In briefVideo text editing with trajectory-aligned glyph rendering for accurate characters
SummaryThis paper addresses replacing or adding text in videos while keeping the rest of the footage unchanged. The authors propose a trajectory-aligned glyph rendering reference that guides each frame's text position and perspective, and a depth-normalized recognizer feature supervision that targets stroke errors missed by diffusion loss. They also release VTEdit, a benchmark of 288 real-scene clips with 440 annotated text trajectories. On VTEdit, the method outperforms image and video editing methods and commercial models in text accuracy and background preservation, reaching a sentence accuracy of 0.9408.
In briefReplayLens: Auditing How Agents Use Logged Outcomes
SummaryThe paper introduces ReplayLens, a black-box audit that tests whether an agent's changed decisions come from recorded scores, action names, or record positions in storage. It alters one relationship in the stored history at a time while keeping the rest of the interface fixed. Experiments on LLM interfaces, bounded memory, sequential experiment planning, and code debugging show that endpoint accuracy alone cannot reveal these dependencies.
In briefAGILE-GS: Fast Anchor-Guided Camera Viewpoint Selection for Active 3D Gaussian Splatting
SummaryThis paper presents AGILE-GS, a method for choosing the next best camera view when building 3D Gaussian Splatting scenes. It optimizes a virtual anchor camera pose that marks where the model is most uncertain, then scores candidate views against it and uses a greedy step to keep a small, non-redundant shortlist. Tested on standard benchmarks and closed-loop embodied acquisition, both variants match or exceed baselines while cutting selection latency by one to two orders of magnitude.
In briefGradLev: Parallel Test-Time Training via Costate Prediction
SummaryThe paper introduces GradLev, a method for parallelizing test-time training (TTT), where a model updates its weights after each token during inference. Because sequential gradient updates make parallel training difficult, the authors show that online gradient descent can be computed exactly in parallel given layer inputs and activation gradients (costates). A causal auxiliary network predicts costates for all tokens at once, associative scans compute adapted weights and propagate gradients, and a consistency loss supervises the predictor. The auxiliary predictor is discarded at deployment, and the model updates token by token as usual.
In briefGPARA: Graph-based posterior-aligned refinement and active measurement for diffusion-model reconstruction
SummaryThis paper proposes GPARA, a method for actively choosing measurements and refining reconstructions when using a frozen diffusion model as a prior. It learns a context-dependent graph surrogate over diffusion prediction residuals to estimate the value of candidate measurements and update the reconstruction. Experiments on physical field and computer vision reconstruction tasks show consistent improvements over the evaluated baselines in refinement and active acquisition.
In briefNatural-image autoencoder features for fMRI trait and state prediction
SummaryThe paper introduces FReD, which extracts fMRI representations from a frozen Deep Compression AutoEncoder pre-trained only on natural images, paired with task-specific readouts. On four resting-state datasets covering six trait-prediction targets, linear probes on these frozen features generally outperform fMRI foundation model representations and remain competitive with fully fine-tuned foundation models. On three task-fMRI state-prediction tasks, the approach performs comparably to the strongest foundation models evaluated.
Juhyeon Park, Yeonwoo Kim, Peter Yongho Kim et al.
In briefHidden Activations Are Not Enough I: Knowledge Matrices as Higher Representations
SummaryThis paper treats the knowledge matrix of a trained feedforward network as a higher-level representation of its inputs, with rows that sum exactly to the logits. The authors analyze what determines this matrix, what it is invariant to, and what its geometry measures. They use it to compare ResNet-152, DenseNet-121 and GoogLeNet and to study adversarial attacks. The paper also reports two negative results, including that penultimate-layer features outperform it on AlexNet/CIFAR-10 in most detectors and attacks.
In briefSlideDP: Scaling Host-Resident LLM Fine-Tuning Across Multiple GPUs
SummaryThis paper presents SlideDP, a synchronous data-parallel runtime for fine-tuning large language models that exceed GPU memory, using host-resident state on multi-GPU systems. It keeps one authoritative host state and pipelines parameter delivery, gradient aggregation, and CPU updates across GPUs. In matched-batch tests, it reports 1.46 to 2.64 times the geometric-mean throughput of SlideFormer, MegaTrain, and ZeRO-Offload, and approaches GPU-resident FSDP2 throughput on four H100s for Qwen3-14B.
WhyExtends host-resident layer streaming to shared-host multi-GPU data parallelism with an analytical model; evaluated against several baselines on large models, a substantial improvement over prior work
In briefGT-PSSM: A unified probabilistic method for detecting anomalies in multivariate time series
SummaryMany anomaly detection methods for multivariate time series rely on deterministic models and point-wise errors, which can be unreliable because real data is noisy and random. The paper proposes GT-PSSM, which combines a probabilistic state-space model with a Graph Transformer. It jointly models stochastic dynamics, long-range temporal dependencies, and interactions between variables. The authors aim for more robust anomaly detection.
In briefRoutePrism: Tracing How Processing Order Shapes Agent Memory
SummaryThe paper introduces RoutePrism, a diagnostic protocol that builds agent memory twice from the same records in two different orders and traces differences in sources, compiled contexts, and answers. A matched four-condition intervention shows that restoring one record displaced by reordering recovers over 60 percentage points of lost accuracy, while substituting a non-supporting record of equal length does not. The protocol is evaluated on PersonaMem-32K and LongMemEval-S, with the core intervention replicated across five answer models.
In briefWorldGraph: Modeling Worlds Natively as Evolving Graphs
SummaryThis paper studies graph world modeling (GWM), where the evolution of a graph itself is treated as the world's dynamics rather than as a tool for organizing internal states. The authors build GWM-Zero, a benchmark of node-, edge-, and graph-level transitions across eight temporal graph datasets. They propose WorldGraph, which pairs a state-aware graph transformer with transition-aware GRPO using dynamic grouping and structure-aware verifiable rewards. Experiments report that WorldGraph consistently outperforms the listed baselines across all three transition granularities.
In briefA Graded Measure of How Stable LLM Beliefs Are Within Their Wider Belief System
SummaryThis paper introduces graded belief stability, a measure of whether a language model's belief in a claim persists when considered alongside its other epistemic commitments. It estimates conditional belief probabilities with a Direct Conditional estimator that uses internal model representations. Across 12 LLMs and three domains, after matching on individual belief probability, lower-stability beliefs showed greater mean behavioral movement under conversational challenge in 83.3% of model-domain settings.
In briefTableSeek: An agent that actively searches table collections for evidence while keeping table structure intact
SummaryTableSeek is a search framework for finding evidence in heterogeneous table collections. An LLM agent iteratively follows sparse clues, inspects previews that preserve table schema, and refines its search, using cells and schemas as anchors while keeping full tables as evidence units. The authors report competitive end-to-end performance against strong retrieve-and-rerank pipelines on heterogeneous table benchmarks, without retriever training or a precomputed semantic index.
In briefTransfer Calibrated Prediction-Powered Inference: adapting a source-domain predictor to a target domain with gold-standard samples to sharpen confidence intervals
SummaryPrediction-powered inference (PPI) and its extension PPI++ tighten confidence intervals by combining a small set of gold-standard labels with predictions from a large AI model, but their efficiency depends on low residual variance, which may fail if the predictor was trained on a different domain. This paper proposes TC-PPI, which uses cross-fitting to adapt the source-domain predictor to the target domain with gold-standard samples, supporting methods such as sparse linear calibration, LoRA, and fine-tuning. The jointly tuned estimator Joint-TC-Cross-PPI++ stays unbiased and is at least as efficient as classical inference, PPI, and PPI++, guarding against negative transfer. The authors give high-dimensional MSE bounds for calibration and report empirical gains over baselines on several real-world applications.
Aditya T. Vadlamani, Jae Ho Chang, Srinivasan Parthasarathy et al.
In briefSPINET: a sheaf-based network that designs protein sequences from motion trajectories
SummaryThis paper introduces SPINET, a model that predicts protein amino acid sequences from molecular dynamics trajectories rather than a single fixed structure, so the design can account for how proteins move. It represents residue interactions within each frame using cellular sheaves, integrates information across frames with recurrent units, and predicts all amino acids in one pass. On mdCATH, SPINET reaches 56.7% top-1 sequence recovery, above the strongest static (44.5%) and ensemble (40.7%) baselines, and its predicted sequences have a median TM-score of 0.760 with target-trajectory conformations.
Jens Lundsgaard, Colin Mikulski, Zhixuan Yan et al.
In briefMeasuring how close Danish, Norwegian and Swedish are, using pretrained language models
SummaryThis paper proposes a computational framework with three measures (semantic distance, orthographic fragmentation, and masked-prediction predictability) to quantify distances among Danish, Norwegian (Bokmål), and Swedish. Using 150 trilingual parallel sentence triplets from Tatoeba, the authors find consistent rankings across two models (LaBSE and mBERT), with Danish and Norwegian closest. The authors state this matches historical linguistic findings, and note that extension to other language groups requires further validation.
In briefSelf-evolving agents that refine their tool sets using output-based likelihood scoring
SummaryThis paper studies how an agent can continually refine a persistent, task-specific set of tools based on accumulated output experience, with model parameters kept fixed. The authors introduce LOTS, which scores each tool by measuring how much the likelihood of the generated answer changes when that tool's observed output is removed, then aggregates these scores per task. Across three benchmarks, LOTS improves task performance while substantially reducing tool context, and the learned configurations transfer across different models.
In briefFunctional Hand Type Priors for 3D Hand Pose and Action Recognition from Egocentric Monocular Video
SummaryThis paper addresses the difficulty of recognizing dynamic hand actions in egocentric video when only geometric or physical cues are used. The authors define a practical taxonomy of hand types based on function, label existing datasets with it per frame, and propose a framework that uses hand-type semantics as a prior. The pipeline combines feature extraction, an Egocentric Knowledge Module for 3D hand pose, object category and hand type, and an Egocentric Action Module that aggregates information over time. The model reportedly outperforms state-of-the-art methods on the FPHA and H2O benchmarks.
Wonseok Roh, Seung Hyun Lee, Won Jeong Ryoo et al.
In briefGeometric Encoding for Spatial Reasoning in Vision-Language Models
SummaryThis paper presents Geometric Code, a pipeline that extracts depth, camera pose, and object information from monocular video and converts it into an explicit spatial code of object positions, dimensions, and distances. The code is provided to vision-language models as text, without training or fine-tuning any component. On VSI-Bench, the authors report that adding the spatial code improves average accuracy of 2B and 4B open models over a frames-only baseline, with the largest gains on numeric estimation tasks such as absolute distance.
In briefExpertoRhythm: A Model That Estimates Cuffless Blood Pressure From a Single PPG Signal
SummaryThis paper presents ExpertoRhythm, an attention-enhanced 1D U-Net that reconstructs the arterial blood pressure waveform from a single-channel PPG signal and derives systolic and diastolic pressure from it. The authors add a composite loss combining range-weighted SmoothL1 reconstruction with a window-range regularizer to better capture high-dynamic pressure segments. On the UCI cuffless BP dataset with 942 subjects, it reports MAEs of 2.46 and 1.46 mmHg for systolic and diastolic pressure, and states that it meets AAMI criteria and achieves BHS Grade A.
In briefCAST: A Method That Speeds Up Interactive World Models While Preserving Reconstruction Quality
SummaryInteractive world models must respond quickly to controls while keeping scenes consistent, and existing speed-up methods can struggle to recover skipped features. This paper proposes CAST, which chooses anchors by interaction sensitivity, reconstructs skipped residuals with a phase-aware method, and routes historical KV features by downstream reconstruction responsibility. On Matrix-Game 3.0 and HY-World 1.5, the authors report speedups of 2.15x and 3.48x, with visual quality close to the native model and the highest VBench scores among compared methods.
In briefBeyond Geometry: Benchmarks and Reasoning for Detecting 3D Logical Industrial Anomalies
SummaryMost existing 3D industrial anomaly detection focuses on local geometric deviations, while many anomalies break object-level design or assembly rules, termed 3D logical anomalies here. The authors introduce ILGAD, the first scalable benchmark for logical anomalies in industrial point clouds, with 2,774 samples across 15 categories and point-level annotations. They also propose a consistency reasoning framework that checks local geometry, structure coverage, and spatial relations against normal design, and report strong object-level detection and point-level localization on ILGAD, Anomaly-ShapeNet, and IEC3D.
In briefEfficient Video Object Detection via Analytical and CNN-Based Motion-Vector Propagation
SummaryThe paper proposes a hardware-software method that reuses codec motion vectors between detector calls to cut latency and energy in video analytics. It compares an analytical propagation model (Analytical-MV) with a learned convolutional neural network model (CNN-MV). Analytical-MV reaches an F1 score of 0.909 with 9.03 ms mean latency and 0.177 J per frame, reducing latency by 25.9% and energy by 36.4% versus per-frame detection.
Ashiyana Abdul Majeed, Mahmoud Meribout, Neethu Joseph
In briefSame Tasks, Different Apps: Why Mobile GUI Agents Fail to Generalize
SummaryThis paper introduces AnyAppBench, a category-controlled live Android benchmark that tests whether mobile GUI agents can complete the same user goal across different apps. It covers 10 functional categories, 100 task templates, and 520 task-application pairs over 52 apps. The authors find that success on an original app does not reliably transfer to new apps with the same goal across 13 agents. Providing high-level sub-goal decomposition yields only small, category-dependent gains, and failure patterns shift with the target interface.
WhyIntroduces a new cross-app generalization benchmark for mobile GUI agents, validated on 520 task-app pairs across 13 agents with human-checked failure taxonomy.
In briefWord similarity datasets for six Indian languages, with baseline scores for word representation models.
SummaryThe paper presents manually annotated monolingual word similarity datasets for six Indian languages: Urdu, Telugu, Marathi, Punjabi, Tamil, and Gujarati. The datasets were built by translating and re-annotating English word similarity datasets. The authors also report baseline scores for word representation models on the new Urdu, Telugu, and Marathi datasets.
Syed S. Akhtar, Arihant Gupta, Avijit Vajpayee et al.
In briefYou Can't Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning
SummaryThis paper studies concept unlearning in text-to-image diffusion models, which aims to suppress a target concept while keeping related ones. The authors argue that the trade-off stems from the geometry of concept representations and prove that overlap between concepts sets a lower bound on the damage strong erasure causes. Testing thirteen unlearning methods, they find none achieves both strong erasure and strong preservation of neighboring concepts.
WhyFormalizes concept entanglement geometrically with a proven lower bound, validated across thirteen methods and multiple models, offering novel and fairly strong evidence.
Yian Wang, Ali Ebrahimpour-Boroojeny, Hari Sundaram et al.
In briefWaggle: Learning One Shared Local Rule for Self-Organizing LLM Agent Swarms
SummaryThis paper introduces Waggle, a shared anonymous policy that lets interchangeable LLM agents choose task actions, semantic messages, and local commitments from bounded local views. The authors also propose Swarm-Consistent Distillation (SCD) to train this rule. According to the paper, the same learned rule stays effective as population size and interaction budget change, retains over 96% of substrate-specific oracle quality, and transfers without retraining.
In briefEvo2Team: When Evolved Skills Transfer Across Multi-Agent Teams
SummaryThis paper studies whether routing and communication skills evolved for one multi-agent team transfer to another, using teams of 4 to 32 agents and GPT and Qwen models. It introduces Evo2Team, which selects, adapts, and confirms source skills for a target team, and compares it with six frozen selectors across 28 transfer directions. The authors conclude that skill transfer must be judged by the actions agents take, the tasks those actions reach, and the quality and cost of the final deployment, not by selection alone.
In briefStateGuard: A Framework for Managing Analytical State Validity in Long-Horizon Data Agents
SummaryThe paper addresses a problem in LLM-based data agents: during long multi-stage analyses, constraints, variables, and conclusions stay implicit in interaction histories, so stale results can be silently reused and propagate errors downstream. The authors propose StateGuard, which externalizes analytical progress into a state graph and maintains validity through evidence-grounded verification and hierarchical intervention. Experiments on three long-horizon data-analysis benchmarks report improved agent performance and reduced dependency-induced error propagation.
In briefPrefLUT: Reusable, Refinable Personalized Color Editing from Pairwise Preferences
SummaryThis paper presents PrefLUT, a framework that learns a user's color-editing taste from ordered pairs of preferred and non-preferred images. The preferences are encoded into a lightweight Reusable User Profile that can be reused across queries and refined with more pairs, without per-user optimization. The model predicts an exportable 3D LUT and edit strength for each image, and experiments on three datasets show effective personalized and general-purpose enhancement. The authors also propose the PCVP evaluation protocol to verify whether edits depend on user preferences and query images.
In briefFrom Attack Success to Attack Severity: Counterfactual Memory Attacks on LLM Agents
SummaryLLM agents increasingly keep persistent memory, so malicious memory writes can keep affecting their future behavior. This paper argues that the severity of a successful attack, not just whether it succeeds, is a separate design goal, and measures it with counterfactual memory regret (CMR), the increase in downstream loss relative to clean memory. The authors propose MemHarm, which selects from a predeclared set of sparse semantic edits using offline paired-loss feedback. CMR-guided selection produces substantially larger downstream loss while keeping most of the success-rate gain.
In briefHow PLDR-LLM Language Models Learn and Predict: Row Collapse, Renormalization, and Reduced Prediction
SummaryThis paper develops a unified account of training and inference in Power Law Decoder Representation language models (PLDR-LLMs). It uses exact finite work identities to split changes in the energy of the learned row-centered map into parameter contributions, interactions, and numerical errors. Experiments show dependence on the observer and optimizer and reject the tested autonomous row-state candidates.
In briefSTITCH-RAG: Tracing Spatio-Temporal Influence over Topic Hypergraphs for Multi-Hop RAG
SummaryThis paper proposes STITCH-RAG, a hypergraph-based retrieval framework for multi-hop retrieval-augmented generation. It combines a semi-merged topic hypergraph, spatio-temporal influence bridging propagation, and continuous scores in localized Personalized PageRank to connect evidence spread across documents. Under the reported protocol, it shows the highest Contain-Acc and LLM-Acc point estimates and higher Recall@8 on HotpotQA and 2WikiMultiHopQA; mixed-domain results are auxiliary because only LLM-judged accuracy is available.
In briefJET: Judge-Guided Evolution at Test Time for Agent Programs
SummaryThis paper introduces JET, a method that evolves an executable judge on labeled source trajectories, then freezes it and transfers it to guide the evolution of an agent's program at test time, without true rewards or model-weight updates. On unseen WebShop tasks, JET achieves about 13% higher mean reward than fixed-rubric guidance from a cold start and 4% higher from a warm start. Cold-start exact success improves by a 36% relative gain.
In briefUsing Large Language Models to Understand Clinical Cognitive Assessment Dialogues
SummaryThis paper presents a de-identified corpus of 33 cognitive assessment conversations, with 8,250 utterances labeled by three speaker roles and 56 dialogue acts. The authors benchmark large language models on fine-grained dialogue-act classification and next-patient-utterance generation, and test out-of-domain instruction data and explanation-augmented training. Instruction tuning gives the best patient-utterance matching and improves classification accuracy, but even the best models struggle to separate closely related dialogue acts. The work makes no diagnostic claims and offers data and an evaluation framework.
Vishalakshi Arumugam, Dan Schumacher, Veronica Rammouz et al.
In briefSpatialSkill: Self-Evolving Skills for Cross-View Spatial Reasoning
SummaryThis paper proposes SpatialSkill, a framework that improves a frozen vision-language model's ability to reason across different viewpoints without updating its weights. It accumulates explicit natural-language reasoning skills from offline trajectories, admitting them only after visual-grounding and executability checks. On CityCube, the method gives consistent gains across four frozen executors, and a 9B executor with SpatialSkill surpasses the strongest closed-source reference in the evaluation.
In briefWhat a stream model adds to flow matching, and whether it helps
SummaryThis paper examines stream-level flow matching, which replaces the linear interpolant in conditional flow matching with Gaussian-process streams and reports lower sample error on some benchmarks. The authors argue that the conditional paths such streams can reach are exactly the Gaussian paths that conditional flow matching already parametrises. They also find that the single kernel couples mean and variance, so the paper's recipe for widening coverage gives no benefit. They further report that the released code does not implement the state-velocity correlation mechanism the paper describes.
In briefProbabilistic forecasting of electricity demand with uncertainty quantification
SummaryMost electricity consumption forecasting research produces only a single point estimate per time step. This study empirically compares four contemporary probabilistic forecasting models on real-world power system datasets. Across all consumption zones, NGBoost achieved the lowest MAE and RMSE, with well-calibrated uncertainty estimates, high prediction-interval coverage and reasonably narrow intervals. The authors conclude it outperforms Bayesian, MC Dropout and Gaussian Process Regression models on the data considered.
Mahesh Neupane, Pragya Dhungana, Pradip Khatri et al.
In briefSlimWise: Pruning Mixture-of-Experts Models Separately for Prefill and Decode to Speed Up Serving
SummaryMixture-of-experts models activate only a few experts per token, but batched decoding touches nearly all expert weights, creating a bottleneck. SlimWise serves prefill with the full model and decode with a pruned model that reuses the prefill KV cache without conversion. Across two backbones and three pruning criteria, this training-free handoff narrows accuracy gaps in many settings, and a low-cost distillation stage reduces the remaining loss. On Qwen3.6-35B-A3B at 50% expert pruning, decode throughput improves by up to 1.81x with minimal accuracy loss.
In briefForecast-necessary causal discovery for nonlinear political panel data
SummaryThe paper presents an inferential workflow for political panel data that combines flexible autoregressive estimation, forecast-necessity testing, functional-form characterization, and same-data linear benchmarking. Applied to the V-Dem panel of 113 countries, it reproduces the core finding on civil society, rule of law, and institutionalized parties, and recovers reciprocal effects from democracy to its institutional supports. It also gives three weak published direct effects different diagnoses, removing one and recovering two.
Michael Coppedge, Dmitry Zaytsev, Valentina Kuskova
In briefEvolution of fairness in a multi-objective reinforcement learning framework (how fairness emerges when agents balance payoff and moral concerns)
SummaryThis paper proposes a multi-objective reinforcement learning framework that models fairness as a dynamic trade-off between material payoff maximization and fairness-driven moral behavior, controlled by a fairness pressure coefficient. Simulations of a two-objective Q-learning ultimatum game show that higher fairness pressure promotes fair outcomes. Under moderate pressure, responders become forgiving and accept low offers, a reversal the authors attribute to competition between payoff-maximizing and fairness-oriented preferences.
In briefGUITAR: Structured Failure Diagnosis of GUI Agents via State Transitions
SummaryThis paper argues that current step-accuracy evaluation of GUI agents ignores the functional structure of interfaces. It proposes GUITAR, which maps visually different but functionally equivalent screens to shared states in a State Transition Graph and analyzes failures over both states and transitions. Across 8 agents and 6 tasks from AndroidControl and Mind2Web, 60.4% of failures occur in 20% of states. Bottleneck-targeted guidance improves SR by 2.8%.
In briefUnknown Is Not Normal: Separating LLM Extraction From Rule-Based Logic for Clinical Risk Scores
SummaryThis paper examines how large language models can compute clinical risk scores from free-text notes without treating undocumented findings as normal. The authors extract each input as present, absent, or unknown, and use deterministic code to compute score bounds, asking only questions that could change the decision. Tested on 1,200 synthetic emergency cases across six calculators, this bounds approach matched ask-all accuracy with about half the questions, while treating missing values as normal caused under-triage.
In briefSpecRegMatch: Robust Semi-Supervised Regression for Predicting Vehicle Interior Noise
SummaryThis paper proposes SpecRegMatch, a semi-supervised regression method for predicting vehicle interior noise levels. It trains a single model using consistency regularization and information maximization, applying various augmentations to embedding vectors and predicted values, which avoids the high cost of training multiple models and repeated data sampling. The authors report state-of-the-art results across several scenarios, with an R² score of 0.434, and note the method is especially useful when labeled data is scarce.
In briefSpectrum-Balanced Feature Matching: A Fix for Biased Knowledge Distillation
SummaryLarge visual foundation models transfer well but are costly to deploy, so feature matching is used to distill their knowledge into smaller student models without labels. The paper argues that the standard L2 feature matching objective favors dominant spectral directions of the teacher's representation and under-optimizes low-variance directions that may carry task-relevant information. The authors propose SpecMatch, which adaptively emphasizes under-optimized spectral directions while keeping dominant ones' relative importance, with negligible added cost. Experiments report gains over conventional feature matching in 40 of 42 settings and across six protein understanding tasks.
In briefReconstructing Implicit Scientific Knowledge: Evaluating LLM Agents by Reproducing Astronomy Papers End to End
SummaryThe paper proposes a framework that evaluates LLM agents by having them reproduce published research end to end, separating execution and verification from computational failure and methodological ambiguity. It applies this to fourteen astronomy studies, finding that eleven of thirteen Nature papers had an ambiguity preventing a uniquely specified reproduction path. In a controlled case, twelve predefined paths gave estimates from 2.16 to 3.53 kpc, and only one matched the published value of about 2.70 kpc. The decisive +0.02 mas parallax zero-point correction was already in the paper, but agents did not recognize its relevance until the effect became visible.</summary>
In briefA differentiable framework that jointly registers bounding boxes over LiDAR point-cloud streams
SummaryThe paper proposes a training-free method that refines sequences of 3D bounding boxes against LiDAR point-cloud streams by jointly optimizing all frame poses with temporal smoothness built into the registration objective, rather than smoothing afterward. On well-observed objects it beats post-hoc smoothers and reduces roughness substantially at matched accuracy. The advantage shrinks as views become one-sided and vanishes for near-edge-on objects, where the decoupled pipeline can be ahead at tight roughness budgets. The authors attribute this limit to orientation alignment failing when the estimated velocity is noisy, and give a ground-truth-free rule for choosing the temporal scale.
SummaryThis paper argues that common semantic-ablation experiments in tabular learning can mislead conclusions about whether semantic knowledge improves prediction. It separates content sensitivity, the performance change when semantics are altered, from predictive utility, the benefit relative to a suitable reference without that knowledge. An audit of 25 comparisons across nine studies found that only one of 18 explicit predictive-utility claims used a control that clearly isolates the semantic contribution.
In briefUsing Environmental Data to Help Identify Individual Animals in Wildlife Monitoring
SummaryThis paper introduces MetaWild, a multimodal dataset for animal re-identification with 20,890 images across six species, paired with environmental metadata from camera trap overlays and scene contexts. It also proposes the Meta-Feature Adapter (MFA), a lightweight module that can be added to existing vision-language model-based re-identification methods. Experiments show that adding metadata consistently improves performance over using visual information alone.
In briefEvaluating Machine Unlearning for Speech Recognition Models
SummaryThis paper examines whether existing machine unlearning methods and evaluation tools suit automatic speech recognition (ASR) models, which relates to right-to-be-forgotten rules. The authors apply several unlearning techniques and compare privacy-utility trade-offs for single-subject unlearning, then test sequential and simultaneous unlearning. Gradient ascent-based methods perform best, while complex methods over-unlearn samples, and simple membership inference attacks are found insufficient for reliably judging unlearning success. Both sequential and simultaneous unlearning yield worse privacy and utility than single-subject unlearning.
Diogo Dinis, Francisco Teixeira, Bhiksha Raj et al.
In briefWhen Is an SAE Feature Interpretable? A Validation Ladder for EEG Models
SummaryThis paper examines how sparse autoencoder (SAE) latents in EEG foundation models are interpreted. The authors find that removing alpha-band activity changes latent activations 7.3 times more than a sham notch, but after normalizing by removed spectral energy the ratio falls to 0.28 and never exceeds one. They propose a validation ladder that checks whether a latent responds, whether the response survives controlling for removed signal, whether it is specific, and whether the property is visible on unperturbed data.
In briefTRACE: A Heart-Signal Model Aligned with Expert Reports for Acute Cardiac Care
SummaryTRACE is a multimodal model that learns representations from electrocardiograms (ECGs) by jointly using unimodal ECG learning and cross-modal learning with cardiologist reports. It is evaluated on public benchmarks for arrhythmia and structural abnormality classification. On a large private acute coronary occlusion (ACO) dataset, the paper reports a 19.0% sensitivity increase or a 62.6% reduction in false positive rate compared with the clinical baseline.
WhyCombines unimodal and cross-modal uncertainty-weighted multitask learning for ECG, with large private clinical validation in acute coronary occlusion showing substantial sensitivity and false-positive
Lovely Yeswanth Panchumarthi, Andrew Lu, Saurabh Kataria et al.
In briefAD-E2E-JEPA: A Predictive World Model Architecture for Faster End-to-End Self-Driving
SummaryThis paper first evaluates existing action-conditioned JEPA world models (LeWM, DINO-WM, JEPA-WM) for end-to-end autonomous driving using goal-conditioned zero-shot planning, without training any driving policy. It proposes AD-E2E-JEPA, which adds a learnable projector with SIGReg regularization to reduce planning patches and embedding size, achieving a 100x inference speedup while retaining planning performance. On the NAVSIMv2 benchmark it reports 67.3/72.9 EPDMS, and the pretrained projector raises downstream imitation learning from 80.2 to 85.4 EPDMS.
WhyAdds a SIGReg-regularized projector to JEPA world models, cutting planning cost 100x with NAVSIMv2 benchmarks and downstream imitation gains.
In briefBeyond One Epoch: Uncertainty-Weighted Regularization for Recommender Models
SummaryThis paper examines the one-epoch phenomenon in recommendation models with sparse embeddings and a shared consumer, where a second training epoch lowers training loss but hurts generalization. The authors attribute this to embeddings carrying label-induced displacement in later epochs, which the consumer can exploit in ways that fail to generalize. They propose uncertainty-weighted sensitivity regularization (UWSR), which penalizes reliance on uncertain embeddings, and report improved test cross-entropy and AUC over one-epoch training on three benchmarks.
In briefK-OPSD: A Verifier-Guided Self-Distillation Method for Training Vision-Language Models on Architectural Drawings
SummaryThis paper introduces K-OPSD, a post-training method that helps vision-language models better understand architecture, engineering, and construction (AEC) drawings. It builds a teacher from the model's own best-of-N outputs, filtered by a process-level verifier, and rescues failed prompts by resampling with a hint revealing the verified answer. The model is then updated on verified completions using a cross-entropy loss, and fine-tuned Qwen3-VL models are evaluated on AECV-Bench and transferred to ArchCAD.
In briefGenoMorph: Pathway-based reasoning framework for predicting genomic diseases with adaptive computation
SummaryThis paper introduces GenoMorph, a multimodal framework that aims to predict diseases from genomic data by reasoning over biological pathways rather than memorized gene-disease links. It combines a frozen DNA foundation model, self-adaptive latent reasoning (LatentSp), and rejection sampling fine-tuning. On an anonymized KEGG-based benchmark, the authors report a weighted F1 of 0.9465, and 0.9725 with fine-tuning, which they say also cuts latency by nearly 60%.
Tanmoy Kanti Halder, Akash Ghosh, Arijit Roy et al.
In briefWhiteCon: A Semi-Supervised Regression Method for Domain Adaptation Using Whitening and Consistency
SummaryThis paper proposes WhiteCon, a method for semi-supervised domain adaptation regression that targets continuous-output tasks, an area it describes as largely unexplored. It combines a domain-specific whitening transform with dual consistency regularization to improve training stability and adaptation across domains. The authors report state-of-the-art results on benchmark datasets under this setting, and the code is publicly available on GitHub.
In briefMaskCoFT: Memory-Efficient MoE Inference via Masked Co-Adaptive Fine-Tuning
SummaryMixture-of-experts language models often exceed single-GPU memory, so expert offloading loads experts on demand, making decoding speed depend on how many experts each token fetches. MaskCoFT fine-tunes routers and experts together using a learnable binary mask that restricts each layer's routing to a subset of experts, with only cross-entropy loss. At inference the mask becomes a soft prior that re-ranks experts while keeping all selectable. In simulations, expert fetches per token fell by 23.7% for Mixtral-8x7B and 10.1% for DeepSeek-V2-Lite, with average accuracy over nine benchmarks staying 0.92 and 0.53 points above the base models.
In briefInformation as a double-edged sword: comparing revealed and hidden lottery numbers in school choice.
SummaryIn school choice, when more students want a school than it has seats, the matching mechanism uses a lottery number to break ties. This paper relaxes the assumption that all students share the same preference ranking and uses a Bayesian Nash equilibrium game model to study how revealing the lottery number affects match rate and social welfare. It finds that revealing the number can either improve or worsen outcomes, and that revealing partial information has non-monotonic effects.
In briefPhysFieldBench: Can AI Models Understand Physical Fields?
SummaryThis paper introduces PhysFieldBench, a benchmark of 24 tasks and 1,160 examples covering controlled equation fields, simulated fields, and observed fields, testing whether multimodal models can infer physical mechanisms, compare control variables, and predict outcomes. Zero-shot performance is low, with the best model scoring 29.3 on a chance-normalized measure and several open-source models near chance. A task-specific supervised vision transformer performs much better, indicating the inputs contain learnable physical information. Reinforcement learning after chain-of-thought supervision generalizes best across tasks.
In briefSNaP: One-Step Posterior Sampling for Noisy Inverse Problems
SummaryThis paper presents SNaP, a one-step MeanFlow posterior sampler for linear inverse problems with Gaussian noise. Its key component is a measurement-adapted Gaussian source whose mean and anisotropic covariance depend on the measurement operator, observation, and noise level. The authors report that SNaP produces diverse, high-quality samples on natural-image restoration and multi-coil MRI with one network evaluation per draw, 30 to 2250 times faster than iterative samplers.
Shirin Shoushtari, Edward P. Chandler, Xiao Shi et al.
In briefFLARE: Flow matching in local axis-angle space for fast stochastic micromagnetic simulation
SummaryThe paper introduces FLARE, a flow-matching method that predicts stochastic magnetization evolution over a finite time horizon. It frames the prediction as conditional transport in anchor-relative local axis-angle rotation space, which keeps each point at unit norm. By conditioning on the physical prediction horizon, FLARE generates full-field stochastic endpoints at multiple target times without stepwise integration. The authors report lower error metrics than the strongest external baseline and a 3,062x best-batch speedup over MuMax³ on a composed 5-ns protocol on one GPU.
In briefCertificate-driven porting of scientific code to GPUs with a self-improving LLM search loop
SummaryThe paper introduces Certificate-Driven Evolutionary Search (CDES), which records restrictions from failed LLM-generated code candidates as certificates and uses them to reject, backtrack, or repair later candidates. It is applied to translating two Geant4 particle-simulation functions from CPU to GPU, evaluated with formal, numerical, physics, and GPU safety checks. The generated implementations reach 13.78x and 23.54x function-level speedups over CPU code, including data transfers. An ablation shows certificate feedback raised the share of candidates passing required correctness checks from 55% to 90%.
In briefWhether a name is a single token and how that affects model concepts
SummaryThe paper shows that names assumed to be matched inputs in bias evaluations are often not equal at the tokenizer level, since some names are single tokens while others split into multiple subwords. The authors introduce NameTrace, which measures concept accessibility from model probabilities over task-specific adjectives. They report that these differences persist across matched race and gender groups and across tasks such as fellowship, hiring, clinical assessment, and lending.
WhyIntroduces NameTrace showing unequal name tokenization persists into task-relevant hidden representations, validated across model families and unseen names.
In briefPerturbation-Robust Policy Learning for LLM Agents via Stable Optimization
SummaryThis paper studies why reinforcement learning-trained LLM agents can be sensitive to perturbations such as hidden-state noise, pruning, and quantization. The authors define perturbation-robust policies and analyze when perturbed policy updates preserve stable monotonic improvement. They propose SPrPO, which applies adaptive, sensitivity-aware perturbations during RL training, and evaluate it on ALFWorld and WebShop across several perturbation types and scales. The reported results show improved robustness to perturbations while keeping policy optimization stable.
In briefCounterexamples showing better local reconstruction doesn't guarantee better final model fidelity
SummaryThis paper studies residual completion, which estimates the contribution of tokens omitted from sparse attention. Using frozen language models with training-free RESA and learned Top-K+φ, the authors find cases where local attention-output reconstruction improves but final KL fidelity is worse than the Exact Top-K baseline. Exact restoration at the same layer improves final fidelity instead.
In briefQuantile head for robot VLA models that predicts several action quantiles in one pass
SummaryThis paper addresses limitations of action heads in Vision-Language-Action models used for robot control: point regression gives only a point estimate, and standard flow-matching needs costly iterative sampling. The authors unify regression and flow matching under a shared objective, derive a quantile objective, and build a Quantile Head that predicts a median and positive gaps to form ordered action quantiles in one forward pass. The jointly supervised median policy reports the highest average success rates among compared methods on LIBERO, LIBERO-Plus, LIBERO-Pro, and two real-robot tasks.
WhyUnifies regression and flow matching into a quantile objective for a one-pass action head, validated on several LIBERO benchmarks and real-robot tasks with comparisons.
In briefKVCMAS: An efficient KV cache correction method for shared context in multi-agent systems
SummaryIn multi-agent systems, agent-specific prefixes change the KV cache built from the same shared context, causing repeated prefilling and high compute and memory costs. KVCMAS represents cross-agent cache deviations with compact low-rank states and chains corrections along the agent workflow without an extra reference prefill. Across language and vision-language workloads it matches or improves accuracy over prior cache-sharing methods and achieves the lowest TTFT under high concurrency. In controlled serving traces it gives a 2.0x TTFT speedup over inference without cache sharing and cuts peak GPU memory by up to 3.7x versus a prior correction method.
In briefDo World Models Learn Global Understanding? A Study of Constraint Propagation in Monoid Worlds
SummaryThe paper frames understanding as learning constraints from local transitions and propagating their consequences to unseen facts. Using monoid worlds, the authors test inverse, commutativity, composition, and periodicity constraints across attention, recurrent, and state-space models. Standard next-state training fails to propagate non-trivial constraints, while compositional training that hides intermediate states reaches 96% accuracy and improves generalization in embodied environments and Wikidata-finetuned LLMs. Generalization drops sharply with proof depth, and longer compositional paths help.
In briefSteering Language Model Goals by Transplanting Value Signals
SummaryThe paper studies whether a model's internal value-axis signal can redirect which goal it pursues. The authors fine-tune Qwen3-8B and GPT-OSS-20B into honest and cheating variants, then shift a host model's activations along a candidate value axis toward a donor's value coordinates at each token. The intervention works in both directions, changing test-gaming behavior, and on solvable coding tasks an honest donor also improves a cheating host's hidden-test performance. The authors report preliminary evidence that the method works across model families.
In briefPReCache: Efficient KV Cache Sharing for Multi-LoRA Agents
SummaryThis paper addresses the redundant memory and computation that arise when multiple LoRA-adapted agents sharing one backbone repeatedly process the same growing trajectory in long tasks. It proposes PReCache, a training-free KV cache sharing framework with two designs: PreLRShared precomputes each agent's compact low-rank cache to avoid repeated prefill, and ReBaseShared rebuilds the shared base cache from adapter-free hidden states to improve accuracy. Across several models and agent benchmarks, PreLRShared reports up to 3.1x TTFT speedup and 2.3x per-request throughput gains, while ReBaseShared shows an average accuracy drop of only 1.1 points versus no cache sharing.
In briefThe Statistical Cost of Causal Discovery with Feedback
SummaryThis paper studies how many samples are needed to learn cyclic causal structure from observational data in linear non-Gaussian models. It recovers strongly connected components and the edges between them, and derives the first information-theoretic sample lower bounds for this task. The authors propose BlockExo, which matches the structural bound under suitable conditions without knowing key structural parameters, and simulations support the results.
WhyGives first information-theoretic sample lower bounds for SCC recovery in cyclic linear non-Gaussian models, with a residual-independence identification principle; experiments are simulations only.
In briefThinking Outside the Box: Information Retention in Sliding-Window KV Inference
SummaryThis paper studies whether information from outside the current window can persist in a fixed-size rolling key-value cache when pretrained causal transformers process text incrementally at inference time. Experiments on five open-weight models (Qwen, Llama, Mistral, and Muse Glimmer) show that keeping previously computed states improves retrieval compared with recomputing the final window from raw tokens. Muse Glimmer and Mistral 7B show the strongest latent information relay, recovering information after its source tokens have left the cache.
In briefARCH-B: A Benchmark for Architectural Representation and Cross-Media Recognition
SummaryThis paper introduces ARCH-B, a benchmark of 354 four-choice questions testing whether multimodal models can recognize the same building across photographs, floor plans, elevations, sections, and renderings. The questions span 11 cross-representation categories and were built from 3.9 million building-linked architectural images. Across 25 models, accuracy ranged from 10.45% to 83.90%, compared with a 35.35% human baseline from 5,830 non-expert responses, with models weakest at floor plan-to-photograph matching.
Kieran Sagar Parikh, Jose Luis Garcia del Castillo y Lopez
In briefKafila: running large language models across a trusted group of mixed everyday computers
SummaryThe paper presents Kafila, a system that links several consumer computers owned by a small trusted group into a ring so they can together run a language model too large for any one machine. Its planner measures each device's memory bandwidth, capacity and reachability, then splits the model and places the output layers before serving starts. The authors report that it shortens the slowest pipeline stage compared with even and memory-proportional splits, and that gains in throughput depend on how much time goes to computation rather than network.
Murtaza Rangwala, Richard O. Sinnott, Rajkumar Buyya
In briefSCOPD: Sparse-Context On-Policy Self-Distillation for Efficient Vision-Language Models
SummaryReasoning vision-language models process many visual tokens, which makes inference costly. The paper proposes SCOPD, where a student model reasons from pruned visual tokens while a full-context teacher supervises the same on-policy prefixes, requiring no ground-truth responses or extra inference computation. At 10% visual-token retention, the Vanilla model keeps 86.37% of unpruned performance across 13 benchmarks, rising to 90.49% with SCOPD and 92.43% with SCOPD+.
Ahmadreza Jeddi, Enming Zhang, Jasper Gerigk et al.
In briefSingularities of Non-negative Matrix Factorization and Their Use in Bayesian Inference
SummaryThis paper studies the local geometry of non-negative matrix factorization (NMF), a singular statistical model, and derives an upper bound on its real log canonical threshold (RLCT). Assuming the true matrix has a strictly positive factorization with inner dimension H0, the RLCT is shown to be at most {(H-H0)min(M,N)+H0(M+N-H0)}/2. When H equals H0 and the true matrix's ordinary rank, the exact value H0(M+N-H0)/2 is obtained. These results bound leading coefficients of Bayesian generalization error and free energy.
In briefVision-language signals in safe RL: safety gains without collision prediction
SummaryThis paper proposes VLM-Safe-RL, which adds frozen CLIP vision-language signals to PPO-Lagrangian through reward shaping and an augmented multiplier update for safety-constrained driving. On MetaDrive Hard, the catastrophe rate drops from 31.6% to 19.4%. However, the analysis finds no evidence that the CLIP signals anticipate collisions, and the VLM term has a negligible effect on the Lagrange multiplier.
In briefDual-agent LLM framework (CLEAR-Med) for traceable analysis of structured clinical data
SummaryThe paper presents CLEAR-Med, a framework in which one agent turns natural-language questions into executed SQL queries, and a separate validation agent accepts, repairs once, or abstains. On a 25-query development benchmark over a 532-record neonatal hypoxic-ischemic encephalopathy table, the query agent answered 66.4% of responses correctly, versus 12.0% for an ungrounded ChatGPT baseline. A prospective evaluation of the full validation and abstention pipeline is left as future work.
Erfan D. Dehkalani, Seetha Shankaran, Abbot R. Laptook et al.
In briefUOPD: An uncertainty-aware method that lets a teacher step in during on-policy distillation for multi-turn AI agents
SummaryThe paper proposes UOPD, a method for on-policy distillation of multi-turn agents. When the teacher is uncertain about a student's action, the teacher's action is executed and the student is trained to imitate it via supervised fine-tuning; otherwise the standard OPD loss is used. Evaluated on ALFWorld, WebShop, and Search, UOPD reportedly improves WebShop score by up to 15.8% over standard OPD.
In brief3D Point Tracking with State Space Models: Tracking Scene Points in Absolute Meters on a Single GPU
SummaryThe paper proposes a monocular, pose-free 3D point tracker that tracks points of a dynamic scene in absolute metric units. It composes frozen dense optical flow for 2D correspondence and a monocular metric-depth network, then learns the depth residual with a compact Mamba-3 state space model conditioned on DINOv3 appearance features. On the TAPVid-3D minival benchmark, the best configuration reaches a mean metric Average Jaccard of 0.256, the highest among methods evaluated under identical conditions.
In briefADPTNet: A Non-Linear State-Space Model for Sequence Modelling
SummaryThis paper proposes ADPTNet, a sequence model intended as an energy-efficient alternative to Transformers that aims to be data-adaptive, capable of long-range memory, GPU-parallelisable, and non-linearly recurrent. It combines linear attention with Riemannian optimisation to build local topological conjugates, with dynamical systems proofs guaranteeing control of its timescales. Reported results include gains over Hawk and Mamba on selected tasks, and a spiking variant reaching 83.56% on Spiking Speech Commands.
In briefConformal Factuality Control for Multi-Hop Retrieval-Augmented Generation
SummaryThis paper tests whether claim-level conformal factuality control, designed for retrieval-augmented generation, still works when retrieval and reasoning span multiple dependent steps. The authors evaluate it on HotpotQA, Natural Questions, and TriviaQA using Llama 3.1 8B and GPT-4o-mini. Stricter targets raise the share of responses whose retained claims are fully supported, but only a small fraction of claims is kept, so reliability must be read together with retention and abstention.
In briefFaithful Activation Verbalization: Cutting Hallucinations When Explaining LLM Internal Representations
SummaryThe paper introduces AVPO, a two-stage method that first reconstructs source text from a model's hidden activations, then checks that text with a separate frozen question-answering model. The reconstructor is trained with direct preference optimization using rewards for semantic recoverability and lexical fidelity. Across six text families, AVPO improves gist-level and detail-level information recovery over the strongest baseline by up to 17.1 and 9.3 percentage points. The authors also report that on out-of-distribution cases it recovers high-level semantics while fabricating fewer details.
In briefRe:Cognize: A Framework for Open-Set Comic Character Re-Identification
SummaryThis paper introduces Re:Cognize, a framework in which models recognise comic and manga characters as the story unfolds, learning identities along the way rather than using a fixed cast. The authors evaluate five Re-ID backbones and their memory-augmented baseline MECHA across three manga corpora. They find that updating the gallery from the model's own predictions degrades performance, and propose VOUCH, which updates memory only when the page provides supporting evidence.
Aaditya Baranwal, Madhav Kataria, Yogesh S Rawat et al.
In briefPosterior Regimes and Latent Deception: Variational Bayesian HMMs for Sequential Fraud Detection
SummaryThe paper presents three Hidden Markov Model variants (Baum-Welch, VBEM, and a neural variational extension) that model each customer's transactions as movement through latent behavioural regimes, one identified as fraud-associated. On IEEE-CIS transaction data, the neural version reaches 14.4 times fraud enrichment in its identified regime. Its AUPRC trails an XGBoost baseline, which the authors argue is structural, so the model is best suited as a calibrated triage and interpretability layer.
Joseph Uririoghene Obukofe, Anthony O'Hare, Chioma Sandra Dike
In briefFinding repeated concepts in audio explanations to reveal shortcut learning in classifiers
SummaryThe paper proposes a pipeline that isolates audio segments explaining a classifier's decisions, captions them with an ensemble of Large Audio-Language Models, and uses a Large Language Model to extract recurring concepts. These concepts can be audited by humans to uncover potential shortcut learning. Evaluated on datasets curated from AudioSet Strong with controlled spurious correlations, the approach reliably uncovers learned shortcuts, such as a model relying on laughter to predict applause.
In briefFast Integration of Tensor Fields on Trees via Structure-Adaptive Methods
SummaryThis paper introduces Structure-Adaptive Tree Field Integrators (STAD-TFIs), a class of near-linear algorithms for integrating general tensor fields on trees with distance-dependent kernels. The method uses decompositions based on path backbones and single-vertex separators, together with two-dimensional fast Fourier transforms, to compute interactions more efficiently than standard tree field integrators. The authors provide theoretical analysis and empirical tests on synthetic trees, accelerated Sinkhorn relaxations of optimal transport on real meshes, and topological attention transformers for vision tasks.
In briefJev in Medicine: Benchmark Tests of a Fast Model That Answers Only From Preset Options
SummaryThe researchers evaluated Jev 1.13, a non-generative model that picks from predefined answers, on four medical benchmarks covering question answering and case-based diagnostic reasoning. GPT-6 Sol, with and without reasoning, served as the reference. Jev matched GPT-6 Sol with medium reasoning on PubMedQA but scored clearly lower on MetaMedQA, DiagnosisArena-MCQ and NEJM cases. The authors state that task-specific validation is required before any clinical use.
In briefLightweight test-time adaptation for vision-language models using position-aware layer queries
SummaryThis paper proposes Layer Query Network (LQN), a lightweight method that adapts frozen vision-language models to out-of-distribution test samples in a single forward pass. It uses Position-Aware Distillation to mimic intermediate-layer spatial tokens and Location Consistency Regularization to replace costly image augmentation with coordinate sampling. The authors report a 9.8% Top-1 gain on OOD ImageNet for zero-shot CLIP ViT-B/16, along with improvements in fine-grained classification, faster convergence, and extension to segmentation tasks.
In briefSR4-Fit: An Interpretable Rule-Based Machine Learning Framework for Trustworthy Decisions
SummaryThe paper proposes SR4-Fit, an algorithm for classification and regression that is interpretable by design and produces compact, stable rule sets. Using U.S. Census Bureau American Community Survey demographic data, it predicts U.S. House election outcomes. Across fourteen benchmark datasets, it outperforms RuleFit and decision trees in accuracy, stability, and compactness while remaining competitive with black-box models.
Shyam Sundar Murali Krishnan, Dean Frederick Hougen
In briefWasserstein causal forests for estimating effects on whole outcome distributions
SummaryThis paper proposes Wasserstein Causal Forests (WCF) for settings where each unit's outcome is a probability distribution, and defines finite-grid treatment effect measures, including a reference-distance contrast. In simulations, WCF is most accurate on the conditional-law metric in most designs but is less accurate than forest baselines for multimodal settings. Applied to Project STAR, small classes raise within-grade math achievement by 0.161 standard deviations on average, with larger gains in the upper part of the score distribution, while overall dispersion is essentially unchanged.
In briefEstimate, Don't Imitate: Reusing Expert Policies for Camera-Based Robot Control
SummaryThis paper addresses deploying simulation-trained manipulation policies that must act from partial observations such as noisy camera images. Instead of training a visual policy to copy a privileged expert, it freezes the differentiable state-based expert and learns only a visual state estimator, supervised by state error and an action-consistency loss backpropagated through the expert. Across five goal-conditioned manipulation tasks, retaining the expert outperforms direct pixel-to-action imitation, and on a physical Panda robot it reaches 76% success without retraining the expert.</summary>
Denis Shcherba, Adrian Abel, Eckart Cobo-Briesewitz et al.
In briefMaat: Deterministic Contract-Based Oversight for Multi-Agent LLM Workflows
SummaryThe paper presents Maat, a runtime layer that checks handoffs between LLM agents against a versioned workflow contract, using no language model in validation or scoring. Tested across six controlled domain workflows with 522 trials, it shows gains on contract-detectable defects in most workflows. The authors also report that 37% of governed halts were false alarms caused by validator defects, and that validator configuration and halt attribution must themselves be tested.
In briefA computer vision method using YOLOv8 to detect visual fraud in phishing websites
SummaryThis paper proposes an image-based approach to phishing detection that judges rendered webpage screenshots as a human eye would, rather than relying on URLs, HTML, or domain data. A YOLOv8 convolutional neural network classifies screenshots as phishing or legitimate based on layout, logo placement, color scheme, and login-form structure. The system reached 92% accuracy on a held-out test set and processed a screenshot in about 100 milliseconds. After data augmentation targeting lighting, compression, and scaling variation, the false-positive rate fell by 11% relative to the pre-augmentation baseline.
In briefLTV-CTDNet: A Forecasting Model for Short-Term Turning Counts
SummaryThis paper presents LTV-CTDNet, a model that forecasts short-term turning-movement traffic counts at intersections to support signal control and corridor operations. It separately predicts nonnegative approach totals and within-approach turning proportions, then reconstructs movement forecasts from these components. Evaluated on seven months of 15-minute LiDAR data from eight locations in Nashville, Tennessee, it reached a movement-level MAE of 1.8189 and RMSE of 3.8072, and produced no negative forecasts, whereas unconstrained models generated negative values in about 10.6% to 29.2% of raw cells.
Md Atiqur Rahman Mallick, Kamrul Hasan, Robert T. White
In briefWhen Known Physics Helps Neural PDE Models: Residual Constraints Beat Generic Priors on Nonlinear Dynamics
SummaryThis paper compares several structural priors for neural PDE surrogate models under a common protocol. At equal tuning budget, a known-equation residual constraint outperforms the best generic regularizer, with the advantage persisting for nonlinear equations such as Burgers, KdV, and Allen-Cahn but vanishing for linear heat and advection-diffusion. Under grid under-resolution, enforcing the residual becomes harmful, while cross-family pretraining and in-context conditioning did not beat the from-scratch baseline.
In briefZeroBot: Robot Learns Manipulation from Scratch in Minutes via Generative Real2Sim
SummaryZeroBot is a framework that lets a robot learn a manipulation task from scratch in minutes, without human demonstrations, pre-trained policies, or known object models. From a single view of an object and a goal pose, it uses image-to-3D generative models to create a complete mesh for large-scale parallel reinforcement learning in simulation. On real-world tasks including grasping, pushing, articulated object interaction, and multi-stage manipulation, it reports an 87% success rate with an average training time of 119 seconds.
Ivan Kapelyukh, Xiaohan Zhang, Stephen James et al.
In briefFIRE: Stabilizing Feedback-Based Self-Distillation for LLMs via Fisher-Informed Recalibration
SummaryThis paper addresses the unstable optimization of feedback-based on-policy self-distillation, where a language model acts as both teacher and student and may suffer performance collapse during training. The authors propose FIRE, a dual-branch framework that uses re-weighted on-policy SFT for correct outputs and recalibrates feedback-conditioned targets for incorrect ones. According to the authors' experiments, FIRE yields considerably more stable self-distillation while maintaining strong downstream performance.
Seohyun Lee, Dong-Jun Han, Seyyedali Hosseinalipour et al.
In briefEHRAdapt: Adapting language models to patient records using semantic priors for rare clinical events
SummaryThis paper introduces EHRAdapt, an adapter that maps electronic health record event tuples of time, modality, and code directly into a frozen language model's embedding space, avoiding long text serialization. Each event vector combines a frozen biomedical semantic prior with a learned low-rank residual to handle rare events with few observations. Trained on records from about 4 million patients with only the adapter parameters (0.1% to 0.6% of the total) updated, the full model outperforms its ablations on held-out next-event prediction across three backbones and also performs better on infectious-disease and syndromic classification.
