In briefUsing 2D foundation model priors to discover new object categories during test-time adaptation in 3D LiDAR segmentation
SummaryThis paper addresses recognizing previously unseen object categories in 3D LiDAR semantic segmentation for autonomous driving, where models must adapt at test time. It proposes a two-stage framework: the first stage aggregates local 3D features by geometric density and generates reliable pseudo labels under a global class-wise prototype consistency constraint. The second stage formulates category discovery as an optimal transport assignment and fuses semantic features from 2D vision foundation models with 3D features to create a more separable feature space.
In briefWave Physics as Representation: Tensor-Structured Neural Reconstruction for Ocean Acoustic Fields
SummaryThis paper addresses reconstructing three-dimensional ocean acoustic fields from sparse measurements. The authors derive a structured spatial-angular tensor representation (the S/c representation) from the Helmholtz equation and normal mode propagation, and use it as an inductive bias in TIDAL, a neural reconstructor. TIDAL does not require the known environmental model parameters that conventional physics-based methods need, and it achieves the best reconstruction performance and fastest inference across three structured sampling geometries.
In briefXiaomi Auto World Model 2.0: A Causal World Model for Driving and Embodied AI
SummaryThis paper presents Xiaomi Auto World Model 2.0, a unified streaming causal world model for Physical AI that generates real-world physical evolution. It combines View-Temporal Attention with multi-view rotary position encoding, and uses WorldRec 2.0 to reconstruct 3D Gaussians from multi-view, multi-temporal observations as explicit scene context. The model is trained in pre-train, mid-train, and distillation-based post-train stages, and is applied to closed-loop simulation, closed-loop training, and world action model tasks for state prediction and action generation.
In briefMMCoT³: Interleaving 3D Scene Evidence with Text for Better 3D Reasoning
SummaryThis paper introduces MMCoT³, a framework that interleaves textual reasoning with task-relevant 3D scene evidence so the model can check observations during multi-step reasoning. To handle cases where one textual mention matches several candidate objects, it uses an ambiguity-aware evidence aggregation mechanism built on query-guided cross-attention. The authors also release PointCoT, a 3D chain-of-thought instruction tuning dataset of about 176.2K instances, and report state-of-the-art results across six benchmarks.
In briefLearned optical compression for single-pixel phase imaging
SummaryThis paper presents a single-pixel quantitative phase imaging method that jointly optimizes a diffractive encoder and a digital neural decoder to compress reconstruction-relevant phase information into a compact measurement vector. A reconfigurable phase-modulation plane displays object-independent encoding patterns in sequence, each producing one scalar intensity measurement. Experiments at a visible wavelength achieve about 40-fold dimensional compression, using only 2.55% as many intensity measurements as reconstructed pixels.
WhyThe authors propose jointly optimized diffractive encoding and neural decoding for single-pixel quantitative phase imaging, reporting experiments on handwritten digits and grating phase objects with c
In briefHow much random rotation improves GPTQ: sharp worst-case rate gaps at high rate
SummaryThe paper studies the rate gap between GPTQ, a Cholesky-based Babai nearest-plane quantization algorithm, and WaterSIC in the high-rate, entropy-coded regime. It shows that after a Haar random rotation, the k-th pivot converges almost surely to a classical Lyapunov-exponent formula, which yields an asymptotic expression for the worst rotated rate gap. The authors conclude that rotation cannot in general replace waterfilling, though it nearly closes the gap on real layers.
WhyThe authors characterize rotated GPTQ pivots via a classical Lyapunov-exponent formula, deriving a tight worst-case rate gap; they report validation on 96 Qwen3 and Llama-3.1 Hessians, mainly affectin
In briefTesting whether biological world models use actions, using design-based methods
SummaryAction-conditioned world models are often trained on randomized experiment data, and their dependence on actions is typically checked by removing actions and measuring the accuracy loss. The paper shows that under autocorrelated designs such diagnostics can report action effects that do not exist. It proposes a conditional randomization test written as an e-value and analyzes how pooled scores over many sites sharing one action sequence can amplify small model errors.
WhyThe authors show action-perturbation diagnostics yield false effects and propose design-based e-value tests with pooling amplification, reported on synthetic and optogenetic ERK experiments.
In briefGatedVLA with Camera-Based Emergency Braking: A Technical Report for the AlpaSim End-to-End Closed-Loop Challenge
SummaryThis report describes GatedVLA, an entry in the Physical AI autonomous driving track. It pairs a vision-language-action policy with a flow-matching action expert that outputs 2.5 seconds of acceleration and curvature controls, using a soft visual mask on the expert's cross-attention. In local closed-loop tests on 150 scenes, the visual mask raised the mean score from 0.479 to 0.517. Adding a rule-based automatic emergency braking module raised it to 0.599 and reduced the average number of scenes with at-fault collisions from 34.0 to 11.0.
In briefA hybrid method for detecting unusual patterns in multi-sensor time series data
SummaryThis paper proposes a hybrid deep learning method for detecting anomalies in multivariate time series, such as data from power plants, smart houses, and servers. The method captures both temporal and spatial information by combining a variational autoencoder with 2D convolutional layers and a prediction-based network using 2D convolutions and a gated recurrent unit. The authors report experiments on public data against state-of-the-art methods, showing the approach's potential.
In briefA hybrid anomaly detection method for identifying electric vehicles from smart meter data
SummaryElectric vehicle charging creates load imbalance and voltage instability in power grids, so distributors need to identify which users own an EV. This paper frames the task as unsupervised learning and proposes a hybrid anomaly detection method that combines prediction-based and reconstruction-based components. The model captures temporal and stochastic representations of smart meter time series. The authors report that it outperforms state-of-the-art supervised and unsupervised methods on real-world data.
Ali Babalhavaeji, Hui Song, Mohammadreza Radmanesh et al.
In briefEEG-HealthLLM: Adapting large language models to assess health from diverse brain signal recordings
SummaryThis paper proposes EEG-HealthLLM, a unified and parameter-efficient framework that applies large language models to health assessment from electroencephalography (EEG) signals with different configurations. It uses task-specific EEG adapters to preserve each task's spatio-temporal characteristics, plus a shared LoRA injected into a backbone such as Phi-3 to capture common high-level assessment semantics. The authors also build EEG-HealthBench by reorganizing five existing EEG resources into subject-independent tasks. The method reaches an average balanced accuracy of 69.58% while reducing the language-side adapter footprint by 80.0%.
In briefOne-step physics-informed generation via long-short flow maps
SummaryThis paper proposes LSFM-DG, a physics-informed generative model that splits one-step generation into a long flow map for global evolution and a near-terminal short flow map with closed-form solutions and physics potentials. As the short flow map shrinks to zero, it supplies direct distributional and physical targets for learning the long flow map. The authors report stronger data fidelity and physical consistency without distillation.
In briefNear-optimal √K regret for online learning in games where opponent actions are hidden
SummaryThis paper studies online learning in two-player Markov games where the learner cannot observe the opponent's actions and the opponent may use any unknown policy in each episode. The authors present an algorithm that achieves Õ(H^{3/2}S√(AK)) empirical Nash-value regret against arbitrary opponent policies, settling an open question about whether the √K rate holds in the worst case. The method combines global and local planning, coordinating them to control the gap between local loss estimates and the total loss of the executed policy.
WhyThe authors resolve an open question on worst-case √K regret, proposing an algorithm combining global and local planning with an H^{3/2}S√(AK) bound.
Junyan Liu, Haipeng Luo, Lillian J. Ratliff et al.
In briefdKV Forcing: A Differentiable KV Cache That Separates Generation and Memory Representations in Video Models
SummaryThis paper studies causal video generators, which produce two key-value states per segment: one for denoising the current segment and another written to the cache for later segments. The authors propose dKV Forcing, which learns a separate memory representation through a writer-specific low-rank residual trained on how future segments read from the cache. Compared with Self Gradient Forcing, the method raises VBench-Long aesthetic quality by 6.6 points at 60 seconds and 8.5 points at 240 seconds while keeping subject and background consistency high.
In briefFair Ranking for Value-Weighted Account Blocking
SummaryThis paper studies how to rank accounts for risk screening and recommendation when blocking is limited by a budget on forgone value, balancing loss capture against group disparities in complaints. It proposes weighted binary cross-entropy and group-scaled complaint penalties, evaluated on public recommendation, insurance, credit, and mortgage datasets. At a fixed minimum retained loss capture, the source-level results do not establish a statistically significant average advantage over group-blind pricing.
In briefUA-Depth: Uncertainty-aware depth repair for handling any material in robot grasping
SummaryUA-Depth is a framework that repairs depth maps from RGB-D cameras, which often fail on transparent, reflective, and translucent objects. It iteratively refines depth and per-pixel uncertainty together, using a frozen depth foundation model as a prior for missing depth. Across eight zero-shot datasets it achieves the best mean rank among representative baselines, and the predicted uncertainty also helps category-level 6D pose estimation and 6D grasp synthesis.
In briefCREDO: A method that assigns credit to intermediate evidence steps for temporal question-answering agents
SummaryThis paper addresses temporal knowledge graph question answering, where language model agents retrieve evidence and reason over multiple steps. It proposes CREDO, which uses intermediate evidence to provide local supervision and redistributes outcome-policy weight toward reasoning and answering, even when all rollouts in a group fail. The authors report higher accuracy than the strongest published baseline on MultiTQ and than their reproduced baseline on TimelineKGQA.
In briefSharp generalization analysis of shallow overparameterized neural networks with channel attention for nonparametric regression
SummaryThe paper studies nonparametric regression using an overparameterized two-layer ReLU network with fixed covariance-based channel attention, with both layers trained by gradient descent. It proves that early-stopped training achieves regression risk O(ε_n²), where ε_n is the critical rate of the attention kernel. Under polynomial eigenvalue decay, this bound reaches the minimax optimal rate. The analysis is extended to covariates supported on R^d, achieving a nearly optimal risk under sub-Gaussian covariates.
In briefAnswer Locally or Escalate? Post-Training Expert Routing for Real-Time Duplex Speech Models
SummaryThis paper studies how a real-time full-duplex speech model can decide, at the moment it starts speaking, whether to answer a question itself or hand it to a stronger expert. The authors freeze MiniCPM-o 4.5, use only its speech input and output, and fit a logistic failure readout on its existing hidden states at the listen-to-speak transition. This gate needs no extra generation and its thresholds can be retuned without changing the duplex policy. On an internal benchmark, selective handoff improves answer accuracy by 27.1 points.
In briefTesting AI vision-language models on handwritten Urdu Nastaliq script
SummaryThis paper compares six systems without fine-tuning on 2,465 single-line images from the Urdu Nastaliq Handwritten Dataset, including Tesseract, a printed-Urdu recognizer, and open-source and commercial vision-language models. Both commercial models show lower error than the printed-Urdu recognizer and the open-source models. A second study examines whether these models can correct the output of another OCR system, with gains depending on how legible the supplied text already is.
In briefBanglishDepYT: Depression detection in native Bengali social media
SummaryThis paper introduces BanglishDepYT, a Bengali depression dataset collected natively from YouTube, with 191,000 unlabeled comments and 1,000 dual-annotated samples. The authors propose EPH, an emotion-preserving preprocessing method that keeps emoticons, Banglish tokens, and negations, and show it markedly improves performance on informal text. They also apply domain-adaptive pretraining to SahajBERT, which achieves the highest mean F1 and MCC, and use LLM-assisted pseudo-labeling with semi-supervised training to expand annotations.
WhyThe authors combine emotion-preserving preprocessing with domain-adaptive pretraining, reporting comparisons and significance tests on native Bengali YouTube comments.
In briefZero-shot testing of open vision-language models as fast decision models for radiology tasks
SummaryThis paper compares two multimodal "System 1" decision models, Clef 27B and d1 3B, with a conventional LLM, Qwen 3.8-27B, across seven radiology tasks involving reports, radiographs, CT and MRI, without any task-specific fine-tuning. All models extracted findings from reports with AUROC above 0.95, and Clef outperformed its Qwen backbone on some image tasks. The smaller d1 model ranked below Clef on six of seven tasks while using 3 to 9 times less energy per request. The authors conclude that System 1 models can handle these radiology tasks accurately at lower compute cost, but no model performed well on judging motion-artifact severity.
In briefEgoMo3R: Joint Reconstruction of Egocentric Human Motion and Scene
SummaryThis paper presents EgoMo3R, a joint optimization framework that reconstructs 3D human motion and scene geometry from casual egocentric video. It combines low-level visual-geometric cues such as monocular depth, optical flow, and point tracks with a diffusion-based human motion prior, using an alternating optimization scheme in which each component refines the other. The authors evaluate it on the highly dynamic EgoExo4D and Nymeria datasets and report that it outperforms specialized baselines in both scene reconstruction and egocentric motion estimation.
In briefA paper defining the scope of computational aesthetics, split into special and general levels.
SummaryThe paper clarifies the scope of computational aesthetics, a field combining computer science, AI, and aesthetics. It proposes two levels: special computational aesthetics (SCA), focused on aesthetic measurement and process modeling, and general computational aesthetics (GCA), which adds artistic generation guided by computational aesthetic models. The authors also distinguish the field from related concepts such as AIGC, AI art, and computational creativity.
In briefFinForesight: A Diagnostic Backtesting Benchmark for Financial Forecasting Research Systems
SummaryFinForesight is a diagnostic backtesting benchmark for evaluating deep research systems on financial forecasting across Chinese, U.S., and Hong Kong markets. It covers four forecasting tasks (valuation, return ranking, event direction, and future fundamentals) plus two capability tasks on counterfactual reasoning and financial calculation. Retrieved evidence is restricted to information available on each forecast date. The authors find that pretraining-era scores usually exceed post-cutoff scores, and stronger retrieval controls narrow the gap between post-cutoff backtests and live results.
In briefFocusing on Key Image Regions and Adapting Verification to Load: Load-Adaptive Speculative Decoding for Vision-Language Models (Uni-LaSpec)
SummaryThis paper presents Uni-LaSpec, a block-parallel speculative decoding framework for vision-language models. It reuses the verifier's attention to select a few key visual tokens for the drafter and chooses the verification structure per request based on GPU load. Across six VLMs and 17 benchmarks, selecting eight visual tokens retains at least 98.6% of full-image acceptance, and at concurrency one with greedy decoding it reaches up to 5.11x text and 3.24x multimodal throughput over autoregressive decoding.
In briefMultimodal house price prediction combining listing attributes, text, and photos
SummaryThe paper proposes a multimodal framework that combines structured property attributes, textual listing descriptions, and property images to predict house prices. A large language model extracts spatio-temporal and contextual features from text, while an unsupervised vision-language method groups photos into classes such as indoor, outdoor, mixed, or floorplan without manual labels. The features are fused into a unified representation and fed to an optimized regressor. Experiments on a large real-world property dataset show better results than traditional methods and existing multimodal baselines, and ablation studies support the contribution of the textual and categorized visual features.
Khandokar Md. Rahat Hossain, Souvik Das, Tahsina Hashem et al.
In briefSonny-ENS: a machine-learning ensemble weather forecasting model explained
SummaryThis paper presents Sonny-ENS, an ensemble of three 68.9-million-parameter neural networks using Monte Carlo dropout, trained on the continuous ranked probability score in 2.7 A40 GPU-days. On 2022 data, it beats the ECMWF ensemble on 96% of 570 targets and has 8.3% lower mean CRPS at lead times of 1 to 10 days. The authors also find that dropout adds uncertainty mainly in the tropics, and that forecast spread tracks baroclinic instability in the atmosphere.
WhyThe authors report that a dropout-based ML ensemble trained in 2.7 GPU-days beats ENS and GenCast on 2022 targets, with ablations and sparse-autoencoder analysis of spread sources.
In briefSurrogate losses for fairness-aware classification: consistency and calibration
SummaryThis paper examines whether minimizing surrogate loss functions yields optimal decisions for fairness-aware classification under demographic parity and equal opportunity. The authors define fairness-aware Fisher consistency and characterize it through conditions on the surrogate losses' derivatives. They also derive a calibration inequality linking excess fairness-aware risk to excess surrogate risk. An application to consumer credit default prediction shows that appropriately calibrated fairness regularization can substantially reduce demographic disparities while maintaining predictive performance.
In briefHalal-NeSyNet: Clause-Scoped Neuro-Symbolic Model for Halal Food Ingredient Auditing
SummaryThe paper introduces Halal-NeSyNet, a model that judges whether packaged foods are Halal from ingredient declarations and E-number additives. It combines a DistilBERT text encoder with a rule-derived symbolic vector through cross-attention and gated fusion, trained with an asymmetric loss that penalises false Halal labels. Evaluated on a 3,199-item benchmark, it outperforms text-only baselines, but because labels and features come from one rule engine, the results measure fidelity to a rule policy rather than religious certification.
In briefNatureFusion: Reframing infrared and visible image fusion as natural complementary enhancement
SummaryThis paper proposes NatureFusion, a method for fusing infrared and visible images that treats fusion as producing a natural, clear visible-style image rather than copying both sources. It uses a generative foundation model as a teacher to create supervision, fine-tunes a pretrained generative model with conditional flow matching via a lightweight fusion adapter, and adds a detector-supervised task adapter to strengthen infrared's complementary role. Comparisons with fourteen other methods show the best perceptual quality, and detection and segmentation models pretrained on visible images achieve higher accuracy on its outputs.
In briefMedNet and U²-ViT: A Nested Vision Transformer Diffusion Framework for Generating Diverse Medical Images
SummaryThe paper addresses the shortage of diverse, high-quality medical images by introducing MedNet, a multi-modal text-to-image dataset covering 61 disease categories and 160 classes across humans and animals. The images are drawn from trusted public repositories. The authors also propose U²-ViT, a diffusion-based text-to-image architecture that nests Vision Transformers within a U-Net structure. Experiments report that it achieves the best Fréchet Inception Distance (FID) score compared with state-of-the-art models.
In briefDyRadio: a spatiotemporal diffusion method for rebuilding dynamic radio maps
SummaryThe paper presents DyRadio, a framework that reconstructs sequences of radio maps in dynamic environments from sparse measurements taken at different times. It uses a propagation-prior-guided, noise-conditioned spatiotemporal diffusion model to jointly reconstruct consecutive maps and an iterative predictive likelihood method to estimate unknown measurement noise. Experiments report higher reconstruction accuracy and noise robustness, with PSNR gains of 1.18 to 5.75 dB over the best baseline across all evaluated settings.
In briefWhen Sparse Routing Helps: How Expert Separation Shapes Mixture-of-Experts Recovery
SummaryThis paper studies how expert separation and routing sparsity jointly affect expert recovery in sparse mixture-of-experts models. It analyzes location Gaussian mixtures of experts with a fixed number of experts, showing how separation geometry differs for dense and sparse models. With Top-1 routing, experts can be recovered at a rate independent of the separation quantity.
In briefAutomated photo-identification of fire salamanders from their spot patterns
SummaryThis paper presents a pipeline for identifying fire salamanders from photographs using zero-shot local-feature matching (SuperPoint and LightGlue), requiring no species-specific training. It was evaluated on 2,440 field photos from Haifa and Tel Dan in northern Israel, with yellow spot patterns isolated and similarity scored by high-confidence keypoint matches. On held-out individuals, the correct individual ranked first for 95.2% of queries, with an ARI of 0.958. Unmodified, it reached 95.1% and 91.7% top-1 accuracy on two European fire salamander datasets.
Maayan Yesharim, Hava Goldstein, Gabi Kolodny et al.
In briefSCOUT: Robot planning around hidden objects using imagined future scenes
SummarySCOUT is a planning framework that helps robots manipulate objects in cluttered spaces when foreground items block their view. It reconstructs hidden geometry using an object-centric scene representation, and an LLM proposes candidate actions that are evaluated with Monte Carlo Tree Search over a learned action-conditioned world model. On Reveal and Placement tasks, SCOUT outperforms VLA policies, embodied agents, and 3D world models, and it was also tested on a real robot in a cluttered shelf environment.
In briefAutoencoder-based deep clustering of 3D point clouds
SummaryThis paper studies deep clustering of three-dimensional shapes represented as point clouds, motivated by a problem in Generative Specification-Producing AIs (SPAIs). It uses a fully unsupervised pipeline in which a FoldingNet autoencoder is first trained for reconstruction, then K-means++ is applied to the resulting codewords. Clustering accuracy, normalized mutual information and adjusted Rand index are reported on ModelNet10 and ModelNet40 as a function of the number of clusters K, with t-SNE visualizations of the codewords and a discussion of how K is chosen.
Nader Alfares, Guangmingmei Yang, David J. Miller et al.
In briefEvaluating diffusion video inpainting by how recoverable each hidden pixel is
SummaryVideo inpainting for object removal is usually scored over the whole hole, which mixes copying visible background with generating new content. The paper proposes recoverability-stratified evaluation, using a point tracker to label each hole pixel by whether and how far away it is observed, and introduces training-free evidence-guided sampling (EGS) that blends tracked colors into a video diffusion remover. On DAVIS, the perceptual advantage of diffusion removers over flow propagation shrinks where content is visible a few frames away, and EGS closes 43% and 56% of the gap to a copy reference on held-out DAVIS when colors come from clean-video tracks. With tracks computed on the masked input, the method gave no gain on MiniMax-Remover and a loss on ROSE.
WhyThe authors propose recoverability-stratified evaluation and training-free evidence-guided sampling, reporting on DAVIS, real paired captures, and a pre-registered confirmation on two removers.
In briefSpending Bits Where Tiny Objects Are: Rate-Utility Coding for Aerial Detection
SummaryThe paper proposes rate-utility region coding (RUO) for bandwidth-limited aerial cameras, deciding for each region whether to drop it or send it at low or high quality based on its expected detection utility and measured byte cost, with one multiplier shared across images. On all 1,610 VisDrone-DET test-dev images, RUO uses 42.4% fewer bits than a tuned threshold-based ROI coder over the shared tiny-object average precision range. It also needs fewer bits for tiny objects than AVIF, HEVC and three published learned codecs.
In briefInvalid Code, Nearly Unchanged Score: A Replay Study of Judge-Based Evaluation
SummaryThis paper repeats an earlier evaluation on 421 coding problems, comparing the base Qwen-2.5-7B-Instruct with two GRPO-trained versions, one rewarded by execution tests and one by a strong language-model judge. The trained checkpoints produce invalid Python on 63% to 67% of attempts, yet the judge's Pass@32 score drops only about 3 percentage points. The authors recommend reporting the share of answers that pass Python's syntax check alongside any judge-based score.
In briefSearch-aware self-distillation with sibling alignment for generative recommendation
SummaryThis paper addresses generative recommendation, which retrieves items by autoregressively generating semantic IDs with beam search. The authors propose SA-OPSD, which transfers privileged self-teacher knowledge to states visited during the student's own beam search. Evolving-Boundary Sampling prioritizes search-critical states near the pruning boundary, while Confidence-Decoupled Branch Distillation separates target-probability calibration from relational supervision among competing branches. Experiments are reported on the Industrial, Office, and Instruments datasets.
In briefPrincipled human-AI collaboration for deciding which AI outputs can skip review
SummaryThis paper presents a statistical framework for deciding which AI-generated answers can be released without human review. Experts evaluate selected outputs before deployment, and their judgments calibrate the verifier threshold and number of attempts, without modifying the generator or verifier. In a MultiHiertt financial question-answering case study, the automation rate rose from 2.50% to 24.17% while the error rate among released outputs stayed around 1.20%.
WhyThe authors calibrate verifier thresholds and attempt counts from expert evaluations, giving finite-sample error control and a minimax lower bound. The paper reports automation gains on MultiHiertt fi
In briefBeyond Validation Loss: Coverage and Reliability in Robot Policy Pretraining
SummaryThis paper studies closed-loop pass@k for robot policies, the chance that at least one of k rollouts from the same start succeeds. The authors find that large-k pass@k reveals differences among posttraining recipes that pass@1 does not, and that pass@k peaks at different points during supervised pretraining, showing a coverage-reliability tradeoff. They propose an offline checkpoint-selection criterion and a generative data augmentation method that expands supervision beyond the demonstration set.
In briefMemory scaling laws and meta-memorization in transformers
SummaryThis paper studies how transformer models memorize data, finding that larger models can store more information, with capacity in bits scaling roughly linearly with parameter count, and that they memorize faster. The authors reproduce both scaling laws using random binary sequences with no structure. They also describe a "meta-memorization" effect, where models trained on many random sequences learn new ones faster, enabling even one-layer transformers to memorize a 100-bit sequence from few examples.
In briefChatDify: Automatically Building Dify Workflows from Plain-Language Descriptions Using Multiple AI Agents
SummaryThis paper introduces ChatDify, a multi-agent framework that turns natural language descriptions into complete Dify LLM application workflows. It uses an Actor-Critic design with a Planner, an Executor that synthesizes parameters and compiles DSL, and a Reviewer that verifies outputs and refines them in a closed loop. On 100 production-inspired tasks, it achieved success rates of 0.96, 0.91, and 0.75 with three different LLM backends.
In briefFinite-time sample complexity of online-trained conditional diffusion models
SummaryThis paper studies the finite-time training and sampling error of conditional diffusion models trained with online stochastic-gradient denoising score matching. It analyzes the score network after a finite number of gradient updates, separating approximation complexity from the online sample budget and propagating both errors through the reverse diffusion process. The authors derive a conditional score approximation error rate and a total-variation bound decomposed into initialization, time-discretization, approximation/localization, and finite-time optimization errors.
Mudit Gaur, Prashant Trivedi, Sasidhar Kunapuli et al.
In briefGenerative modeling with continuous flows: a sample-complexity analysis of flow matching
SummaryThis paper studies the theory of flow matching, an ordinary-differential-equation-based approach to generative modeling. Unlike earlier analyses that assume exact empirical risk minimization, it accounts for a finite dataset and a finite number of stochastic gradient updates. Under several regularity and stability assumptions, it proves a high-probability guarantee that separates approximation, statistical, tail, and finite-time optimization errors.
Mudit Gaur, Prashant Trivedi, Shuchin Aeron et al.
In briefText-assisted event-camera place recognition using multi-scale spatiotemporal optimal transport
SummaryThis paper proposes MSTOT-EPR, a method for visual place recognition using event cameras. It encodes event streams and scene text separately, aligns them across modalities with optimal-transport constraints, and fuses them with a cross-modal bottleneck module. Multi-scale spatial and inter-frame temporal optimal transport modules are added to extract spatiotemporal event features for scene recognition. The authors report experiments on the NYC-Event-VPR and EPRBench datasets.
In briefCan Large Language Models Carry Agentic Skills into Goal-Directed Open-Ended Dialogue?
SummaryThis paper proposes Tool-Mediated Interaction (TMI), a training-free protocol that represents dialogue actions as tool calls and counterpart replies as tool results, forming a continuous reasoning-action-observation loop. The authors evaluate TMI on social interaction, emotional support, persuasion, and price negotiation using two models across 15,678 main-experiment episodes. TMI improves several goal outcomes over conversational baselines with and without thinking, while reducing reasoning tokens and cross-turn phrase reuse.
In briefTopoDPO: Structure-aware preference alignment for aerial referring segmentation
SummaryThis paper introduces TopoDPO, a topology-aware dense preference alignment framework for segmenting objects in aerial images based on referring expressions. It mines informative preference pairs using connected-component discrepancy and macro-corner recall, and applies the preference penalty to each local margin before spatial aggregation. On the 4590-instance Topo-RS-Hard structural stress set, TopoDPO improves IoU and reduces Betti-0 error compared with image-level direct preference optimization.
In briefRe-examining LLM homogeneity from a multidimensional perspective
SummaryThis paper measures similarity among large language model outputs with 25 interpretable metrics across five dimensions: semantic, lexical, syntactic, structural, and stylistic and presentational. Across 18 contemporary models, the authors find that models share global meaning but differ in lexical traits, structural patterns, and interaction conventions. Comparing releases from 2024 to 2026, models become more similar to their own earlier versions but not more similar to one another.
In briefBand2Net: EEG-only session-condition decoding via fMRI-privileged transport distillation
SummaryThe paper proposes Band2Net, which uses fMRI as privileged information available only during training, while deployment relies on portable but poorly localised EEG. EEG is encoded into five frequency-specific connectivity graphs and fMRI into seven canonical network tokens, with a balanced entropic transport problem assigning each EEG band its own fMRI target. In subject-grouped cross-validation on the PsiConnect data set, Band2Net reaches 0.801 ROC-AUC for EEG-only session-condition decoding, 5.0 AUC points above an EEG-only student model.
Nishi Kanta Paul, Md Shihabul Islam Shovo, Dwarikanath Mahapatra
In briefGSExplorer: Interactive world generation that keeps a persistent 3D scene.
SummaryGSExplorer is a framework for interactive world generation that maintains a persistent 3D Gaussian scene state across long exploration. It pairs a condition-driven video generator with a SLAM-based reconstruction backend, using inconsistency filtering and cross-chunk bundle adjustment to preserve scene consistency. On WorldScore evaluations for single-step and three-step settings it achieves the best overall performance, with a Camera Control score of 99.07 in single-step extrapolation.
In briefMoPort: A View-Aware 2D Interface for Motion Language Models
SummaryThis paper presents MoPort, an interface that lets motion language models take 2D joint sequences from monocular video as input instead of 3D motion. It combines Paired View Alignment, which makes 2D observations from different viewpoints map to consistent latents, with Camera-Modulated Encoding, which uses estimated camera extrinsics to add viewpoint geometry. The method is evaluated on HumanML3D, KIT-ML, and the real-world RGB dataset Fit3D.
In briefHow Annotators Disagree on Emotion Labels and Why It Matters
SummaryThis paper studies the types of disagreement that arise in emotion annotation, using multilabel annotations of 1,677 sentences from Hemingway's The Old Man and the Sea and comparing them with individual annotations from BRIGHTER and GoEmotions. It identifies three recurring disagreement structures: systematic overlap between emotion categories, differences in how broadly annotators apply labels and in how they treat category boundaries, and sparse participation or abstention that can look like disagreement. The authors argue that similar aggregate disagreement can come from different annotation processes, so these underlying patterns should be understood before aggregating or modeling annotations..
In briefTask-driven sparse-view X-ray CT for measuring layered-object axial gaps
SummaryThis paper proposes a computational imaging approach for sparse-view planar X-ray tomography that designs acquisition around task-relevant information instead of full image recovery. Using the axial termination gap of back-drilled PCB vias as a test case, the method estimates the gap directly from polychromatic projection data via a physics-based sensitivity model, without 3D reconstruction. Numerical experiments show that under a fixed photon budget, task-driven measurement changes little between 12 and 64 views, while reconstruction-driven measurement is strongly view-dependent with much larger errors.
In briefCUFormer: Modality-aware CT-ultrasound point cloud alignment for bone surgery navigation
SummaryCUFormer is a registration framework that aligns intraoperative ultrasound bone-surface point clouds with preoperative CT models to support radiation-free orthopedic navigation. It uses a Modality-Aware Transformer with modality-specific adapters and weights correspondences by predicted imaging visibility. It achieves the lowest mean registration errors among evaluated methods on UltraBones100k, UltraBonesHip, and SpineDepth. On UltraBones100k, its RMSE is 2.54 mm, 65.8% lower than the best baseline.
In briefAbacusGS: Gaussian Splatting Without Per-Splat Depth Sorting
Summary3D Gaussian Splatting normally sorts every splat by depth each frame, which dominates frame time in CPU-sorted web and mobile viewers. AbacusGS stores the blending order in the scene layout by grouping splats into octree boxes of at most 128 splats, stored in order along an axis chosen at write time. Each frame, only visible boxes are sorted, and a dot product per box sets the read direction. On 13 standard scenes it loses 0.13 dB PSNR on average against an equally fine-tuned 3DGS model, and renders 4.5 times faster than per-splat sorting in an OpenGL CPU-sorting viewer.
In briefLearning gradients of homogeneous functions with a network built to respect homogeneity
SummaryThe paper proposes a homogeneous gradient network architecture, built on GradNets with Euler's homogeneous function theorem, that is guaranteed to output the gradient of a homogeneous function. It shows that monotone homogeneous GradNets can universally approximate gradients of k-homogeneous convex functions. In experiments on max energy subspace projection and amortized inner product search, the method outperforms baseline ICNNs.
Mayur S. Gowda, Srinivasa Pranav, Shreyas Chaudhari et al.
In briefAR-RRG: Anatomy-Guided, Retrieval-Driven Radiology Report Generation
SummaryThis paper presents AR-RRG, a method that automatically generates radiology reports from medical images using anatomical guidance and retrieval. It combines an anatomy-aware lesion detection network with a retrieval-driven report generator to reduce hallucinations and missed lesions. The authors report that it outperforms state-of-the-art methods on MS-CXR, MIMIC-CXR, and IU X-Ray across mainstream metrics.
In briefFrom Tool State to Scientific Action: Auditing Decision Provenance in Biomedical Research Agents
SummaryThe paper introduces BioAgent-DF, a method that swaps paired reference and candidate values at scientific computation points to test whether research agents rely on different evidence while reaching the same verdict. Experiments show agents can change which evidence they cite without changing their verdicts, so audits should track both evidence and final answers. In hERG cardiac-safety assay allocation, silent numerical faults changed testing decisions, and restoring the agent's earlier execution state with certified scores recovered all 39 changed decisions without committing any incorrect action.
In briefInformedXRD: A reproducible benchmark and evaluation framework for powder diffraction symmetry classification
SummaryThis paper presents InformedXRD, an evaluation framework for machine learning models that classify symmetry in powder X-ray diffraction data. It provides a machine-readable mapping from 230 space groups to 99 extinction groups, reproducible benchmarks built from frozen data snapshots, and a reporting protocol with frequency baselines, label-permutation tests, and hierarchy-aware metrics. The authors show that evaluation design changes which conclusions are justified, and they release all mapping tables, code, scripts, and checkpoints.
Edward G Friedman, Elizabeth J. Baggett, Abhishek Shetty et al.
In briefA survey of AI-assisted and AI-oriented integrated sensing, communication, and computation for 6G.
SummaryThis survey reviews integrated sensing, communication, and computation (ISCC), a joint design approach for 6G networks. It distinguishes AI-aided ISCC, where learning methods support joint optimization, from AI-oriented ISCC, where the functions are configured to support learning and distributed intelligence. The paper also covers co-design of AI and ISCC, compares design approaches, architectures, metrics, and evaluation practices, and identifies open research problems.
In briefWriting Facts So Models Can Reason With Them: Overcoming the Two-Hop Curse
SummaryThis paper examines the "two-hop curse," where a language model fine-tuned on a new fact can repeat it but cannot use it in a second reasoning step. The authors argue the reasoning ability is intact and the problem lies in how facts are written during fine-tuning. They propose Legible, a loss term that aligns the bridge phrase's representation with the model's own reading of the new entity. Experiments on MQuAKE, RippleEdits, and several model families show it outperforms the compared editing methods.
In briefWhy Vision-Language Models Can Read Text in Images but Still Fail to Use It
SummaryThis paper finds that vision-language models can transcribe text in images word for word yet still fail to answer questions from it, a gap the authors call the read-use gap. Recognizing and using visual text need different resolutions, and putting the evidence into words narrows the gap. The authors propose ALOUD, which uses visual resolution to find and recognize evidence and language tokens to use it, roughly doubling a frozen model's accuracy on long documents compressed 5x and 10x.
WhyThe authors identify a read-use gap in VLMs and propose ALOUD, reporting roughly doubled accuracy on compressed long documents versus a frozen Qwen3-VL-8B and comparisons with several methods.
In briefLabel-free finding and test-time suppression of CLIP's text-reading circuits
SummaryThis paper studies why CLIP-based vision-language models are vulnerable to typographic attacks, where rendered text overrides an image's semantic content. The authors propose a head-decomposed sparse dictionary learning architecture and label-free scores to identify attention heads that carry typographic information. At test time, they reduce attention from patch tokens in these heads, which improves object classification under typographic attacks across five CLIP variants and four benchmarks with up to 1% general accuracy loss.
In briefAuditing conformal calibration and abstention for molecular property prediction under scaffold shift
SummaryThis paper examines selective prediction with split conformal calibration for molecular property prediction, using a gradient-boosted fingerprint model on lipophilicity and BBBP under a scaffold split. It finds that coverage depends on which molecules are used for calibration, and that the abstention score tracks local label variance more than chemical novelty. At a fixed assay budget, gating on interval width or ensemble variance did not improve hit precision over ranking by point prediction.
In briefOpenULF: A modular toolkit for data-driven ultra-low-field MRI using classical and quantum sensors
SummaryThis paper presents OpenULF, an open-source simulation toolkit for data-driven reconstruction in ultra-low-field MRI (field strengths below 10 mT). It generates paired training data and integrates a Bloch-equation-validated physics simulator, SQUID and coil receiver models, reconstruction networks, and an automated design-space exploration engine. The authors trained 774 models and found a performance crossover between SQUID and coil sensors near 450 μT. They conclude that the best sensor depends on the dataset and field strength rather than SNR alone.
In briefImproved Sample Complexity Analysis for First-Order Bilevel Reinforcement Learning
SummaryThis paper studies bilevel reinforcement learning, where an upper-level objective is optimized through a policy's response to a lower-level reinforcement learning problem, as arises when learning a reward model while adapting a policy to it. It proposes a single-loop, fully first-order method for continuous states and finite actions, using multilayer neural actors and a nonlinear critic. Under stated conditions, the coupled iterates are shown to remain contained, yielding an improved sample complexity of about O(ε^-2) for an expected squared gradient norm of O(ε) for the upper objective.
In briefFedUDG: A federated learning method for generalizing to new domains using unlabeled clients
SummaryThis paper proposes FedUDG, a federated domain generalization method that can use unlabeled client data, so models generalize to unseen environments without centralizing client data. It combines a federation-level parameter-functional differentiation strategy, which keeps a semantic anchor updated only with labels, with a per-client student-teacher framework using high-confidence self-training and label-free representation distillation. The authors also introduce DP-FedUDG, a differentially private variant that quantifies the trade-off between privacy and out-of-distribution generalization. The paper reports that FedUDG outperforms state-of-the-art baselines across four evaluation protocols.
In briefReducing Hyperparameter Search for Steering LLMs Using Activation Geometry
SummaryThis paper studies representation steering with difference-in-means (DIM), a method for controlling LLM behavior. The authors use the discriminability index d', a cheap geometric measure of how well contrastive activations separate along the steering direction, across five behaviors on 3 models. They find d' reliably identifies effective layers for steering in multiple-choice question answering, removing the need for a layer-wise sweep. It also remains useful for layer selection in open-ended generation, but predicting success on individual examples is inconsistent there.
Nishkal Hundia, Swastik Agrawal, Navita Goyal et al.
In briefTopoSteer: a topology-based way to choose which layer of an LLM to steer
SummaryThis paper studies activation steering, which controls a large language model at inference time by adding a vector to its internal representations, with a focus on choosing the layer where that vector is added. The authors propose TopoSteer, which scores each layer by the total H_0 persistence of the cloud of per-example displacements, measured as the length of its Euclidean minimum spanning tree. Across several language models and steering tasks, TopoSteer yields more effective steering than discriminability-based layer selection and does not need held-out validation.
Tuc Nguyen, Marcos Guti\'errez-del-Pozo, Eduardo Paluzo-Hidalgo et al.
In briefA spectral-convolutional transformer that estimates multipath geometry for mmWave positioning and radio SLAM
SummaryThis paper proposes the Spectral-Convolutional Factorized Transformer (SCFT), which jointly estimates multipath delay, angle of arrival, and angle of departure from noisy channel estimates and beam-pair descriptors. It also classifies paths as line-of-sight or non-line-of-sight. On realistic ray-tracing data, SCFT reduces GOSPA error by 13.4% compared with ResNet. In snapshot positioning and radio SLAM, it lowers position RMSE by 18.5% and 65.0% relative to ResNet.
In briefHandling irregular state changes in online medical consultations via belief estimation
SummaryThe paper notes that in online consultations, patients often volunteer information or shift topics, disrupting the clinician's questioning. Across 16 medical dialogue datasets, the authors find real online interactions have lower contextual continuity and higher unpredictability than offline and simulated benchmarks. They propose PRISM, which splits triage into belief estimation, policy optimization, and response realization, and report gains over frontier LLMs and clinical agents in decision accuracy and response quality.
In briefGraph Neural Network Surrogate for Variation-Aware Clock-Mesh Analysis and Design-Space Exploration
SummaryThis paper proposes a graph neural network surrogate model of clock meshes used in high-performance VLSI systems. The model predicts power, sink delay and slew, and skew distribution under on-chip variation, trained on SPICE data. It achieves much faster analysis than the approximate methods used in prior work while improving accuracy, enabling more efficient clock-mesh design-space exploration.
Muhammad Hadir Khan, Wajid Ali, Arya Thottiparambil Manikandan et al.
In briefCausal optimal-transport attention with positional priors for length extrapolation
SummaryThe paper studies how causal masking affects optimal-transport attention and characterizes the positional prior it induces, showing attention is diluted as context grows for a fixed balancing strength. Preregistered experiments with 25M- and 100M-parameter language models report better long-context loss extrapolation than softmax. An explicit distance-bias method, CABLE, is reported to outperform RoPE with YaRN at three training lengths.
WhyThe authors analyze a causal unbalanced OT attention, its positional prior, and report preregistered 25M and 100M model experiments against RoPE with YaRN.
In briefSee, Think, Use: Diagnosing Perception and Reasoning Failures in VLMs
SummaryThis paper proposes the See, Think, Use protocol, which traces each vision-language model answer to perception, reasoning, or the use of perception, so developers can tell which stage to improve. A localization probe measures seeing, a text-replacement probe measures thinking, and use is the rate at which answers follow the model's own perception across counterfactual image pairs. Across two model families and six datasets, the helpful intervention depends on the task family, and allocating compute by the measured stages comes close to the best uniform policy at lower cost.
In briefTraining-free method for sorting animal detections into resolve, clear, or defer
SummaryCamera-trap and aerial surveys produce large image sets, and animal detectors still leave many candidates that need expert review. The paper proposes training-free active prototype triage, which reuses reviewed detections as evidence to resolve foreground classes, clear background false positives, or defer unsupported cases to experts. Experiments on camera-trap and aerial datasets find that BioCLIP2 and DINOv3 provide complementary evidence for foreground and background decisions. Deferring weakly supported cases avoids the higher errors of forced decisions.
Anirudh Potlapally, Brijesh Nanda Brunda Yogananda, Jianyang Gu et al.
In briefDIMWIT: Directly Invoking Models Without Intensive Tuning, a method letting pretrained models make resource decisions
SummaryThe paper proposes skipping the development of task-specific resource-management policies and instead asking pretrained models to make decisions directly from the current system state, desired outcome, and available choices. The approach is evaluated on microservice autoscaling, DRAM sizing, and video encoding, using one frontier LLM and two small decision models. The authors argue that making individual decisions may be much simpler than building a general strategy for them.
Saarth Deshpande, Neeraja J. Yadwadkar, Venkat Arun
In briefAn auditable agentic pipeline ranking genes for GWAS variants, checked against independent functional evidence
SummaryThe paper presents V2G, an agentic pipeline that records evidence retrieval and returns ranked candidate genes alongside the evidence, for linking GWAS variants to causal genes. The authors test on 135 held-out loci whether language-model reranking improves on the order of the evidence it receives. They also examine whether the gains come from the retrieved evidence or from recall of known associations.
In briefCan AI agents learn by watching tutorial videos without retraining?
SummaryThis paper proposes Tutorial Agentic Learning (TAL), a method that lets language model agents learn from tutorial videos without updating their parameters. TAL organizes the videos into a structured, persistent memory called a tutorialbook through a two-stage adversarial process, and retrieves from it using a table-of-contents tree. The authors introduce VidLearn, a question-answering benchmark built from 72 chess tutorial videos, where TAL's performance improves as more videos are used. TAL also reaches a stronger opening position than the vanilla model in 69% of head-to-head games.
In briefContactCut: Choosing Contacts to Fix False Merges in Brain Connectomics
SummaryThis paper addresses false merges in connectomics, where separate neurons are wrongly joined in 3D reconstructions, by selecting which contacts to cut so the result is a valid 3D partition. It proposes ContactCut, with CC-Auto for learned single-contact selection and CC-Search for bounded paired-contact search and learned ranking of compact alternatives. On historical two-way corrections from MICrONS and FlyWire, CC-Auto improves FA-3D Part. mIoU over graph-based and learned baselines. A ranker trained on MICrONS transfers to FlyWire with negligible degradation.
In briefCONFIRM: A Framework for Trustworthy Reporting of Autonomous Neuroimaging Results
SummaryLLM agents can autonomously run neuroimaging analyses, but existing systems rely on LLM judgment to decide which results to report, and such judges can endorse site-confounded findings. The paper presents CONFIRM, which separates running analyses from deciding what to report by fixing each hypothesis and its pass criteria in a pre-specified plan, then filtering results through five deterministic gates. It also introduces NeuroClaimBench, with 489 executable neuroimaging claims and 150 controls, on which CONFIRM falsely supports at most 3% of controls versus 23–38% for baselines.
In briefReproducible Check of Bayesian Fuzzy Regression Uncertainty on Tabular and Building Energy Data
SummaryThis paper reimplements a first-order Takagi–Sugeno–Kang (TSK) fuzzy regression model and rebuilds its Bayesian, BIC-based, and spike-and-slab estimators with corrected code. Across 30 outer splits on three targets from two UCI datasets, the corrected methods reach very similar R² values. On a building meter dataset covering 158 buildings across 17 sites, Bayesian TSK shows no demonstrated accuracy or calibration advantage over a persistence baseline.
In briefI-PTC: Interactive code-based tool calling that keeps state for AI agents
SummaryThis paper presents I-PTC, a framework in which models call tools through code snippets, keep reusable intermediate state, and update the tool interface through API corrections. The authors also introduce PTC-Bench, a benchmark of long-tail, production-style user requests with large environments and many dependent API calls. Experiments report better performance on complex multi-step tasks with large tool sets, and substantially lower token consumption than conventional tool-calling and original PTC baselines.
In briefRenormalized Cellular Intelligence: Does Simplifying a Model Keep Collective Capability?
SummaryThe paper introduces Renormalized Cellular Intelligence (RCI), a framework for specifying and auditing whether multi-agent systems keep their collective capability when reduced to simpler descriptions across scales. It requires dynamical closure, decision compatibility, and preservation of the intended baseline comparison, and bounds the resulting error in relative capability gain. Using cellular systems as models, the authors evaluate three reduction candidates on five frozen learned models; the final candidate cuts the per-cell state from eight to five dimensions but fails the closure and decision-preservation requirements needed for a complete certificate.
In briefFOCUS: Closed-Loop Attention Feedback for Efficient Video Understanding by Vision-Language Models
SummaryThis paper presents FOCUS, a training-free framework that reduces how often a vision-language model (VLM) is called on video frames and directs each call toward relevant content. A smaller VLM from the same family detects meaningful scene changes to decide when to invoke the larger model, and a spatial focus map guides which visual regions are emphasized. The model's text-to-vision attention is fed back to guide later observations. On the 500-video EgoSchema subset in streaming mode, FOCUS reaches 74.8% accuracy with 25.4 VLM calls per clip, compared with 74.6% using 55 calls for uniform sampling.
In briefFast minimax prediction method for stationary renewal processes
SummaryThis paper addresses next-symbol prediction for stationary binary renewal processes, whose minimax Kullback-Leibler risk is known to scale as n^(-1/2). The authors present a polynomial-time predictor that attains this rate for any unknown finite-mean interarrival distribution. It builds on a classical type-uniform renewal code and uses an exact dynamic program to compute censored prefix marginals. The statistical rate and its lower bound are prior results; the contribution is computational efficiency.
In briefWhen Should an Agent Refresh? Adaptive Index Reuse for Long-Horizon Reasoning
SummaryThis paper proposes Adaptive Index Reuse (AIR), a method that reduces the computation of sparse attention in language agents working with long, growing contexts. Building on IndexCache, a small controller learns for each segment of an agent episode how many layers should refresh their index selection, using signals such as tool events and a compute budget. On several benchmarks with a 128K context, AIR improves success rates over the strongest static schedule at matched cost and lowers time per successful task.
In briefA study on finely graded, robust detection of human-AI co-authored text
SummaryThis paper proposes a discrete-class framework for text provenance detection that sorts human-AI co-authorship into levels of AI involvement, separating surface-level polishing from changes to discourse structure. It shows how adversarial paraphrasing blurs the decision boundaries of co-authorship detectors. Augmenting training data with adversarial paraphrases from a state-of-the-art paraphrasing attack model yields consistent robustness gains and good cross-generator transferability.
In briefFlyWheel: Fast Attention Kernels for Google's TPU Chips
SummaryThis paper presents FlyWheel, a set of fast attention kernels for Tensor Processing Units (TPUs), covering both softmax and linear attention. It reduces vector computation by reformulating softmax and triangular inversion, improves reuse of matrix operands through coordinated data layouts, and overlaps independent computations to hide latency. On TPU v6e and v7x, FlyWheel speeds up softmax attention by up to 4.14x over RPA and 1.80x over Splash Attention, and Gated DeltaNet prefill by up to 3.61x. When integrated into vLLM serving Qwen3.8-27B, it improves prefill throughput by up to 1.56x with comparable accuracy.
In briefJoint generative model of solar plant fault events and sensor data, with uncertainty estimates
SummaryThis paper presents a multimodal generative model that simulates a solar power plant to support predictive maintenance and fault prediction. It encodes event sequences with a Transformer Hawkes Process and time-series data with an iTransformer, fusing them through co-attention. Two decoders then generate future sensor trajectories and event predictions, producing uncertainty-aware stochastic scenarios. On a dataset from the Red Rock Solar Site in Arizona, the approach improves aggregate event-prediction F1 over baselines.
In briefLatent Harness: compiling agent behaviors into activation shifts after each block
SummaryThe paper proposes Latent Harness, which distills an external agent behavior condition into a rank-8 activation shift applied to every decoder block of a frozen model, using 0.03-0.05% of parameters. At inference no harness text is included and no extra tokens are added. The authors report that compiled behaviors recover the source-harness gain in 8 of 10 preregistered settings.
In briefPrivacy sensitivity inflation from cross-record precision routing in quantized fine-tuning
SummaryThe paper shows that LLM.int8, used in private fine-tuning of frozen quantized language models, routes outlier activation columns based on a batch-wide union, so one record can change how other records are computed and exceed the standard sensitivity bound. The authors bound this sensitivity inflation and propose a record-isolation contract along with record-owned precision dispatch that uses the same matrix multiplications as batched LLM.int8. Audits of the native CUDA stack find records exceeding the bound in practice.
WhyThe authors show LLM.int8 cross-record precision routing inflates DP-SGD sensitivity, propose a record-isolation contract and dispatch, and report native CUDA experiments.
In briefRAG-TRACE: Checking Whether AI-Written ECG Reports Rest on Real Measurements
SummaryThis paper examines how reliable retrieved diagnostic criteria are in AI-generated electrocardiogram (ECG) reports. The authors analyzed 1,083 expert-validated PTB-XL patient records with a library of 47 criteria cards and compared reports generated with and without retrieval. Keyword retrieval introduced criteria that could not be tested or were contradicted by measurements in 46.8% of GPT-4.1-mini reports. A measurement-based verification rule removed most unsupported criterion applications, but the authors conclude that diagnostic validity requires separate, independent assessment.
In briefRadPose: ISAC-Based RF 3D Human Pose Estimation
SummaryThis paper presents RadPose, a learning-based monostatic integrated sensing and communication framework that lets mmWave communication devices estimate 3D human poses without dedicated sensing hardware. It extracts channel impulse responses from standard OFDM signals using full-duplex operation to produce 3D radio point clouds. A cross-frame transformer with spatial and temporal attention, followed by graph-convolution-based inverse kinematics refinement, handles sparse, noisy data. The authors report accuracy competitive with state-of-the-art radar methods and robust performance in unseen and occluded scenarios.
In briefSpecUltra: a single foundation model for mixed-sensor spectral images
SummarySpectral images from different sensors differ widely in band count, wavelengths, and resolution, which makes unified modeling and transfer difficult. The paper proposes SpecUltra, a unified foundation model that uses 3D convolutional tokenization and wavelength-aware 3D RoPE, and pretrains it on the newly built multi-sensor HSI dataset MultiHyper-200K. To bridge hyperspectral and multispectral data, it introduces Spectral Kernel Transformation, which builds MSI embedding kernels from HSI-pretrained kernel slices. Experiments on HSI and MSI classification and segmentation benchmarks show state-of-the-art or competitive results.
In briefERIA: An Instruction-Optimized Agent for Clinical EEG Reporting
SummaryThis paper presents ERIA, a locally deployable framework in which a 27B language model controller analyzes whole-recording EEGs by calling specialized tools for background abnormality, sleep-wake state, seizure-like activity, and interictal discharges, alongside deterministic signal analyses. The controller weights and tools stay fixed; only its natural-language analysis policy is optimized with GEPA using feedback from expert-written EEG reports. The optimized policy improves report quality, transfers to external EEG benchmarks, and remains competitive with models evaluated on those benchmarks.
In briefVocabulary-aligned explanations for computer vision models
SummaryThe paper argues that optimizing vision models only for accuracy does not make them learn human-understandable, domain-specific concepts. It introduces FLUENT, a post-hoc framework that ranks expert-written logical rules over concepts and scores them with uncertainty-aware confidence. Selected rules are illustrated with occlusion-based highlights and localized examples from reference images.
Erfan Mirhaji, Nikhil Krishnaswamy, Jill Zarestky et al.
In briefSame Path, Less Time: Training-Free Speed-Up for Robot Action Models Under Inference Delay
SummaryVision-language-action models predict chunks of future robot actions, but robots lose time while the next chunk is computed and while moving at the slow speed of cautious demonstrations. The paper proposes Path-preserving Time Scaling (PTS), a training-free execution layer that treats robot motion as a time-reparameterized reference path and updates the path only when new observations arrive. On LIBERO with two frozen models under injected inference latency, PTS raises task throughput over the strongest deployable baseline by 33% for pi0.5 and 104% for SmolVLA while matching or improving success rate.
