In briefUnlearning at Speed: Fast Neuron Dampening for Removing Identities from Face Recognition Models
SummaryThis paper proposes Rapid Activation Dampening (RAD), a fast method for removing a person's data from large-scale open-set face recognition models. RAD estimates differences in average neuron activations between identities to be forgotten and retained, then suppresses the neurons most associated with the target identity using a sensitivity-controlled dampening mechanism. The authors also introduce the Forget Score (FS), a probabilistic measure of forgetting in the representation space. Experiments on several recent large-scale face recognition architectures and datasets show that RAD forgets effectively, preserves overall model utility, and substantially speeds up unlearning.
Rouqaiah Al-Refai, Leo Scheubel, Philipp Terh\"orst
In briefCHANNELS4WORLD: A large-scale environment dataset for training wireless foundation models
SummaryThis paper introduces Channels4World, a dataset of 100,000 distinct outdoor radio-frequency scenes and a procedure for generating them. Each scene links base-station and user positions with multipath and interaction records at 3.5 and 7 GHz. A scene-level train-test split is provided, and the authors show that models trained on broader scene collections generalize better to unseen scenes.
Joonkyu Jang, Hyung-Joo Moon, Kwang Soon Kim et al.
In briefFlex-DDCM: Zero-Shot Task-Oriented Image Compression via Diffusion Trajectory Guidance
SummaryThis paper proposes Flex-DDCM, a framework for image compression that serves both human viewing and machine analysis without compression-specific training. Built on Turbo-DDCM, it adapts the coded diffusion trajectory through task-guided noise selection, embedding task guidance into the compressed representation. Experiments show flexible trade-offs among pixel fidelity, perceptual quality, and downstream task performance, with competitive results against trained task-oriented codecs and faster decoding than Diff-ICMH.
In briefVietnamese legal ontology linking each legal answer to the article it relies on
SummaryThe paper builds a knowledge graph from Vietnam's open legal corpus, connecting question-answer pairs and legal documents so that every node and edge traces back to its source record. The graph has 11.5 million nodes and 16.0 million edges. It finds that 36% of the 78,361 answers with resolvable citations cite a law no longer in force, and identifies further citation problems such as mismatched article numbers and shared document numbers.
In briefEfficient offline policy optimization when all policies are linearly realizable
SummaryThis paper studies offline policy optimization in finite-horizon Markov decision processes, assuming every deterministic Markov policy has a linear action-value function over known features with bounded coefficient norm. Using independent trajectories from one unknown behavior policy under all-policy occupancy coverage, it presents an algorithm performing one convex regression per layer. It achieves sample dependence of order the inverse square of accuracy, with computational complexity polynomial in log(2+B).
WhyThe authors claim polynomial-time offline policy optimization needing only deterministic-policy linear realizability, with a tight sample dependence and a two-action lower bound, mainly impacting offl
In briefGenRM-Flow: Generators as Process-aware Reward Models in Flow Matching
SummaryThis paper proposes GenRM-Flow, a framework in which a preference-tuned video generator acts as a reward model directly in the noisy latent space. The reward is derived from the relative difference in continuous-time velocity-prediction errors. The authors show theoretically that DPO, IPO, and RWR fine-tuning all lead to the same optimal policy. The extracted reward can replace external vision-language model scorers and yields measurable gains in video generation.
In briefModeling individual and item differences in atypicality inferences with ACT-R
SummaryThis paper examines why listeners infer that highly predictable events are atypical when they are mentioned, and then construct explanations for them. The authors propose a resource-rational cognitive model in which comprehenders choose interpretation strategies based on individual exploration tendencies and item-level predispositions. The model's parameters are estimated from Raven's Matrices scores, event typicality ratings, and language-model-based accommodation likelihoods, and it predicts the variation in inference rates across individuals and items.
In briefBreaking down Earth embeddings into interpretable features using sparse autoencoders
SummaryEarth observation models produce satellite-based embeddings whose dimensions mix unrelated physical concepts, making them hard to interpret. The paper proposes sparse autoencoders to decompose these embeddings into sparse, nearly single-concept features, and introduces an evaluation protocol for satellite pixels covering reconstruction, concept alignment, and spatial scale. Sparse autoencoders reconstruct AlphaEarth and Tessera embeddings better than baselines at matched sparsity, and their features align with far fewer reference concepts than raw dimensions.
In briefSearch-by-Answer Bias: A Flaw in How Retrieval Models Pick Passages
SummaryThis paper finds that dense retrieval models show a search-by-answer bias, favoring passages that mention answers the model already knows over passages supporting other valid answers. Using controlled factual contrasts from LAMA/TREx, the authors tested eight retrievers, including DPR, Contriever, E5, and Qwen3-Embedding. Higher token-prediction scores for an answer consistently predicted higher retrieval scores, even when the passage expressed a wrong relation.
In briefOrthogonal updates for stable, adaptive minimax optimization
SummaryThis paper proposes an orthogonal update framework for minimax optimization to reduce instability and oscillation in primal-dual training. It introduces AdaSGDA, which adapts stepsizes using accumulated gradient norms so the method needs no problem-dependent tuning in the nonconvex-strongly-concave setting. The authors prove AdaSGDA attains an O(T^{-1/4}) convergence rate, and their variance-reduced variant AdaSGDA-VR attains O(T^{-1/3}). Experiments show competitive performance against existing minimax methods.
In briefHow concept meaning carries across languages: reading, transferring, and erasing shared subspaces in Bangla idioms
SummaryThis paper studies how cross-lingual concept subspaces in frozen encoders retain information and how classifiers and erasers transfer through them. Using 10,359 Bangla idiom senses, it compares class-mean directions against random-subspace and permuted-label baselines. Class-mean directions show the most consistent evidence of shared coarse meaning, while erasure effects do not always match what reading reveals.
In briefComparing density estimation for transfer learning across Banach, Hilbert, and Euclidean spaces
SummaryThis paper argues that common domain adaptation methods assume Hilbertian geometry and measure distribution differences with symmetric L2 divergences. It proposes Banach Manifold Transfer Learning (BMTL), which embeds features on a geodesic manifold and uses an L1-norm density matching regularizer that corresponds to minimizing Total Variation distance. On MNIST-USPS, COIL20 and Office+Caltech10, the L1 variant outperforms an otherwise identical L2 ablation on every task, for example 92.44% versus 89.80% on USPS→MNIST.
In briefDoes Predictive Performance Translate to Causal Performance? Evaluating Tabular Foundation Models for Causal Effect Estimation
SummaryThis paper evaluates whether tabular foundation models based on prior-data fitted networks (PFNs) can serve as nuisance learners for estimating average causal effects within doubly robust estimators. The authors compare PFNs with conventional learners, including Super Learner ensembles, across synthetic and semi-synthetic settings with varying confounding and effect heterogeneity. PFNs performed best in complex settings for both accuracy and confidence interval coverage, though poor coverage appeared at the smallest sample size and in high-dimensional settings without cross-fitting.
Andrew R Weckstein, Issa J Dahabreh, Marinka Zitnik et al.
In briefPreventing Representation Collapse in Self-Supervised Learning Using Score-Based Entropy Maximization
SummaryThis paper introduces EBReg, a self-supervised framework that jointly learns an encoder and a latent score model to regularize the latent distribution and prevent representational collapse. The learned score supplies a diversity-preserving signal, and the same model can refine corrupted representations at test time. The authors report 75.57% linear-probe accuracy on ImageNet-1K with ViT-L and an average relative accuracy gain of about 5% on ImageNet-C.
Adarsh Kappiyath, Silpa Vadakkeeveetil Sreelatha, Xiatian Zhu et al.
In briefSegDis: Separating Foreground Segmentation from Type Classification for Salient and Camouflaged Object Detection
SummaryThis paper addresses joint salient object detection (SOD) and camouflaged object detection (COD), where most existing methods output only a binary mask and cannot tell the object type. The authors propose SegDis, which splits the output into an explicit foreground probability and a type probability conditioned on foreground, handled by separate mechanisms. This allows foreground-recovery errors and type-assignment errors to be diagnosed separately. On USC12K with a PVT-v2-B2 backbone, SegDis reaches 68.98% FG-mIoU and 78.35% mIoU at about one-eighth of USCNet's parameters.
In briefLLM-guided multi-criteria deep reinforcement learning for dynamic, efficient 5G UPF selection
SummaryThis paper proposes a deep reinforcement learning method for selecting user plane functions (UPFs) in 5G networks, using a dynamic multi-objective reward function. The method uses large language models to translate network operator intents and takes network data analytics function (NWDAF) abnormality alerts as input. Validated on a real-world 5G emulation platform, it balances latency, power consumption, and traffic distribution, outperforming heuristic baselines and fixed DRL by 16.89%.
Abdelrahman Soliman, Ahmed Refaey Hussein, Lei Lei
In briefD3Tree: Learning to Prune Draft Trees for Faster Verification
SummaryThis paper presents D3Tree, a method for pruning block-diffusion draft trees used in speculative decoding. A small learned head predicts each candidate node's conditional survival, and thresholding these scores keeps an ancestor-closed subtree whose size adapts per request and decoding round. Under heavy load, a batch-wide node budget is allocated by predicted survival. The authors report it is fastest in 70 of 72 load settings across six vision-language verifiers and 17 benchmarks.
In briefBiVerify: Linking Backward and Forward Checks for Math Self-Correction
SummaryThis paper proposes BiVerify, a framework that lets language models check math solutions in both directions. It first tests the answer against the original constraints, then checks the reasoning process, switching directions and re-verifying based on the results until a candidate passes both. Tested on five math reasoning benchmarks with several DeepSeek, GPT, and Gemini models, it beats the strongest public baseline by 13.5 percentage points on average. It also reports better overall performance than either verification direction used alone.
In briefMG-Cache: A Multi-Granularity Caching Method for Faster Diffusion Transformer Inference
SummaryThis paper proposes MG-Cache, a training-free caching method that speeds up inference for Diffusion Transformer (DiT) models. It exploits redundancy across both timesteps and Transformer layers. The method uses two stages: an offline timestep analysis to find redundant timesteps for cache reuse, followed by an adaptive layer-wise strategy that caches stable Attention and MLP submodules based on input stability and layer importance. Experiments on mainstream open-source DiT models show significant acceleration while staying close to uncached outputs, with up to a 4.18x speedup on HunyuanVideo.
In briefOOM: Memory-execution co-designed full-parameter training of large MoE models with offloading
SummaryThis paper presents OOM, a system for full-parameter training of large mixture-of-experts (MoE) models on a single multi-GPU node by offloading model state to host memory. It co-designs expert execution and memory allocation, jointly planning expert residency and activation retention. On four H100s, OOM trains with 1M-token contexts where FSDP2, ZeRO-Offload, and MegaTrain run out of GPU memory at 512K. On four RTX 4090s, it achieves geometric-mean throughput of 1.58x and 4.08x over CPU-offloaded FSDP2 while using as little as 1/3.0 of its host memory.
In briefNuJEV: structured visual judgments for consistent cell nucleus counting
SummaryThis paper presents NuJEV, a method for locating and classifying cell nuclei in microscopy images using the Qwen3-VL-2B model, rather than generating counts as free-form text. It jointly trains a visual LoRA and a parallel head that predict nucleus presence, centers, and types, with counts derived from these predictions. On 2,732 PanNuke test images, it reports a five-query mean absolute error of 1.844 overall and 5.268 on dense images.
In briefAOP-Smart-Finder: A human-AI framework for finding adverse outcome pathways and mapping evidence from large literature collections
SummaryThis paper presents AOP-Smart-Finder, a high-throughput human-AI hybrid framework that combines large language models, AOP-Smart retrieval-augmented generation, and Python automation to discover biological events and relationships from large-scale literature. It builds Adverse Outcome Pathways (AOPs) and maps supporting evidence. The framework was tested on DEHP and amyotrophic lateral sclerosis (ALS) cases, and after manual curation and Weight of Evidence evaluation, the ALS case yielded a mechanistic network of 16 key events and 23 key event relationships.
In briefSeparating skill context for agents that follow standard operating procedures
SummaryStandard operating procedures are written concisely for people, so LLM agents that turn them into executable skills lack the tacit knowledge needed for automation. The paper manages SOPs and contextual knowledge separately and uses a contract-aware evaluation that separates genuine violations from differences resolved by that knowledge. On eight Indeed-derived workflows with a single LLM, the functional-equivalence rate rose from 12% to 58%. Under three SOP revision perturbations, previously extracted contextual knowledge covered 79.5% of newly introduced automation gaps.
Mitsuhiro Nakayama, Toshiaki Sota, Takehiko Yamaguchi et al.
In briefSpatial Harness: physics-engine-grounded spatial reasoning
SummaryThe paper introduces Spatial Harness, which represents scenes as explicit, executable world states so that spatial reasoning is grounded in physics-engine feedback. A Perceiver places objects and cameras in a shared coordinate frame using constraints from the question and visual estimates, and a Reasoner executes viewpoint changes in the engine. On MindCube, it raises Qwen3.5-27B from 52.70% to 71.14%, and reaches 98.25% given identical annotated poses versus 71.81% for model-internal reasoning.
In briefFrozen-Weight Agent Self-Improvement: Value and Limits of Failure Traces
SummaryThis study examines whether the notes surrounding an AI agent, its harness, can learn from the agent's failure traces compared with a matched search that never sees them. Using an open 9-billion-parameter model with fixed weights and notes written by a larger teacher model, notes based on failures improved accuracy by 4.0 percentage points on generated customer-service decisions. A process that only tried and kept better-scoring notes accounted for about 60% of that total gain. In full customer-service conversations, the same notes showed no demonstrated improvement.
In briefOPSRL-SSP: Optimistic posterior sampling with near-optimal regret for shortest-path learning
SummaryThis paper studies online reinforcement learning for stochastic shortest path problems whose transition dynamics are unknown. It proposes OPSRL-SSP, an optimistic posterior-sampling algorithm that runs in epochs, uses a logarithmic time-dependent schedule for posterior samples, and adds a pseudo-state with an optimistic boundary value. The authors prove a high-probability regret bound whose leading term matches the minimax lower bound, and extend the analysis to general zero-cost SSPs.
WhyThe authors propose an optimistic posterior-sampling SSP algorithm with regret whose leading term matches the minimax lower bound, addressing an open problem for posterior sampling.
In briefWhere Should a Gate Look? Comparing Input and State Gating in Diagonal State-Space Models
SummaryThis paper examines whether memory-update gates in diagonal state-space models should depend on the current input, the recurrent state, or both. The authors prove that for scalar models, state gating solves induction-head tasks for arbitrary vocabulary size and sequence length, while input gating fails on a finite instance under standard stability assumptions. Experiments show state gating keeps its advantage on induction-head tasks in higher-dimensional models, input gating performs better on selective copying, and combining both gives the strongest overall performance.
Riccardo Zattra, Giacomo Baggio, Leonardo De Marchi et al.
In briefMoT3DVG: A benchmark for locating objects in outdoor 3D scenes using motion-aware language descriptions
SummaryThis paper introduces MoT3DVG, a dataset for outdoor 3D visual grounding, which means finding objects in 3D scenes from language descriptions. It covers 850 scenarios with 31,128 frames from nuScenes and provides 144,568 prompts describing how dynamic objects move over time. The authors also propose DynaVG, a framework that models local motion and global temporal context, and report state-of-the-art results on the dataset.
In briefCVaR Deep Optimal Stopping: A Deep Learning Method for Risk-Sensitive Stopping Decisions
SummaryThis paper proposes a framework and algorithm for optimal stopping that minimizes risk by targeting the Conditional Value at Risk (CVaR) of the stopped reward. The method, called DIOS (Distribution-Informed Optimal Stopping), exploits structural properties of stopping problems to avoid state augmentation and uses a pathwise gradient estimator. Across 24 main Bermudan max-call configurations, DIOS achieves the highest mean CVaR and the shortest recorded training time among the tested CVaR baselines.
In briefDiffusion-Based Recovery of Missing SLO Data for Predicting Post-Cataract Surgery Vision
SummaryThis paper proposes a conditional diffusion method that recovers proxy representations of missing scanning laser ophthalmoscopy (SLO) images in cataract patients' preoperative data. The recovery is guided by fused OCT image and clinical text features, and the result feeds a multimodal network that predicts postoperative best-corrected visual acuity. On 2,141 samples with a 58.80% SLO missing rate, the method reached 0.9714 accuracy within 0.1 LogMAR and a mean absolute error of 0.0423.
In briefAuditing Forecast Interval Coverage When Sensors Fail and Records Are Incomplete
SummaryThis paper presents a measurement-record audit that checks whether prediction-interval coverage falls short during sensor outages, a shortfall that can be hidden by acceptable overall performance. The method links outage duration at forecast issue time, availability of original targets, and time-uniform statistical evidence, projecting covered and scored count bounds onto a cumulative conditional expected coverage ratio. Selected traffic and air-quality results are reported, including that in METR-LA only 7.6% of issued long-gap forecasts have originally observed targets.
In briefRank-guided data-free quantization method for Vision Transformers
SummaryVision Transformers are powerful but demanding to run on mobile devices. This paper proposes RankQ-ViT, a data-free quantization framework that compresses ViTs without access to original training data. It generates synthetic calibration samples with greater feature diversity, guided by patch similarity matrix rank. Experiments on ImageNet, COCO, ModelNet40, and ShapeNet show effectiveness across multiple architectures and tasks, especially at low-bit quantization.
WhyThe authors replace similarity entropy with rank-guided, block-aware self-contrastive sample synthesis, reporting results on ImageNet, COCO, ModelNet40 and ShapeNet.
In briefVerify Before You Fetch: A method for reliable LLM memory compression with bounded attention error
SummaryThe paper presents VerKV, which keeps a small set of query-agnostic pivotal tokens and a sketch of the KV cache on the GPU while storing the full cache in CPU memory. At inference, it estimates attention error and skips host memory access if a user-specified (ε, δ) tolerance is met, otherwise fetching more tokens from the CPU. The authors also introduce the PoissonComp sampling method and report 90-100% of dense quality at 32K context with 8-32x lower GPU memory use.
Devanshu Jha, Gaurav Kumar, Agniva Chowdhury et al.
In briefPROLA: A Low-Rank Adaptation Method for Predictive Weather Downscaling
SummaryThis paper introduces Predictive Spatiotemporal Weather Downscaling (PSWD), a task that predicts future high-resolution weather fields from recent coarse weather states while refining both spatial and temporal resolution. The authors use a shared pretrained backbone with fixed numerical scaffolds and lightweight residual heads, split into spatial and temporal stages. They propose PROLA, which allocates a fixed low-rank budget between principal directions of pretrained weights and their orthogonal complement. Experiments show PROLA outperforms representative low-rank adaptation baselines and also performs well across other datasets and tasks.
In briefReGKG-2025: a GKG-compatible rebuild of GDELT's Global Knowledge Graph for the June–July 2025 data gap
SummaryA June 2025 outage in GDELT's collection pipeline left a gap in its Global Knowledge Graph of news metadata. This paper recovers 281,620 article records from Common Crawl and the Wayback Machine, covering all 1,665 missing 15-minute batches, and converts them to the GKG 2.1 format with provenance and manual review records. The dataset suits retrospective content analysis and source-level aggregates, but its tone and theme distributions differ from native GKG and it cannot support exhaustive counts.
In briefA method that fits curved low-dimensional shapes to data while adapting to noise patterns
SummaryThis paper proposes AMLE-SQMF, an alternating maximum likelihood method for fitting quadratic manifolds to high-dimensional data. It jointly estimates the manifold and the noise distribution, modeling residuals with a generalized Gaussian whose shape parameter sets the fitting loss. The approach avoids prescribing a noise model in advance. Experiments on synthetic and real datasets show it fits manifolds effectively and remains robust across diverse noise distributions.
In briefScrapeBench: Testing Whether AI Agents Follow Website Scraping Rules
SummaryThis paper introduces ScrapeBench, the first agentic benchmark for measuring whether AI agents comply with website scraping policies such as terms of service and robots.txt. On 1,626 popular websites that prohibit scraping, most agents complied less than 5% of the time, even when they checked those policies. On 269 sites that allow scraping, some agents refused anyway, with Claude Opus 4.6 over-refusing 98.5% of the time. A human baseline with 180 participants showed compliance rates similar to the agents but much higher over-refusal rates.
In briefTransArt3D: Reconstructing Movable 3D Objects from Two States Using Video Transition Priors
SummaryThis paper presents TransArt3D, an optimization-based framework that reconstructs an articulated 3D object from images of only two endpoint states. It uses frames generated between the endpoints by a video diffusion model to jointly estimate part structure, a 3D motion field, and a shared joint. A canonical appearance fusion step then maps both endpoint images onto a single surface. On PartNet-Mobility, the authors report better geometry and joint accuracy than methods given the same two endpoints, and better results than methods given six observed states.
In briefNCCL-BOT: Using Online Bandits to Correct NCCL's Cost Model
SummaryNCCL chooses an algorithm-protocol pair for each collective operation using an analytic cost model that never uses measured latency. NCCL-BOT is a tuner plugin in which a per-context bandit, warm-started from this model, learns from measured latencies. In a replay of a forced-arm H100 sweep, the bandit reduced the contexts where the model's choice loses more than 10% from 19 to 1. Live tests on two latency-emulated node sets showed gains in some cases, but the authors do not show a net end-to-end win.
In briefEASE-KV: Diagnosing Why Stored Memories Aren't Always Used Effectively
SummaryThis paper presents EASE-KV, a framework that studies why reusable key-value (KV) memories do not always become useful evidence for a frozen language model. It separates three stages: whether required relations are kept during encoding, whether preserved memories get appropriate influence during allocation, and whether outcome supervision can tell better allocations from worse ones. Causal encoding improves compositional accuracy by 25.0 points over block-diagonal encoding, while the tested bridging methods fail to recover lost context. With stored states unchanged, evidence-guided allocation raises LongMemEval accuracy from 31.2% to 38.4%.
In briefSingle-radar 3D imaging of both sides of a robot's path
SummaryThe paper presents a motion-assisted method that reconstructs structures on both sides of a mobile robot's trajectory using one forward-mounted millimeter-wave radar. It uses a near-field model based on spherical propagation paths, combined with sliding-aperture back-projection and top-K power fusion, followed by multi-layer height extraction and adaptive height refinement. In controlled, office/study, and corridor experiments, mean lateral-forward localization errors were 4.65, 5.63, and 5.98 cm, with peak-to-background ratios above 13 dB.
In briefTransformers Splitting Object Detection and Captioning Across Edge Devices
SummaryThis paper proposes a unified transformer framework that performs object detection and dense region captioning on distributed edge networks. It shares one visual encoder across detection, region extraction, and language decoding, avoiding a separate image-encoding path. A workload-allocation method assigns model stages and input streams to connected edge nodes. On a six-device testbed, the unified INT8 model runs at 38.1 FPS on a Jetson Orin Nano, and under four concurrent streams it reaches 35.2 FPS with 28.4 ms mean end-to-end latency.
In briefA unified system that finds and reads license plates in one network
SummaryThe paper proposes one network that detects license plates and recognizes their characters on low-power edge hardware, sharing a single feature encoder and decoding characters directly from aligned feature maps without cropping plate images. It is trained on about 37,000 manually verified images covering plate layouts from all 50 US states. On an NVIDIA Jetson Orin Nano with an INT8 TensorRT engine, it reports 97.8% mAP50, 90.7% full-sequence accuracy, and 147 FPS. Compared with a YOLOv8s+CRNN baseline, it reports 7.1x higher throughput and a 77.8% reduction in inference memory.
In briefWarBias: A bilingual Ukrainian–English benchmark for measuring language model stereotypes about displaced people and veterans
SummaryThis paper introduces WarBias, a Ukrainian–English benchmark that tests whether language models hold stereotypes about internally displaced people and veterans. It contains 1,007 stereotype, counter-stereotype, and unrelated sentence triplets, plus 240 multiple-choice questions per language, with each stereotype claim given a human severity rating. In a baseline evaluation of six models, four preferred stereotypical Ukrainian sentences about at least one group, most strongly for displaced people, while English results were mostly balanced.
In briefHow Efficient Is Your Agent? Budgeted Online Predictive Bounds
SummaryThis paper evaluates AI agents online, as tasks arrive one by one, by constructing an upper bound on each task's time-to-success before it runs. Because the budget is limited, some runs stop early, leaving their time-to-success unobserved. The authors introduce PROBE, which calibrates these bounds online while adaptively allocating the budget. They prove PROBE gives a coverage guarantee over any finite horizon while meeting the budget constraint, and experiments on software engineering, retrieval-augmented generation, and instruction following show closer-to-target coverage and tighter bounds than a calibrated baseline.
In briefAn adaptive recommender that uses context and stacked attention to track users' changing tastes over time.
SummaryThis paper proposes a recommendation system that adapts to users' preferences as they change over time. It treats the items each user interacts with in each time interval as contextual information, and introduces two attention mechanisms: a contextual item attention module and a multi-head attention module. Experiments on three real-world datasets show that the proposed model outperforms several state-of-the-art recommendation methods.
Jung-Hsien Chiang, Chung-Yao Ma, Chi-Shiang Wang et al.
In briefVariance-reduced accelerated stochastic optimization with Adam-style preconditioning
SummaryThe paper proposes scalar and diagonal versions of VR-Adam-SHANG for smooth stochastic convex optimization. The methods combine STORM-type recursive gradient estimation, Adam-style preconditioning, and SHANG's accelerated interpolation and correction, using at most two stochastic gradient evaluations per iteration. Under unbiased sampling, bounded variance, and mean-square smoothness assumptions, the authors derive convergence bounds for gradient estimation error and the expected objective gap. In experiments at matched sample-gradient budgets, at least one of the two versions achieved the lowest mean terminal objective gap in every tested setting.
In briefTask-adaptive anonymisation for privacy-preserving sharing of clinical speech data
SummaryThis paper introduces CASA, a speech anonymisation agent that adapts its voiceprint and content anonymisation to the privacy and clinical utility requirements of each clinical task. It uses recursive self-improvement to iteratively evaluate and refine its strategies. Validated across 14 clinical tasks in five languages, CASA preserved disease-related classification performance while meeting speaker-anonymisation requirements on all tasks, and also resisted speaker re-identification and removed sensitive personal information.
In briefBiRefNet-V2: Rethinking Bilateral References for High-Resolution Image Segmentation
SummaryThis paper presents BiRefNet-V2, a model for high-resolution dichotomous image segmentation that separates foreground from background at the pixel level while preserving fine structures such as thin filaments and sharp edges. It introduces locality-preserving detail injection, a learned boundary stream with focused supervision, and a dual-branch decoder, within a modernized framework using a DINOv3 backbone. The authors report state-of-the-art results across all 13 evaluation sets spanning DIS5K, camouflaged object detection, and high-resolution salient object detection, with training reduced from 600 to 50 epochs.
WhyThe authors extend BiRefNet with a semantic gate and learned boundary stream, reporting gains across 13 evaluation sets including DIS5K.
In briefDuSS-PET: label-free self-supervised low-dose PET reconstruction using pseudo-dual-domain consistency
SummaryThis paper introduces DuSS-PET, a self-supervised framework for reconstructing low-dose PET images without requiring paired low-dose and full-dose scans. It builds on the Noisier2Noise paradigm with a Bernoulli corruption strategy and combines an adaptive Swin transformer, an image enhancement module, and a consistency constraint between image and pseudo-sinogram representations. Experiments on multi-center datasets report performance that exceeds other self-supervised methods and is comparable to or better than fully supervised approaches at various dose levels, with good generalization across scanner vendors.
In briefPrimitives and Paths: predicting what post-training can do from a base model's existing skills
SummaryThis paper asks whether reinforcement learning teaches language models new abilities or only amplifies those from pre-training. It proposes a two-dimensional state space for post-trainability, measuring whether complete solutions can be sampled and whether the component skills can be executed. The authors identify amplification, composition-limited, and acquisition-limited regimes. Experiments on synthetic composition tasks with three small language models find that mastering primitives alone does not guarantee RL can learn compositions, while modest composition supervision enables substantial gains.
In briefAuto-GlaKG: Automated Biomarker Reasoning for Explainable Glaucoma Detection
SummaryThis paper presents Auto-GlaKG, a framework that detects glaucoma from fundus images. A ResNet50 network predicts the cup-to-disc ratio and six morphological signs, and a knowledge graph reasoner combines these with fixed clinical rules into a glaucoma score. On 689 images using group-aware five-fold cross-validation, it achieves a pooled AUC of 0.951, with sensitivity 0.970 and specificity 0.779.
In briefLearning calibrated probabilistic forecasts with Koopman mode decomposition and the continuous ranked probability score
SummaryThis paper proposes a probabilistic forecasting method that combines machine learning with the linear dynamics of Koopman mode decomposition, producing predictive distributions by pushing noise samples forward. The model is trained with the continuous ranked probability score to target calibrated forecasts. It is validated on two nonlinear ODE examples and two PDE examples, vortex shedding and cavity flow.
In briefALETHEIA: A Multi-Frequency Eddy Current Thermography Dataset for Neural Operator Learning
SummaryThis paper presents the ALETHEIA dataset, the first 3D benchmark for learning data-driven solvers for nondestructive testing. It simulates eddy-current-induced heating in conductive solids and the resulting heat propagation, with over 4,700 high-resolution samples across 10 excitation frequencies from 1 to 100 kHz. Real infrared thermography data from cracked rail specimens are included, and four tasks are benchmarked with various neural operators.
In briefHow self-supervised speech models learn phonemes and prosody layer by layer during pretraining
SummaryThis paper uses frozen-encoder probes on every Transformer layer across pretraining checkpoints of HuBERT-Base and wav2vec 2.0-Base to study when and where acoustic, prosodic and phonetic information emerges. It covers five tasks, from absolute pitch and lexical stress to phoneme boundaries and 38-way phoneme identity. Acoustic-prosodic cues and phone boundaries are decodable from the earliest checkpoints, while phoneme identity is built during pretraining and moves to upper-middle layers under HuBERT's stage-2 targets.
Masaru Tanibata, Shun Takahashi, Hiroki Ouchi et al.
In briefHow style rewriting undermines LLM confidence in fake-news detection, studied layer by layer
SummaryThe paper examines how a large language model's confidence in judging news veracity changes when only an article's writing style is altered while its facts stay the same. Tracing predictions and confidence through the network layers across three model families and three fake-news benchmarks, the authors find that sensational or emotionally charged rewrites sharply reduce accuracy while the models remain just as confident. The authors conclude that an LLM's stated confidence about veracity cannot be trusted as a safety signal when an article's style may have been manipulated.
In briefFRESCO: A New Consistency Control Method for Asynchronous Pipeline Parallel Training
SummaryThis paper examines how asynchronous pipeline parallel training suffers instability from inconsistent gradient computations, finding that error signals depend on forward-time features as well as parameter states. The authors introduce FRESCO, which casts asynchronous updates as a constrained optimization problem seeking minimal update changes. In large language model pre-training, FRESCO outperforms representative asynchronous baselines.
In briefLoMASA: A local multi-agent sales assistant that hands off uncertain B2B quote decisions
SummaryThe paper introduces LoMASA, a training-free multi-agent sales assistant that runs entirely on one desktop computer to help salespeople re-coordinate B2B quotes when customers change quantities, discounts, payment terms, or costs. A 35B mixture-of-experts model makes each decision in one pass and reads confidence from its option probabilities, while a 27B dense model drafts customer replies and escalation briefs when it is unsure. On 500 simulated episodes with 1512 decisions built from public supply-chain data, the salesperson decides only 19.5% of cases, 38.0% are settled automatically, and 42.5% are escalated to a manager by hard rules. On automatically settled decisions, accuracy matches a rule baseline (0.894), and fewer commitments later fall below the margin floor (2.61% vs 4.53%), at the cost of more manager escalations.
In briefTLRS-DPO: Token-level risk scoring to improve direct preference optimization for language models
SummaryThis paper proposes TLRS-DPO, an extension of Direct Preference Optimization that adds fine-grained token-level control while remaining offline and reference-based. It converts log-ratios on rejected responses into localized risk scores, uses a curriculum schedule that gradually strengthens this compensation, and applies risk-adaptive screening to penalize only high-risk tokens. The authors report that compared with strong DPO baselines, it improves alignment and safety-relevant behavior with minimal impact on helpfulness.
In briefMCSA: Multi-cutoff spectral aggregation for equivariant GNN atomic potentials
SummaryThis paper proposes MCSA, a plug-and-play module for equivariant graph neural network interatomic potentials that aggregates radial kernels with different cutoff radii to capture multi-scale geometry. It fuses cutoff-dependent radial weights before angular tensor-product operations, exploiting compact kernel support while preserving rotational equivariance. Experiments report consistent gains in energy, force, and stress prediction, and improved accuracy for phonon dispersion and lattice thermal conductivity.
In briefDo feature interactions break explainability? Provable explanations for any-order neural additive models
SummaryThis paper studies how to compute provably minimal and sufficient explanations for neural networks with restricted feature interactions, lying between fully general networks and purely additive models. The authors prove that the problem is Σ2^P-hard even when interactions have constant order, and identify two tractable settings, one based on sparsity in the interaction structure and one based on a stronger interaction-aware notion of sufficiency. Experiments show these explanations are significantly smaller and faster to obtain than those from standard algorithms, and come from models with higher accuracy than standard neural additive models.
WhyThe authors extend provable explanations to any-order feature interactions in neural additive models, prove Σ2P-hardness at constant order, give two tractable settings with algorithms, and report smal
Idan Refaeli, Yizhak Yisrael Elboher, Guy Katz et al.
In briefEbbFlow: PINN-Inspired Progressive Curriculum for MeanFlow Training
SummaryEbbFlow is a training method for MeanFlow, a model that learns interval-average velocity for few-step image generation. It treats temporal samples as collocation points, starting with diagonal supervision and then gradually adding interior intervals with a growing maximum length. In class-conditional ImageNet 256×256 experiments with DiT-B/2 and no classifier-free guidance, EbbFlow reached a 50k-sample FID of 39.65 at one evaluation and 36.02 at two, compared with 44.82 and 39.86 for MeanFlow.
In briefEdge Computing for Computer Vision in IoT: Feasibility and Directions
SummaryThis paper examines whether edge computing can serve as a foundation for computer vision in Internet of Things systems. The authors run experiments under real-world, dynamic conditions and evaluate related technologies, strategies and process flows. Research directions include multi-access edge topologies in 6G networks, model quantization and satellite-enhanced communication, along with zero-shot learning, multimodal perception and decentralized generative AI. The study evaluates LoRaWAN and satellite IoT communication, analyzes computer vision workloads on edge hardware, and outlines future directions for low-latency, energy-efficient applications.
In briefscBIG++: Predicting how single cells respond to gene or drug perturbations across all genes
SummaryThis paper introduces scBIG++, a framework that predicts single-cell gene expression changes caused by genetic or chemical interventions. Unlike methods limited to highly variable genes, it models expression for all retained protein-coding genes while accounting for differences among cells. Experiments across six genetic and chemical settings, covering 6,539 to 19,451 genes, report competitive recovery of perturbation effects and full-gene expression reconstruction.
In briefReplay Reliability in AI Agent Testing: Randomness, Hidden State, and Test Checks in AppWorld
SummaryThis paper audits all 71 failed runs in a fixed historical AppWorld corpus using 284 isolated replay attempts. Reconstructed settings admit 42 cases, and resetting the random-number generator before each interaction admits 68; database and test-oracle checks narrow the gap to four cases. The authors identify hidden-state differences that persist even when outputs and verdicts match.
In briefSeparating where and when: a factored explanation method for video action recognition
SummaryThis paper proposes a perturbation-based post-hoc explanation framework for video action recognition that splits explanations into a spatial appearance mask and a temporal motion mask. The two masks are jointly optimized against a frozen model, with a separation penalty and an area constraint. Evaluated on EPIC-Kitchens-55, EGTEA Gaze+, and Something-Something V2, the method reports lower deletion AUC than leading baselines and masks about five times sparser.
In briefDeterministic Narrative Gravity: Linking Theory of Mind and Historical Simulation via ABM-LLM Design
SummaryThis paper presents a deterministic multi-agent framework that couples large language models with an agent-based modelling pipeline grounded in Theory of Mind and Cognitive Archaeology, for historical social simulation. The system runs locally at temperature zero and verifies exact output identity with SHA-256 hashing. Across seven model configurations from 4B to 31B parameters, smaller models at or below 8B produce repetitive dialogue, while models at 12B and above keep scenarios open and expand them with period-consistent details. The full pipeline is released as open-source software to support independent replication.
In briefDynOLAT: Controllable Full-Body Video Relighting via OLAT Prediction
SummaryThis paper proposes DynOLAT, a framework that predicts temporally consistent One-Light-At-a-Time (OLAT) basis videos from flat-lit inputs and linearly combines them according to a target environment map. This makes lighting direction and relative intensity explicitly controllable. The authors also introduce the first large-scale dynamic full-body OLAT dataset, with over 24 million frames from 224 identity-outfit combinations. Experiments show it responds more faithfully to target lighting changes than the tested baselines while preserving appearance details and temporal coherence.
WhyThe authors predict temporally consistent OLAT basis videos and linearly combine them via environment maps, plus a 24M-frame dynamic full-body OLAT dataset; reported responses beat baselines.
In briefViscot-steer: Transferring visual reasoning across vision-language models without training
SummaryThis paper introduces Viscot-steer, a training-free activation steering method for transferring visual chain-of-thought (thinking with images) abilities between different large vision-language models. It splits the capability into grounding and reasoning vectors, aligns them to the target model's activation space, and applies them during bounding-box prediction and reasoning-trace generation. The authors report that across several benchmarks it outperforms the strongest transfer baseline by an average of 1.86 percentage points on primary metrics.
Ningning Wang (1), Beier Zhu (1), Xingyu Zhu (2) et al.
In briefNot Yet in the Profile: Uncertainty-Aware Dynamic User Profiling Learned from Future Interactions
SummaryThis paper proposes UAPM, a dynamic user profiling method that separates confirmed user traits from ambiguous new behavior, using a Profile, a Buffer, and a Rulebook. Training uses recorded future interactions, which are not needed at deployment. On three Web personalization tasks, UAPM outperforms multiple baselines under the same downstream input budget, improving on the best same-budget baseline by up to 5%.
In briefWhen majority voting hurts small language models and how to skip it
SummaryThis paper studies chain-of-thought prompting with self-consistency, which samples several reasoning traces and takes a majority vote, on small language models. It finds that applying self-consistency to every prompt can lower accuracy because it flips some correct greedy answers to wrong ones. A gate that applies it only to selected prompts cuts most of the compute while keeping accuracy close to the always-applied baseline. Experiments use GSM8K and MATH-500.
In briefSource-anchored autoregressive generative image compression at extremely low bitrates
SummaryThis paper introduces Source-Anchored Visual Autoregression (SAVAR), a generative image codec built on Visual Autoregressive modeling for bitrates below 0.01 bpp. It transmits early-scale tokens for image-specific layout and a compact caption for global semantics. On Kodak and CLIC2020, SAVAR lowers LPIPS by about 8–9% and DISTS by up to 22% relative to matched ARPC operating points at a lower bpp. It reconstructs a 512×512 image in about 218 ms on one NVIDIA L20X GPU.
In briefDhara: A search engine for evidence in Indian Knowledge Systems texts
SummaryThis paper presents Dhara, a retrieval engine for Indian Knowledge Systems collections that combines hybrid search, Entity Notes, fuzzy verse retrieval, and neural reranking. Five comparisons on 325 questions in Dhara Bench show BGE-M3 outperforming BM25 across categories, and Qwen reranking improving several metrics over BGE at about 6.72 times the serial computation time. The authors also compare the pipeline with Gemini. They conclude that retrieval choices should depend on evidence requirements and resource budgets.
In briefTEAM: A Transformer for Graphs Whose Nodes Carry Text and Images
SummaryThis paper proposes TEAM, a framework for graphs whose nodes have multimodal attributes such as text and images. It treats each node's modalities as separate tokens, organizes local structure with a Modality Encoding Tree, and uses a Modality-Source Context Transformer to enable fine-grained cross-node, cross-modal interaction. The authors report that TEAM achieves the highest mean scores among evaluated baselines on graph-learning metrics and on G2Text and G2Image metrics. On Toys, it exceeds the strongest baseline by 3.37 points in Accuracy and 2.93 points in Macro-F1.
In briefWhy guidance on the first step hurts speaker similarity in fast speech synthesis
SummarySpeech synthesis systems that pair a language model with a flow-matching decoder are now run with only two to four steps, which lowers speaker similarity. The authors trace most of this loss to classifier-free guidance applied on the first step, where it shifts all trajectories by nearly the same vector. They propose two training-free samplers, FSU and FSU-Mid, that recover most of the lost similarity while using one fewer network evaluation than the official two-step sampler.
In briefWhen Nearby Experts Can Be Combined: Cross-Task Model Merging in Neural Thickets
SummaryThe paper studies whether task experts produced by random perturbations of a shared checkpoint keep their gains when merged. The authors formulate expert-pair selection around joint retention, measured by strict joint success and worst-task gain, and propose a quality-constrained spectral selection method with cross-task screening, layerwise geometry, separate validation, and Base fallback. Across six checkpoints and 21 task pairs, the method improves strict joint success over Top-1 averaging and leads most model-scale comparisons.
In briefHow engineering papers changed after AI chat assistants: a study of 1.8 million articles
SummaryThis study examines 1,839,962 articles from 256 engineering journals published 2015–2025, comparing wording and structure changes after conversational language models became public on November 30, 2022. In abstracts, 31 words associated with language model output appeared 82% more often than their pre-2023 trend projected, reaching 55.4 per 10,000 words in 2025, while four control words fell 16%. The design is descriptive rather than causal.
In briefClusLoad: Cluster-Guided Model Sharing for Scalable Transformer Load Forecasting
SummaryThis paper addresses the difficulty of forecasting loads for large numbers of distribution transformers, where device differences are substantial and maintaining a separate model per device is costly. The proposed framework learns compact temporal representations with T-Rep, groups transformers with a Gaussian mixture model, and trains one PatchTST model per cluster. On real 15-minute data from nearly one thousand transformers, it reduces MAE by about 20% versus the direct PatchTST baseline, reaches 95.66% high-load warning accuracy, and needs only seven cluster-specific models.
In briefMap2SLAM: Turning Streaming 3D Map Predictions into Real-Time SLAM
SummaryStreaming 3D foundation models estimate camera poses and dense geometry online, but their outputs cannot be revised, so loop closures cannot correct accumulated drift. Map2SLAM freezes the LingBot-Map model, trains only a lightweight retrieval head, and treats its predictions as observations feeding an external persistent SLAM state. A front-end tracks each frame and selects keyframes, while an asynchronous back-end re-observes past frames to correct the trajectory and dense geometry. It reports state-of-the-art results on KITTI, VBR, and Oxford Spires, with over 40% lower mean ATE than the best uncalibrated baselines, and real-time tracking above 28 FPS.
In briefThe Race Is Decided Early: A Kernel Law for Predicting Lock-In in Majority-Vote Test-Time RL
SummaryThe paper studies test-time reinforcement learning with majority-vote rewards, where each unlabeled question locks into one answer during training, often a wrong one on hard questions. The authors show that this outcome is decided early by a short race driven by sampling noise and derive the probability of each outcome. For networks with shared parameters, they give a closed-form kernel-based prediction, and LoRA experiments on three models show that pre-training predictions forecast lock-in well.
In briefFull Fine-Tuning Isn't the Problem: Why Adam Fails and SGD Works for Few-Shot CLIP
SummaryThis paper examines how to optimize CLIP models when only a few examples are available. Across 11 datasets, 3 CLIP backbones, and 5 shot levels, full fine-tuning with plain SGD outperforms common parameter-efficient methods, while Adam and AdamW collapse. The authors attribute this to unreliable preconditioning from small support sets, which Adam's update rule amplifies. SGD avoids this problem and stays close to the pretrained CLIP model.
WhyThe authors attribute few-shot CLIP adaptation failure to optimizer-induced preconditioning rather than full fine-tuning, reporting experiments across 11 datasets, 3 backbones, and 5 shot levels plus
In briefFast approximation of ℓp Chamfer distance using lopsided embeddings and structured JL transforms
SummaryThis paper presents efficient algorithms for approximating the Chamfer distance between two d-dimensional point sets of size n under the ℓp norm for 1 < p ≤ 2. For general p, it uses lopsided ℓp-to-ℓ1 embeddings with weak guarantees and shows these preserve the Chamfer distance, with Fast Matrix Multiplication speeding up the embeddings in high dimensions. For the Euclidean case, it applies a Toeplitz-matrix-based Fast Johnson-Lindenstrauss Transform to re-analyze a prior ℓ1 algorithm in ℓ2, achieving runtime O(nd(log log n + log(1/ε))/ε²), improving on the earlier O(nd log n/ε²) bound.
In briefZarrFM: A Standard Tensor Storage Design for Pretraining EEG Foundation Models
SummaryLarge EEG datasets differ in sampling rates, channel layouts, window lengths, and file formats, which makes preprocessing and data reading slow. The paper introduces ZarrFM, which stores an EEG corpus as a single Zarr v3 hierarchy with indexed metadata, canonical channel spaces, masks, stable subject splits, and block-aligned sharding. Across 200 recordings, it cuts metadata queries from 601 to 14 and reduces storage from 42.4 to 22.7 MB per hour compared with EDF. Block-aligned sampling reaches 3,131 windows per second, up to 11 times faster than Parquet.
In briefNot All Layers Matter Equally When Adapting Images to Video
SummaryThis paper studies how to adapt frozen image foundation models to video tasks using lightweight adapters. The authors find that temporal sensitivity increases with network depth, and that the growth pattern depends on the pretraining paradigm. They propose STRAP, a training-free placement rule that uses spatial adapters before a transition boundary and temporal adapters after it. On four fine-grained action recognition benchmarks, STRAP improves accuracy by up to 11.8% over uniform placement while reducing trainable parameters by 29 to 48%.
WhyThe authors propose training-free adapter placement from layer-wise temporal sensitivity, reporting up to 11.8% accuracy gains and 29-48% fewer trainable parameters on four fine-grained action recogni
Thinesh Thiyakesan Ponbagavathi, Constantin Seibold, Alina Roitberg
In briefDISCOVER: Online variance-guided data selection for budget-limited multimodal GRPO
SummaryThis paper presents DISCOVER, a framework that helps vision-language models improve multimodal reinforcement learning under a limited annotation budget. It clusters samples semantically, queries representative examples for reward feedback, and estimates the utility of unannotated prompts during training. Prompts are selected for annotation using a score that accounts for reward variance and difficulty. Experiments on ViRL39K and VLAA-Thinking with Qwen3-VL-2B show gains in multimodal mathematical reasoning under a 10% unique-sample budget over several baselines.
Peyman Gholami, Shivam Chandhok, Saber Malekmohammadi et al.
In briefTALK: One-Shot Batch Design for Protein Variant Effect Prediction
SummaryThis paper addresses the high cost of experimental labels in protein variant effect prediction, where the variants that can be assayed may differ from those needing prediction. It proposes TALK, which estimates each candidate's contribution through posterior coupling between accessible and target variants and reduces redundancy within a batch. In retrospective ProteinGym and FLIP protocols under limited measurement budgets, TALK generally improved prediction across the target space.
In briefSurgWorld: learning surgical representations beyond the visible view via latent prediction
SummarySurgWorld is a self-supervised framework for surgical video that learns to predict hidden surgical dynamics from partial observations. It introduces Out-of-Sight Dynamics Modeling, which predicts latent representations conditioned on spatial or temporal shifts, and a Geometric Visual Similarity objective that preserves patch-level spatial structure. Combined through an agglomerative training scheme, the model is evaluated on 17 surgical benchmarks and reportedly outperforms existing vision foundation models on workflow recognition, action understanding, and dense perception.
In briefGraphene: A lightweight image upscaling model that sharpens details using Ising-inspired refinement
SummaryThis paper presents Graphene, a lightweight Transformer for image super-resolution, which increases image resolution. It adds two refinement modules: an Ising Refinement Block that adjusts local features using signed couplings between tokens, and a History Attention module that blends current features with earlier stage outputs. On the Urban100 benchmark at 4x scaling, Graphene and Graphene-light outperform IET and IET-light by 0.24 dB and 0.38 dB while using fewer parameters, FLOPs, latency, and memory. Real-world comparisons show it recovers fine details and textures across diverse degradations.
In briefFrom Game Generation to Game Playing: Automatically Generated Training for Game-Playing AI Models
SummaryThe paper presents a framework in which an LLM-powered agent builds executable game environments and tasks, and programmatic experts use simulator state to produce matched instructions, observations, and keyboard and mouse actions. The authors post-trained a 4B vision-language model with behavior cloning on 150.3 hours of synthetic Minecraft and Roblox gameplay to create the Script2Play policy, without human demonstrations or reinforcement learning. Evaluations on several public game benchmarks show it follows gameplay instructions and generalizes across game engines.
In briefConfidence-driven adaptive semantic transmission for task-oriented communication
SummaryThis paper tackles the trade-off between communication cost and classification reliability in wireless remote image monitoring. The proposed Confidence-Driven Adaptation (CDA) lets an edge device run a lightweight classifier and choose among three transmission modes: a short predicted-class text, a compressed Region of Interest, or an uncompressed Region of Interest, depending on confidence. Numerical evaluations under bandwidth-limited, lossy, latency-constrained links show higher task accuracy at the receiver than non-adaptive baselines.
Suman Das, Prajnyajit Mohanty, Siddhartha S. Borkotoky
In briefCrossLiDAR: Controllable Multi-Sensor LiDAR Sequence Generation
SummaryThis paper presents CrossLiDAR, a transformer-based framework that generates controllable, spatio-temporally consistent sequences from multiple LiDAR sensors. It uses a rectified flow objective with a Multi-Modal Diffusion Transformer in a range-image latent space, with an inter-sensor module and a ControlNet-style branch for structural conditioning. The authors report average improvements of 29.56% in JSD and 15.36% in MMD over state-of-the-art baselines on nuScenes and the Waymo Dataset.
In briefLongCap: Timestamped captions for hours-long audio via chunk-wise processing
SummaryThis paper presents LongCap, a framework for generating timestamped descriptions of audio recordings lasting hours. It splits audio into chunks interleaved with structured event captions, using relative timestamps and retained context to extend to longer lengths. The authors introduce LongCapBench, a human-annotated benchmark, along with three metrics: Timeline Error Rate, Granularity Error Rate, and Caption Similarity. LongCap-Diar, trained on recordings up to 40 minutes, is reported to extrapolate to 16 hours of concatenated audio.
In briefOptimal polynomial-query approximation for consistent submodular maximization with short lookahead
SummaryThe paper studies consistent submodular maximization, where elements arrive one at a time and an algorithm keeps at most k arrived elements with bounded changes per update. It shows that with polynomially many evaluations, seeing future arrivals raises the best achievable ratio from 2-√2 to 1-1/e, and no polynomial-query algorithm exceeds 1-1/e. The paper gives deterministic and randomized algorithms with explicit bounds on recourse and the number of evaluations.
WhyThe authors raise the best polynomial-query ratio for consistent submodular maximization from 2-√2 to 1-1/e, with tight bounds and deterministic and randomized algorithms.
In briefHARP-Bench: Testing Robot Manipulation Alongside Active Humans
SummaryHARP-Bench is a benchmark for evaluating robot manipulation policies in workspaces shared with active humans, treating human activity as part of the task rather than background. It organizes human-present settings into four conditions (Cueing, Contribution, Interruption, and Co-presence) and provides a demonstration-generation pipeline for dynamic scenarios. Evaluations of VLA models, imitation-learning methods, and heuristic policies show that human presence challenges current policies, and higher task success does not consistently improve safety.
In briefWhat You Show Is Not What You Get: Rethinking How Visual Segmentation Follows Examples
SummaryThis paper studies how visual in-context segmentation models use example images to segment new images, and finds two problems in SegGPT: background color mapping deviations and predictions spilling from a requested part into its parent object. The authors propose Semantic Contrast Prompt Selection (SCPS), which gives the requested part and the rest of its object different colors and selects demonstrations using retrieval scores and quality estimates from labeled development images. Averaged over seven retrieval methods on three benchmarks, SCPS improves target IoU and reduces leakage into the rest of the object compared with the original binary selectors.
In briefScalable neural safety certification using monotonicity
SummaryThis paper presents a data-driven framework for verifying safety and synthesizing safe controllers for unknown monotone dynamical systems, using only simulator queries. For monotone systems with upper-closed unsafe sets, the authors prove that safety is fully characterized by the existence of monotone inductive barrier certificates. They introduce Max-Monotone Networks to learn these certificates, and report certifying safety on high-dimensional benchmarks where prior approaches fail to converge or run out of memory.
WhyThe authors exploit monotonicity to reduce neural barrier certificate verification to local boundary checks and report scaling to high-dimensional oscillator and traffic benchmarks.
