-world datasets report improved calibration with maintained predictive accuracy and linear computational complexity.
In briefHow the Way Models Learn Affects Their Unlearning
SummaryThis paper studies how training with memorization versus generalization shapes a model's later unlearning. Using grokking in modular addition to separate the two kinds of models, and a bucketed variant to control their contributions, the authors find that generalization-heavy models suffer greater retain-set damage after unlearning. The same pattern holds for LLM unlearning of verbatim and factual recall.[]
In briefFedDermaSeg: a federated learning method for segmenting skin lesions in images without pooling patient data.
SummarySkin lesion segmentation is a key step in computer-aided skin cancer diagnosis. This study investigates federated learning as a privacy-preserving alternative to centralized training, where medical images are not gathered on one server. The authors simulate a distributed setting using the ISIC 2018 training and validation sets, then evaluate the model on the ISIC 2018 test set and the PH2 dataset. The federated model performs comparably to centralized training and consistently outperforms locally trained models.
In briefSparse2comm: A Robust Cooperative 3D Object Detection Method
SummaryThis paper proposes Sparse2comm, a framework for cooperative 3D object detection among vehicles and roadside infrastructure in autonomous driving. It addresses coupled degradations from limited bandwidth, packet loss, transmission delay, and spatial misalignment by reconstructing dense features from sparse foreground transmissions, then compensating delays and calibrating spatial offsets before fusion. Experiments on DAIR-V2X, OpenV2V, and V2V4Real report improved mixed-degradation AP over the Where2comm baseline.
In briefLess Is More: Leakage-Controlled Study of Skin Lesion Preprocessing with YOLO26
SummaryThis study evaluates dermoscopic image preprocessing and augmentation for jointly classifying and segmenting skin lesions with a small YOLO26 segmentation model. Using the HAM10000 dataset with lesion-disjoint splits to prevent data leakage, it compares minimal processing with online augmentation against more complex preprocessing. Minimally processed images with standard online augmentation gave a better accuracy-efficiency trade-off, while complex preprocessing showed no consistent joint benefit across three seeds.
Truong Viet Vu, Nguyen Chi Hai, Nguyen Phuc Nguyen et al.
In briefSingular value decomposition derived geometrically, showing how its proofs become machine learning algorithms.
SummaryThe paper re-derives the singular value decomposition geometrically, starting from how a linear map turns the unit circle into an ellipse and asking which input directions give the ellipse's axes. It builds the decomposition by maximizing stretch and recursing across dimensions, proving the spectral theorem along the way. The author then links these constructions to the power method and PageRank, gradient descent's stopping rule, and the duality behind kernel PCA.
In briefValid for Free: Homophily-Gated Conformal Prediction for Training-Free Graph Node Classification
SummaryThis paper studies whether a tabular foundation model (TabICL) can classify graph nodes without training on the target graph by reading node and neighborhood features as table rows. The authors show that a frozen predictor makes split conformal prediction exactly valid in finite samples, and that TabICL has lower expected calibration error than GCN with temperature scaling on nine of ten graphs. They also propose HG-DAPS, which gates on in-context labels and reduces prediction set sizes on six homophilous graphs. A pre-registered trap case shows that gating on raw homophily lowers coverage for low-homophily nodes while marginal coverage stays at the nominal 0.90.
Nguyen Duy Long, Phung Minh Hien, Nguyen Trong Viet et al.
In briefAdaptive Power Sampling: Tuning How Sharply LLMs Sample Reasoning Answers Per Query
SummaryThis paper studies power sampling, a training-free way to improve LLM reasoning by sampling from a sharpened version of a base model's output distribution. Existing methods apply the same sharpening to every query, which ignores differences in difficulty. The authors propose Adaptive Power Sampling (APS), which sets the sharpening exponent for each query at test time using answer agreement and the model's self-reward. Experiments on MATH500, HumanEval, and GPQA show APS consistently beats fixed-exponent power sampling without extra training.
In briefReinforcement Learning for Hierarchical Reasoning Rewards with Minimax-Optimal Transformers
SummaryThe paper studies theory behind reinforcement learning post-training of language models on reasoning tasks, modeling the reward as a hierarchical function over the response space. It proposes a Transformer-based actor-critic algorithm that achieves minimax optimal rates in query budget and regularization strength up to logarithmic factors. The authors also prove that sampling from a fixed reference distribution, as in offline reward modeling, can limit regret decay to a logarithmic rate, showing on-policy exploration zooms in on high-reward regions.
WhyProves minimax-optimal regret for on-policy Transformer actor-critic under hierarchical rewards, versus logarithmic decay for offline sampling.
In briefHave I Seen Enough? Frozen Video Models Track Evidence Readiness
SummaryStreaming video-language models must judge whether the evidence for the current question has arrived. The paper finds that unmodified frozen models already encode this readiness signal in a linearly readable form, labelled from timestamped evidence rather than model output. A timing policy built on this readout, called Readiness Gating, improves accuracy at matched video duration, with gains that vary with the task's accuracy headroom.
WhyShows frozen VideoLLMs linearly encode evidence readiness, question-conditioned, validated across seven models and 26 configurations with consistent gains.
In briefIn LLMs, the bias direction used for steering mainly reflects model confidence, not fairness
SummaryThis paper examines what the linear debiasing direction used in activation steering actually encodes in large language models. The authors find the direction is dominated by model confidence, pointing from high- to low-probability token regions, rather than a meaningful representation of bias. Steering along it lowers measured bias mainly by reducing confidence, which on QA benchmarks makes the model abstain from answering and improves fairness metrics as a side effect.
Stephanie Buttigieg, Maeve Madigan, Parameswaran Kamalaruban et al.
In briefSystematic review of human-centered AI safety for youth
SummaryThis paper systematically reviews 100 empirical HCI studies involving children and youth who interact with or are exposed to AI in schools, homes, care settings, and public services. Using the YAIR risk taxonomy and the MIT Mitigation Taxonomy, it maps which risks have been identified, which countermeasures address them, and whether those countermeasures are implemented and evaluated. Most risks are matched only with proposed countermeasures, few are implemented, and fewer are evaluated, with existing evaluations often measuring technical performance rather than protection from harm. The authors identify gaps in coverage and untested safeguards, and propose research directions to strengthen youth AI safety.
In briefDeltaTTT: layerwise optimization for nonlinear recurrent memory networks
SummaryThe paper finds that sequential test-time training, which updates a memory network based on its previous state, does not consistently help nonlinear memories, and a fixed-base parallel baseline can outperform it. The authors attribute this to nonlinear memories being harder to optimize in a single pass over a sequence. They propose DeltaTTT, which replaces joint optimization of a two-layer memory with layerwise learning, giving each layer a local target and a state-dependent delta-rule update while keeping a nonlinear readout and enabling chunkwise parallel computation. Experiments on DeltaNet and LaCT backbones report gains over recurrent baselines in language modeling and retrieval.
In briefAnyBottle: A Method That Keeps Only the Concepts Really Needed
SummaryConcept bottleneck models make predictions by routing them through human-interpretable concepts, but traditionally required concept annotations. Annotation-free versions usually rely on large, fixed concept vocabularies, producing bottlenecks that are hard to inspect. This paper proposes AnyBottle, which needs only a frozen backbone and an unsupervised concept pool, and uses a black-box teacher to iteratively select concepts that explain current failures, trained with nested dropout so any concept prefix can predict. Across six vision and two text datasets, it achieves fewer concepts and higher concept consistency than annotation-free baselines while staying close to the black-box reference.
Wolfgang Stammer, Sukrut Rao, Hevra Petekkaya et al.
In briefCIDOC CRM Periodic Table: A Testable Guide to Adopting Research Ontologies in the Humanities
SummaryThis paper presents the CIDOC CRM Periodic Table, a single-page browser tool that brings together 131 classes and 480 properties from CIDOC CRM 7.1.3, CRMsci, CRMdig and CRMinf. It offers a one-step triple builder and a multi-step path finder, both able to resolve chains across model boundaries. The authors also report preliminary usage signals from 56 months of analytics.
In briefHow High Is 0.6? Judging Interpretability Probe Scores Against Floors and Ceilings
SummaryThe paper argues that a probe score is only meaningful relative to two reference points: a floor, what simple inputs already predict, and a ceiling, what the full input predicts. The gap between them, called headroom, is where a probe can show a model computes beyond simple inputs. The authors prove headroom vanishes under two conditions and test the framework on transformers doing in-context meta-analysis. They also reanalyze the single-cell model scGPT and four LLM probing studies.
WhyProposes floor-and-ceiling headroom to read probe scores, proves when headroom vanishes, and validates on meta-analysis transformers and reanalyzed LLM probing studies.
In briefSteering Follows Geometry, Not Labels: Emotion Control in a Speech Model
SummaryThis paper studies emotion steering in Moshi, an open-source full-duplex speech language model, covering four emotions. It uses mean-difference activation steering, which needs only a few vector additions per frame and no retraining. The authors find emotion is linearly decodable from the model's residual stream, but steering is only partially achievable and uneven: happy, angry and surprise move toward a shared direction, while sad is distinctly steerable. They also report that the shared component across the three emotions cannot simply be projected away equally from all emotions.
In briefTiny neural controllers for safe, real-time robot control on microcontrollers
SummaryThis paper proposes Micro Neural Policies (MNP), very small neural network controllers for resource-limited embedded robots. The authors combine Evolution Strategy with Statistical Model Checking to search for policies that stay safe while shrinking network size. Policies were tested on Cartpole and Quadrotor tasks in simulation and transferred zero-shot to physical systems. Their memory footprint of 0.5 to 7.5 kB allows microcontroller deployment with low inference jitter.
Hongpeng Cao, Riccardo Curcio, Daniele Ottaviano et al.
In briefToward Alignment Scaling Laws: A Framework and First Preregistered Measurements
SummaryThe paper treats AI alignment difficulty as a set of measurable scaling relations, modeling how the alignment burden needed for a fixed safety target grows with model capability for each risk category. It proves several results on which regime dominates and how small-model fits can mislead. Preregistered reanalyses and pilots on Pythia and Qwen models give exponents that differ by risk type, with some remaining undetermined.
In briefRSJEV: Discriminative Remote Sensing Scene Classification with Multimodal Large Language Models
SummaryThis paper presents RSJEV, a one-pass multimodal framework for classifying remote sensing scenes. Instead of generating category names token by token, it directly estimates the probability of each candidate category from joint visual and text representations. Experiments on UC Merced, AID, and NWPU-RESISC45 report better accuracy than CNN, Transformer, Mamba, CLIP, and MLLM-based methods, with lower inference cost using a 0.8B-parameter model.
In briefA scalable method for probabilistic load forecasting that shares patterns across customers while adapting locally
SummaryThis paper proposes a scalable framework for probabilistic load forecasting at customer and transformer levels. A shared model learns common demand behavior, while each load combines a small bank of low-dimensional adaptation components. Experiments on 590 SMART-DS load profiles show better deterministic accuracy and probabilistic quality than several baselines, with low storage and inference costs.
In briefFlowCF: Sparse Counterfactual Explanations for Mixed-Type Tabular Data via Flow Matching
SummaryThis paper proposes FlowCF, a model-agnostic generative method for producing counterfactual explanations of model decisions. It treats counterfactual generation as sparse transport from the factual to the target class, using flow matching extended with a mixed flow operator for mixed feature types and a gating network that minimises changed features. On six benchmark datasets, FlowCF changes 29% of numerical features versus 89% for the best baseline, with 70% smaller displacement, and remains comparable on other desiderata.
In briefHow the choice of divergence shapes the Shampoo optimizer
SummaryThis paper studies how the divergence used to learn Shampoo's preconditioner affects its behavior in neural network training. The authors propose a unified Bregman divergence framework that connects popular divergences and analyze gradient second moments spectrally. They find some divergences better compensate for finite-sample underestimation of the second moment, and validate this explanation with GPT-2 pretraining experiments.
In briefBeyond Perturbation Magnitude: Direction-Dependent Responses in Multimodal Geometric Representations
SummaryThis paper studies how Gram-determinant-based multimodal alignment scores respond to modality degradation such as video blur and audio noise. Using MSR-VTT and DiDeMo, it finds that displacement magnitude alone explains little of the response. It proposes the Directional Geometric Response (DGR), the projection of the displacement onto the local volume gradient, which reports high out-of-sample explanatory power.
In briefMedCORE: Interpretable Medical Image Diagnosis Using Clinical Criteria Reasoning
SummaryMedCORE is a vision-language framework that breaks medical image diagnosis into clinically defined criteria and localizes each one to relevant image regions. It models dependencies between criteria with a Graph Attention Network and aligns them with clinical text descriptors and class-wise visual prototypes. The method was tested on dermoscopy (ISIC 2018), breast ultrasound (BUSI), and diabetic retinopathy (IDRiD) datasets, where it reportedly outperformed CNN, transformer, and other baseline models.
Asim Khan, Samee Ullah Khan, Dwarikanath Mahapatra
In briefPHBA: a hybrid attention design pairing sparse block retrieval with compressed prefix states for long-context models
SummaryThis paper introduces Prefix-State Hybrid Block Attention (PHBA), which replaces local sliding-window attention with top-k block-sparse retrieval. Each retrieved block is paired with a compact prefix state summarizing its preceding context, built by a gated linear recurrence at block boundaries. The authors also provide a hardware-aware Triton implementation. According to the paper, PHBA improves long-context and retrieval performance over strong linear and hybrid baselines while keeping training and inference efficient.
In briefWareFly-VLA: A Vision-Language-Action Framework for Drone Navigation and Human Tracking in Smart Warehouses
SummaryThe paper presents WareFly-VLA, a photorealistic framework and dataset for language-guided drone search, localization, and tracking of workers in warehouses, built in NVIDIA Isaac Sim. The dataset contains 507 teleoperated flight episodes with 8,504 RGB frames, each paired with a target-worker description and four-degree-of-freedom control commands. Four open-source vision-language-action models were benchmarked under a leakage-free protocol, and performance dropped substantially under strict generalization, showing the task is far from solved. The dataset, baselines, and evaluation protocol are released publicly.
Thinh D. Le, Son T. Nguyen, Duong Q. Nguyen et al.
In briefSpatial reasoning on grids with adaptive neural cellular automata
SummaryThe paper notes that many learning approaches still struggle with spatial reasoning, failing to use geometric information about entities and their spatial relations. The authors propose an Adaptive Neural Cellular Automata (aNCA) that uses deformable convolutions to dynamically adapt its perceptive field and iteratively reason over 2D spatial relations on grid-like data such as images. Empirical results on public benchmarks show state-of-the-art, comprehensible performance with strong generalization on image puzzles like Sudoku and shortest-path finding in mazes.
In briefHow Much Evidence Should a Coding Agent's Self-Correction Carry? Adaptive Dirichlet Evidence for Self-Distillation
SummaryThis paper proposes Effective-Evidence Self-Distillation (EESD), which separates transition support from evidence mass when coding agents learn from execution-based self-corrections. A Dirichlet posterior produces an uncertainty-penalized weight for KL-anchored correction learning. The authors report that after one correction-learning round, DeepSeek/CodeARC all-tests Pass@1 rose from 15.0% to 20.4%.
In briefWiki-Talkie: Testing AI Personas Across Languages Using Real Wikipedia Discussions
SummaryThis paper introduces Wiki-Talkie, a multilingual dataset of conversations from Wikipedia Talk pages in five languages, paired with personas derived from real user communities. It evaluates how well agents generate the next conversational turn under different persona conditioning strategies. The results show that a user's comment history outperforms explicit persona information, and that models underproduce negative or extreme sentiment while overproducing references and suggestions. These patterns hold across languages with small differences.
In briefCylindrical Geodesic Flow Matching: a method for translating quasiperiodic physiological signals
SummaryThis paper proposes cylindrical geodesic flow matching, a method for translating quasiperiodic cardiovascular waveforms recorded at one body location into the corresponding target waveform at another. The authors argue that the standard affine path in flow matching distorts intermediate amplitude and instantaneous frequency, and replace it with a closed-form geodesic on the phase-amplitude cylinder. On zero-shot photoplethysmography and limited-support seismocardiography adaptation benchmarks, the method outperforms interpolation baselines and matches or exceeds direct supervised prediction, reducing Hilbert Transform, L2, and Dynamic Time Warping distances by up to about 15%.
Onur Selim Kilic, Afra Nawar, Cem Okan Yaldiz et al.
In briefX-OPM: An explainable automatic on-chip power model for better robustness
SummaryThis paper presents X-OPM, a framework for building digital on-chip power meters that estimate processor power for power management. It uses feature engineering grounded in synchronous digital circuit design, with tree-based models capturing feature interactions and linear models making predictions, plus a human-in-the-loop workflow. On a commercial C906 vector processor, it reports R² above 0.93 across all workloads with sampling windows under 8 cycles, while several baselines fail to generalize. Layout results show an area overhead below 0.1%.
In briefSpinDoctor: Finding Harmless-Looking Prompts That Steer Image Generators Toward Bias
SummaryThis paper presents SpinDoctor, a black-box method for finding seemingly innocuous text prompts that make text-to-image models generate images with a chosen bias while keeping image and prompt quality. In 302 of 304 experiments, the prompts moved outputs toward the target bias, with average improvements of 34.8% and 33.5% on two open-source models. Some bias transferred across models in 78.6% of cases, and part of the results were confirmed by human evaluation. The authors note such prompts could be used to manipulate users.
In briefLanguage-model depression ratings mostly reflect the rater, not the patient
SummaryThe study used 880 language-model raters, combining 11 open models with different prompting and scoring choices, to assess 189 interviews against the eight-item Patient Health Questionnaire. Model choice explained 30.0% of summed-symptom score variance, while stable participant differences explained 10.5%. Two randomly drawn raters with comparable accuracy disagreed on screening decisions for about 40% of participants.
In briefA method for searching molecules with rare properties by testing mixtures
SummaryMachine learning can speed up molecular design, but when target molecules are very rare, existing algorithms perform little better than random guessing. The paper proposes searching large molecular spaces through algorithmically controlled stochastic synthesis, making and testing mixtures of molecules as pools and then deconvoluting the molecule-activity map. Theoretically, this can reduce the experiments needed to find the optimal molecule among d candidates from O(d) to O(log d) or O(1). In simulations on estimated protein fitness landscapes, it finds active molecules with about an order of magnitude fewer experiments than existing Bayesian optimization methods.
