In briefObject-Only Is Not Context-Free: A Counterfactual Test for Background Removal in Object Retrieval
SummaryThis paper checks whether object-centric retrieval, which masks the background before pooling features, really produces a descriptor that reflects only the object. The authors introduce the Keyhole Aperture Probe (KAP), which keeps the object region fixed and varies only the outside pixels to measure a backbone's context dependence. They find that the pooled object region already contains its surrounding context, and across ten backbones object-only descriptors keep only about a quarter to two-thirds of full-image accuracy. A lightweight adapter reduces this dependence and improves instance retrieval on INSTRE and ILIAS, and KAP is released as an evaluation protocol.
In briefEvent-based optical flow constraint using time-sliced plane fitting for motion variation
SummaryThis paper presents an event-based optical flow estimation method. Conventional spatiotemporal planar approximation assumes constant velocity and averages out temporal velocity changes. The authors extend this approximation along the time axis by partitioning the space-time volume into sub-intervals and applying a two-step plane fit to track object edges. A spatial compensation mechanism and a temporal regularization term are added. Experiments indicate that modeling temporal motion variation improves flow accuracy over conventional methods.
In briefAnswering Without Arriving: An Evidence Audit for Embodied Question Answering
SummaryThe paper introduces a multi-room benchmark of 200 test episodes and 800 questions for embodied question answering, with an evaluation ladder that varies the evidence given to the answerer. A closed-loop vision-language agent answers 45.5% of questions without reaching any target, nearly matching the same model given no images, while replaying the reference trajectory raises accuracy to 57.4%. The authors argue that evidence acquisition, answer coverage, and shortcut sensitivity must be established before accuracy gains are credited to memory.
In briefToward Self-Evolving Narrative Systems: LLM-Based Real-Time Adaptation and Long-Term Optimization
SummaryThis paper proposes a hybrid narrative framework that balances pre-authored content with real-time LLM generation using semantic similarity alignment. It also introduces a self-evolving, feedback-driven optimization method that uses in-context learning to iteratively refine prompts without manual prompt engineering. A case study and participant evaluation indicate that the framework improves narrative coherence while preserving player agency.
In briefContinual learning in fully fine-tuned networks without replay, via hippocampus-inspired memory
SummaryThis paper addresses catastrophic forgetting in neural networks that keep learning. It proposes a training-only hippocampal-style memory that indexes the input geometry of each group of synapses sharing an input, and limits how far each synapse can move in each direction. The authors report that this lets fully trainable networks keep learning without cross-task replay, a teacher model, or inference-time memory, across four architecture families and models up to 8 billion parameters.
WhyThe authors index input geometry with a training-only hippocampal memory to limit synaptic movement, avoiding replay; they report gains across four architectures up to 8B parameters and 17.5 points ov
In briefHow language models use time anchors to recall time-specific facts
SummaryThis paper studies how language models choose the correct answer to a fact that depends on a condition such as a year. Using causal interventions on four open-weight models, the authors identify two parallel routes: a subject-mediated route where the anchor changes the subject representation, and a direct route where the final position attends to the anchor and upper-layer MLPs favor the time-appropriate object. Most queries use both routes, and their balance depends on the relation.
In briefDetecting Hidden Emotions Beyond Multimodal Conflicts Using Temporal Context
SummaryThis proposal distinguishes ordinary emotional inconsistency across modalities from deliberate emotion concealment, and infers the underlying emotion when evidence supports it. It plans to build a multimodal benchmark annotated with expressed emotion, internal emotion, and concealment conditions. It also aims to develop an MLLM-based framework that integrates fine-grained facial Action Units, cross-modal features, and temporal context.
In briefTransferring Knowledge from Related Environments to Improve Offline Q-Learning
SummaryThis paper studies how offline reinforcement learning can borrow data from related environments when target data are scarce, as in marketing and healthcare. The authors propose Direct Transfer Fitted Q-Iteration (DFQI), which jointly estimates target and source value-function coefficients within each Bellman update, linked by an L1 penalty. Using sieve approximation, the paper establishes statistical guarantees for target estimation and policy learning in stationary discounted settings with common or heterogeneous transition kernels. Empirical studies show DFQI effectively transfers source information to improve policy learning with limited target data.
In briefReWY: A Faster Way to Compute Gradients for Delta-Rule Attention Models
SummaryThis paper presents ReWY, an algebraic reformulation that reduces redundant computation in the backward pass of delta-rule linear attention models such as DeltaNet, GDN, KDA, and GDN2. The method preserves the recurrence and gradients in exact arithmetic without changing the architecture or training objective. The authors implement the reduced computation graph on NVIDIA GPUs and Ascend NPUs and report speedups in low-level backward computation relative to reference implementations.
WhyThe authors reduce seven local WY matrix products to three by eliminating auxiliary adjoints in chunkwise backward passes, reporting 1.09–2.89× speedups on H100, H200, and Ascend 950DT.
In briefVLSplat: Vision-Language Guided 3D Object Scene Representation via Scene Graphs
SummaryThis paper presents VLSplat, a method for building object-centric semantic representations in 3D Gaussian Splatting. It adds object-level semantic priors from vision-language models to mask-based 2D-to-3D lifting and refines Gaussians using an object-group scene graph. Two language-guided modules suppress semantic noise within object groups and expand object regions by adding structurally relevant candidates from neighboring groups. Experiments on indoor scene datasets are reported to improve semantic consistency and structural completeness of object-level categories.
In briefSpade-SLAM: Dense self-supervised visual SLAM using sparse correspondence anchors
SummaryThis paper presents Spade-SLAM, a dense visual SLAM framework trained in a self-supervised way without ground-truth pose or depth labels. It combines a dense warping-based photometric loss with a sparse correspondence loss from image feature matches to stabilize optimization and reduce drift in ambiguous regions. The authors report that it is more robust and accurate than existing self-supervised baselines and closely rivals systems trained with explicit pose and depth supervision.
Julian Philipp Schmitt, Weirong Chen, Ganlin Zhang et al.
In briefTriPhysBench: A Benchmark for Physical Intelligence in Embodied Agents
SummaryThis paper introduces TriPhysBench, a benchmark of 410 questions from 200 videos that tests whether embodied agents can combine physical perception, physical knowledge retrieval, and physical reasoning. The best of six vision-language models reaches only 50.38% answer accuracy. The authors also propose TriPhys, a framework that learns a unified policy from tool-integrated trajectories, which gives a 17.7% relative gain in answer accuracy for TriPhys-27B over its Qwen3.6-27B backbone on held-out questions.
In briefGeometric Morphology Learning: Giving Foundation Models Structural and Spacetime Understanding
SummaryThis paper proposes Geometric Morphology Learning, a framework that teaches models to learn the geometric or topological structures, shapes, and conformations of objects for representation learning, reasoning, and action modeling. It constructs a neural bundle linking local and global structures and defines differentiable operators such as connection and curvature. The authors argue the framework can benefit foundation models, reasoning and planning, graph modeling, physical simulation, PDE systems, and dynamics.
In briefRECEIPT: Fixing LLM Agents' Assumption That Tool Calls Finish
SummaryThe paper examines the assumption that LLM agents' command-line tool calls return in bounded time with the requested action completed, showing that commands asking a question before acting can instead return success without doing the action or hang until timeout. The authors collected production and public agent trajectories, tested how eight widely used harnesses launch commands, and built two test sets, Litmus and Tripwire. They propose RECEIPT, which keeps input open, reads each process's waiting state from system calls, and reports it back to the agent, reducing false success claims by 73.5% across five LLMs.
WhyThe authors report the first measurement of LLM agents' return assumption, proposing RECEIPT, which reads waiting state via system calls, evaluated on Litmus and Tripwire benchmarks.
In briefWatch Your Step: How Adam's Margin-Maximizing Bias Changes with Learning Rate Schedules
SummaryThis paper studies the implicit bias of Adam on homogeneous models for binary classification and its similarity to sign gradient descent. The authors show that this bias depends on the learning rate schedule, focusing on constant and weakly decaying schedules. In this regime, Adam has no trajectory-independent margin-maximization bias, but instead maximizes a trajectory-dependent weighted ℓ2 margin. In an intermediate regime, an ℓ∞ bias is proven under a gradient directional convergence condition.
WhyThe authors show Adam's implicit bias depends on the learning rate schedule, characterizing it as trajectory-dependent weighted-ℓ2 margin maximization; theoretical results are reported for deep linear
In briefDoes your long-horizon benchmark really test depth? Measuring and synthesizing depth in terminal tasks
SummaryThe paper examines long-horizon terminal tasks used to evaluate and train agents, arguing that turn counts and LLM-assigned difficulty labels do not reliably reflect task difficulty. It defines depth as the number of sequential turns required and audits 734 tasks from six existing datasets, finding that synthetic tasks mostly add work rather than depth. The authors propose DEPTH, a method that generates tasks segment by segment in a real software environment, producing 105 tasks that exceed the audited datasets in depth and work. Fine-tuning Qwen3-8B on runs from these tasks improves its performance on two terminal benchmarks.
WhyThe authors measure terminal-task difficulty by sequential depth and propose segment-wise synthesis, reporting audits of six datasets and gains on two benchmarks after fine-tuning.
In briefTAC-CoT: Tool-Empowered Implicit Chain-of-Thought with Interleaved Explicit Tokens
SummaryThis paper proposes TAC-CoT, a tool-empowered adaptive implicit chain-of-thought framework for LLM agents. It replaces explicit reasoning traces with compact latent reasoning segments, keeping text tokens only for tool calls, observations, and final answers. On agentic math and QA benchmarks, it improves accuracy by up to 8.8 points and reduces model forward passes by up to 6.8x, and it also improves tool-use quality in single- and multi-tool settings.
In briefVision-language model classifies heart-rhythm risk from colour ECG images; does describing T-waves or colour in the prompt help?
SummaryThis paper applies BiomedCLIP, a biomedical vision-language model, to classify Torsades de Pointes risk from pseudo-coloured electrocardiogram images. Using 9,789 images and five-fold cross-validation, it compares colour and colourless images under several text-prompt strategies. After training, colour-enhanced images outperform colourless ones on every metric, while naming colour in the prompt makes no difference.
In briefAgenticMobility: An Open-Source Platform for Simulating How AI Agents Shape Traffic Mobility
SummaryThis paper presents AgenticMobility, an open-source platform for studying how mobility AI agents and other AI agents interact, adapt through repeated experience, and jointly shape travel outcomes. In the current version, language model agents act as travelers choosing routes and departure times and as a manager periodically adjusting road charges in the MATSim simulator. A preliminary demonstration runs 100 simulated days across five configurations. The platform supports experiments varying information access, retained experience, and management feedback. Source code is publicly available.
In briefMachine Learning Beyond Statistical Consistency for Hybrid AI
SummaryThis paper proposes a new definition of machine learning: a machine learns a function when it computes that function successfully, meaning each output has a proper derivation in a formal system. The authors extend Gödel's Representability Theorem to empirical targets and prove that a machine learns a true probability if and only if it meets a Truth Condition (statistical consistency) and an Assurance Condition (it derives its own convergence from the record). The framework is also used to explain the hybrid nature of modern AI, which combines probabilistic and deductive tasks and pairs a statistical proposer with a deductive verifier.
In briefSurgQ-VLM: A Method for Answering Questions About Surgical Videos Using Question-Guided Compression and Answer-Type-Aware Training
SummaryThis paper proposes SurgQ-VLM, a two-stage model for answering questions about surgical videos, built on a LoRA-adapted Qwen2.5-VL-7B. The first stage trains a question-conditioned dual-stream compressor that shrinks each clip's visual tokens 49x relative to raw ViT patches. The second stage applies GRPO post-training with a reward that matches each answer's structure. On a sealed test split of 21,840 questions, the model reaches 0.779 keyword accuracy and a 10x BLEU-4 gain over zero-shot.
In briefWhich layer builds the query: how linear attention sets up retrieval in hybrid language models
SummaryThis paper studies how linear-attention and softmax-attention layers divide the work of retrieving information in hybrid language models. In Qwen3.5 and Granite4.0, the authors identify an Encode–Address circuit in which linear-attention blocks construct the query representation that a softmax-attention reader uses. Replacing this representation redirects retrieval, and the Address computation can be transferred between architectures. The authors conclude that linear attention constructs the query while softmax attention performs the lookup.
In briefDoes reflective LLM feedback help under a wet-lab-scale label budget?
SummaryResearchers replayed 36 ProteinGym assays, acquiring 96 labels per round over six rounds to test whether language-model feedback improves a biological predictor. Full LLM control showed no detectable advantage over random edits in test rank correlation. Exploratory ablations found that LLM model editing is useful, while the added value of repeated reflective feedback remains unestablished.
In briefOnline learning design for safety-critical control systems
SummaryThis paper proposes Control AI, an architecture for Online Active Learning (OAL) inspired by the prediction-evaluation-correction process of biological intelligence. It combines deep neural network representations with real-time error-feedback adaptation from adaptive control. The framework uses prediction, latent transition, and reconstruction models, with residuals between predictions and measurements serving as the learning signal. The paper focuses on the architectural rationale and the progression from a conceptual feedback loop to practical large-scale configurations.
In briefPrivacy-preserving federated estimation of shared loading spaces in tensor factor models
SummaryThis paper studies how to estimate common loading spaces in Tucker tensor factor models when data is distributed across clients under federated differential privacy. Each client applies private stability testing, Huber aggregation and Gaussian perturbation, and the server performs weighted aggregation. The authors give non-asymptotic error bounds governed by a scaled harmonic mean of local error rates and prove minimax optimality up to logarithmic factors under strong factors. They also propose an adaptive private calibration procedure that does not require knowing the noise level, and illustrate performance with simulations and a Meditation Traits EEG dataset.
In briefTraining-free 3D brain MRI motion correction using event-guided reference images
SummaryHead motion causes artefacts in 3D brain MRI that can bias structural measurements. The paper proposes EventMC, a training-free method for interleaved Cartesian acquisitions that iteratively detects motion events and estimates motion using reference images from stable intervals. In simulation, its motion estimates improve mean PSNR by up to 4.0 dB over the learning-based MotionTTT, and a single run is up to 7.5 times faster with less than half the GPU memory. On in vivo motion-corrupted scans, it achieves similar mean image quality to MotionTTT without training.
In briefA systematic study of how robust fine-tuned image classification models are against attacks
SummaryThis paper systematically evaluates the robustness of fine-tuned pre-trained models across different datasets and attack types for image classification. It tests models on MNIST, CIFAR10, CIFAR100, and SVHN before and after various attacks, comparing ResNet50, DenseNet, and MobileNet. The authors find that robustness varies with data distribution and sample size, so model selection should be tailored to the specific task.
In briefRobust remote ECG monitoring with limited energy: comparing anomaly filtering and predictive uncertainty
SummaryThis paper studies whether unsupervised anomaly detection placed before a classifier can make remote ECG monitoring on low-power microcontrollers more efficient. It compares five detectors, including Deep SVDD, against Monte Carlo Dropout and deep ensembles on PTB-XL and BUT QDB under unseen-disease and noise conditions. Deep SVDD performs best across most tested CPU energy budgets, and its estimated energy per input is lower than that of the compared uncertainty methods. The authors conclude that upstream filtering supports efficient monitoring, while noting that independent validation and complete-system measurements are still needed.
Mustafa Fuad Rifet Ibrahim, Maurice Meijer, Alexander Schlaefer et al.
In briefRethinking model aggregation for federated learning under domain skew, from a knowledge-preserving view
SummaryThis paper addresses domain skew in federated learning with FKPL, a two-stage framework. Clients are clustered, domain-specific expert encoders are trained within clusters, then concatenated and frozen, and a global linear classifier is calibrated using class-matched domain prototypes. Reported average accuracies are 73.01% on Office-Caltech-10, 67.00% on PACS, and 64.51% on a four-domain DomainNet subset, and the authors conclude that classifier calibration effectively combines preserved domain representations.
In briefFaster 4D MRI Reconstruction Using Prior-Guided Deep Learning
SummaryThis paper proposes PGPI4MR, a deep learning method that uses information from other respiratory phases or a setup scan to reconstruct 4D MRI without extra scan time. In in vivo tests, it had the lowest held-out error at every scan time and reduced error by 15.5% versus the best comparison method for 0.49 s sub-second MRI. Without retraining it matched the best methods on 3T abdominal scans, and after retraining on thigh data it reduced held-out error by 6.6%.
Mojtaba Safari, Parag J. Parikh, Shansong Wang et al.
In briefAsking Clarifying Questions Instead of Guessing: A Benchmark and Training Method for LLMs
SummaryThe authors find that fewer than 11% of incomplete user queries prompt seven large language models to ask a follow-up question. They introduce VagueBench, a benchmark that creates intent-vague and fully-vague query variants by removing linguistic cues from real queries, and DialogGRPO, an offline reinforcement learning method that trains models to ask questions over a dialogue-state graph. The method recovers missing information on over 88% of vague queries without over-asking on complete ones, and it transfers zero-shot to MultiWOZ 2.4 and AmbigQA.
In briefEfficient Lempel-Ziv Prediction Using Run-Length BWT
SummaryThis paper presents efficient online implementations of Kosoy's Lempel-Ziv Plurality (LZP) next-symbol predictor and a new variant. The first implementation maintains a run-length encoded Burrows-Wheeler transform of the reversed history, reducing update time from Kosoy's O(z log^7 n) to O(log z + log log n). The authors also introduce the Lempel-Ziv Majority (LZM) predictor, which has the same mistake bound and predicts in constant time independent of alphabet size.
In briefAutomated stateful specialization for adaptive agent systems
SummaryThe paper introduces ASpec, a framework that first discovers specialist agent archetypes through evolutionary search and then builds their expertise through accumulated experience. It also proposes a lightweight hierarchical control policy, "retain-then-escalate," which decides when to reuse the existing agent system and when to adapt its structure. Experiments report significant gains on expert-level scientific benchmarks such as GPQA and results matching the state of the art on broader domain tasks.
In briefA self-generated floor on Adam's ε: evidence from power-law matrix sensing
SummaryThis paper studies when the Adam optimizer stalls in nonconvex low-rank matrix sensing with power-law covariance. Lowering the numerical stability term ε₀ does not move the failure boundary past about 10⁻⁸, which the authors attribute to a dynamical floor proportional to the step size α. Reducing the step size lowers this floor directly, while shrinking ε₀ below 10⁻⁸ has no effect.
In briefDesigning fair experiments that give reliable results for every subgroup
SummaryThis paper introduces a Fairness-Aware Contextual Track-and-Stop Design (F-CTSD) algorithm for running experiments across different subgroups. It aims to keep required sample sizes small while providing statistical guarantees on subgroup fairness. The authors analyze the trade-off between fairness and efficiency. Numerical experiments show higher sample efficiency than existing methods and a 4.95% reduction in subgroup fairness violations.
In briefThe (Butter)fly Dream: Fly-Brain Model Tells Real Video from World-Model Dreams
SummaryThe paper places a published fly visual-system model, trained only for optic-flow estimation, inside videos generated by world models. A classifier two-sample test on cell-type responses asks whether real footage can be distinguished from dreams continued from five real frames. After training, all nine dreams were separable from reality (AUC 0.85 to 0.99), far more than before training. Order-only probes further measure how sensitive the responses are to small temporal defects.
In briefA study on whether quantizing recurrent language models is easier than expected
SummaryThis paper studies how to quantize linear recurrent language models, whose numerical behavior differs from softmax-attention Transformers. It separately analyzes weights and activations outside the recurrent state and the recurrent state itself during decoding. The authors propose LoReS, a state quantization pipeline combining value-axis scale grouping, decay-based mixed precision, and stochastic rounding. The method reportedly supports robust 8-bit state quantization across several models and yields end-to-end decode speedups in vLLM.
In briefOverview of recent progress in agent-based computer graphics
SummaryLarge language model agents can now generate graphics code and refine it iteratively through feedback. This report introduces a unified taxonomy for describing agentic systems across diverse computer graphics tasks, and an agentic literature system that helps researchers analyze existing and new papers at scale. It analyzes emerging research trends and discusses open challenges in coordinating agents and tools, human-agent collaboration, long-horizon planning, and evaluating interactive graphics systems.
In briefLearning to Memorize: A Learned Way to Manage Memory in Long AI-Generated Videos
SummaryAutoregressive video models often lose quality when generating long videos, because errors build up across frames. The paper proposes L2M, a learned framework that decides which past frames to keep, remove, or update based on predicted importance. It reports better quality and temporal consistency for long videos than existing baselines under a fixed memory budget.
In briefPrivacy-preserving mood-swing detection from heart rate variability using federated learning on edge devices
SummaryThis paper proposes a privacy-preserving mood monitoring framework that classifies emotional states from heart rate variability features derived from ECG. It is evaluated on three public datasets with 74 participants and validated in a pilot study using a wearable device. Logistic Regression outperformed five alternative classifiers under leave-one-participant-out cross-validation. Federated learning on a Raspberry Pi cluster achieved performance comparable to centralized training without sharing raw physiological data.
Chidike Vincent, Muhammad Aslam Jarwar, Najam Ul Hasan et al.
In briefPredicting irregular clinical time series while accounting for missing data
SummaryThis paper proposes MA-PFN, a Missingness-Aware Prior-Data Fitted Network, for predicting irregularly observed clinical laboratory time series, where the pattern of missing measurements is itself informative. A diffusion model first learns a prior over joint value and observation-mask trajectories from source-domain EHR data, and a PFN-style predictor is trained on synthetic trajectories drawn from it. The model performs imputation and forecasting for new admissions in a single forward pass without target-domain adaptation, and achieves the best results across internal MIMIC-III and external eICU evaluations.
Hadi Mehdizavareh, Gabriele Santangelo, Giovanna Nicora et al.
In briefBiological Congruence: Checking Whether Generated Microscopy Images Agree With Biology
SummaryThis paper proposes biological congruence, an evaluation principle asking whether biological information stays consistent across generated microscopy images, instance masks, and text conditions. The authors build a text-conditioned diffusion model that jointly generates images and masks, then compare agreement across several analyses. Results show that performance depends strongly on the feature measured, and that generative quality, controllability, cross-representation agreement, and population fidelity are complementary properties.
Duway Nicolas Lesmes-Leon, Shumpei Takezaki, Fabian Schmeisser et al.
In briefFADE: A flow-based method for estimating rare multi-receiver signal drops
SummaryThis paper proposes FADE, a flow-based approach for modeling how rare extreme events are dependent across multiple receivers in ultra-reliable low-latency communication. FADE transforms multivariate extremes into a standard extreme-value representation and learns the dependence structure with a flow-based density estimator, without a predefined parametric form. Evaluated on experimentally collected UAV communication data with two ground receivers, it agrees more closely with the empirical joint CDF, reducing MAE by up to 22.4% compared with the Logistic and MEL methods.
Parmida Valiahdi, Niloofar Mehrnia, Mustafa Furkan Beker et al.
In briefZ-Cache: Speeding Up Diffusion Transformers via Self-Reflection
SummaryThis paper proposes Z-Cache, a feature caching method that speeds up diffusion transformers used for visual generation. It predicts features for future caching steps, performs a full computation, and then re-predicts the cached steps using the earlier and current results to correct errors. The authors report generation quality comparable to the original model with faster inference, including a 6.22X speedup on FLUX-dev for text-to-image generation.
In briefOpenNWM: A General Navigation World Model Using Latent Action Pretraining
SummaryOpenNWM is a general navigation world model that predicts future observations based on navigation actions to support planning. It first learns universal latent actions from large-scale action-free videos through visual prediction pretraining, then aligns them with physical motion via action-conditioned post-training. The authors built NavAnywhere, a dataset of 17.5 million visual observations from 15 sources, covering ground robots, humans, and drones across five scene types. Experiments show improved visual prediction and navigation performance.
In briefMEX: A Benchmark Suite for Testing Models of Musical Expectation
SummaryThis paper introduces MEX, a benchmark suite for checking whether computational models of musical expectation match human perception. Such models estimate measures like surprisal and uncertainty from symbolic music or audio. MEX collects and extends existing tasks from music information retrieval and cognitive science on standard datasets, testing whether these measures decrease with repetition, peak at segment boundaries, agree between audio and symbolic models, and predict electrophysiological responses. The authors present it as an evolving interdisciplinary effort rather than a complete characterization, and provide open-access code.
Mathias Rose Bjare, Augustin Bouquillard, Giorgia Cantisani et al.
In briefSelf-evolving AI workflows that break complex tasks into reusable steps
SummaryThis paper proposes a framework in which AI agent workflows break complex tasks into semantic subtasks, each with stated applicability conditions and a procedure that runs code or calls a model when needed. The agent refines the decomposition and updates these procedures using execution feedback. Experiments on bargaining, negotiation and persuasion in the GLEE competition found that most final-decision procedures were implemented in code. The code is publicly available.
In briefNEIGHBOR: Aligning Service Behaviour and Evidence Use in a Community Resident Services LLM Agent
SummaryThis paper presents NEIGHBOR, an LLM agent that answers community residents' repair and service-charge questions by training its service conduct into the model driving its tools, rather than relying only on a system prompt. The agent adapts Qwen3-235B-A22B with LoRA and is deployed on a live resident-service platform in an East Asian city district. In an offline evaluation, a domain expert scored 350 answers, with NEIGHBOR reaching 7.74/10 versus 5.86 for the un-tuned model, improving all six rated dimensions.
In briefReFrame: Silence-Aware Compute Reclamation for Full-Duplex Speech Models
SummaryFull-duplex speech models run dense inference on every audio frame during live conversation, which keeps GPU load high. ReFrame treats this as a silence-aware computation problem, using Mutual Silence Speculation when both sides are silent, Listening Frame Folding while the user speaks, and Frame-Budget Reinvestment to hand freed frame time to external retrieval and language-model workers. The paper reports that across 1,225 FDBench conversations, ReFrame cuts measured GPU execution time by 18.80% to 31.01%.
In briefLegoEnv: Agent environments that evolve by recombining reusable modules
SummaryThe paper proposes building LLM agent environments from reusable atoms (intent, rule, data, and server) and evolving these atoms when an environment fails to give reliable feedback. The authors instantiate this as LegoEnv, which builds 5,465 verified environments over 223 applications. One evolution round raises the share of solvable environments from 71.6% to 88.3%, and training on these environments improves model scores on AutomationBench.
In briefCounting and listing the logically distinct readings of natural-language rules.
SummaryThe paper studies how bracketing of connectives in human-written rules creates ambiguity, modeling each reading as a derivation tree of a deliberately ambiguous context-free grammar. For the positive, repetition-free fragment, it characterizes realizable read-once normal forms and gives a two-label recurrence for their count, verified by exhaustive enumeration up to eleven atoms. It also shows that a skeleton is ambiguous exactly when it uses both conjunction and disjunction, and gives a lower bound on the number of classes.
In briefLUMEN: A geometry-guided diffusion model that estimates spatially varying HDR lighting in scenes
SummaryThis paper presents LUMEN, a geometry-conditioned diffusion framework that synthesizes high-dynamic-range (HDR) environment maps at arbitrary spatial locations from single or multi-view images. It reconstructs a coarse 3D scene, renders partial environment maps to encode visibility and spatial structure, and fuses these with multi-view image features in a diffusion transformer to produce full HDR maps. The authors built a synthetic dataset of 700 indoor scenes with dense spatial supervision for training. Experiments show improved visual fidelity and spatial coherence over prior approaches.
In briefRCFormer: A Reliability-Aware Transformer for Facial Landmark Detection Under Occlusion
SummaryRCFormer is a transformer-based method for detecting facial landmarks when parts of the face are occluded, poorly lit, or posed at large angles. It improves recovery of unreliable facial patches using three components that account for visibility and spatial coherence. The recovered heatmaps serve as reliability-aware structural cues for existing landmark detection networks. The paper reports improved heatmap recovery and strong localization results on the WFLW, COFW, and 300W datasets.
In briefCoGSeq-LTV: A scalable graph-sequence method for predicting customer value across businesses with missing data
SummaryThis paper studies multi-business, multi-horizon value prediction on transaction logs from large super-apps, where item identifiers are partly missing and targets are sparse and cumulative. The proposed CoGSeq-LTV method combines a batch-built user-item graph, business-aware sequence encoders that keep recency and payment context, and cross-business contrastive alignment. It uses a zero-inflated mixture log-normal head to model long-tailed targets. Experiments on the BiBiSheng transaction-log dataset report improvements over representative LTV baselines across businesses and horizons, including a later out-of-time evaluation window.
In briefCan latent prediction error in video world models really signal physical surprise?
SummaryThe paper evaluates whether prediction errors from video world models truly reflect physical surprise rather than appearance or motion. On the IntPhys2 benchmark, it compares latent prediction, masked pixel reconstruction, and simple visual statistics, finding appearance is easy to decode while V-JEPA2 gives only modest information about physical possibility. Experiments on Physion++ and Kubric show that prediction error is affected by factors like frame repetition and color changes, and impossible events can yield lower error than possible ones, so the error does not reliably measure physical surprise.
In briefControl barrier functions for safe robot motion planning in latent space
SummaryThe paper addresses obstacle avoidance in latent-space path planning for robotic manipulators. It replaces a binary collision classifier with a neural control barrier function learned in the latent space and enforces it at planning time with a closed-form first-order correction. On a 7-DoF Franka Panda, the method reaches 75.6% success on a fixed-obstacle task, 12.0 points above the best of 35 tuned classifiers. Success falls to 33.0% in a random-obstacle setting.
In briefDomain-aware federated learning that helps drone swarms train across varied environments
SummaryThis paper proposes domain-aware hierarchical federated learning (DA-HFL) for unmanned aerial vehicle swarms operating in changing and heterogeneous environments. It combines mobility-informed clustering, cluster-level aggregation, cloud-based Fisher-weighted merging, a domain-adversarial shared backbone, and environment-specific heads. Theoretical analysis covers resource feasibility, optimality, and complexity bounds. Simulations across four heterogeneous environments show faster convergence, better accuracy during environmental transitions, and more stable clusters.
Taha Mahmood, Mudassar Shakeel, Ahmed Naeem et al.
In briefShared relations and specific choices: how language models combine common and query-specific effects
SummaryThis paper uses controlled entity-binding experiments on GPT-2 Small, Qwen2.5-1.5B and Llama-3.2-1B to study how a relational representation shared across queries shapes query-specific choices. The shared component transfers to new names, but a transferable representation does not guarantee a transferable functional sign, and its selection effects vary across models and prompts. Pathway-clamping analyses show that opposing pathway contributions can partially offset each other.
In briefRIFT: A residual-aware flow transfer method for one-step generative video compression
SummaryThis paper proposes RIFT, a framework for one-step generative video compression that uses pretrained spatiotemporal priors to recover details lost during coding. It introduces a Residual Geometry Inducer that converts the pretrained flow prediction into a content-adaptive restoration update using decoder-visible codec cues. Experiments on three public video compression benchmarks show that RIFT outperforms competing generative video compression methods.
In briefDetection Instead of Inversion: Using Detection Transformers for T1-T2 Microstructure Quantification
SummaryThis paper treats T1-T2 correlation MRI microstructure quantification as an object detection task. A Detection Transformer (DETR) takes 64 TI/TE signals as input and directly predicts a set of compartments defined by T1, T2, and signal fraction, avoiding the traditional full spectrum reconstruction. On simulated data it achieves a mAP@10 of 76.8%, above the authors' previous model (53.8%) and regularized NNLS (66.1%), and reconstruction is orders of magnitude faster than NNLS.
In briefOCBI: Order-Sensitive Cross-Branch Interaction for Image Manipulation Localization
SummaryThis paper proposes OCBI, a method for locating manipulated pixels in images that extends the Co-Transformers model. It models the difference between two feature composition orders by adding a learnable signed correction to feature transfer, followed by pixelwise adaptive multiscale aggregation before prediction. On five clean test datasets, OCBI reaches a macro PixelF1 of 0.5773, 0.43 points above the baseline. Under JPEG compression, Gaussian blur, and resizing, its average PixelF1 also exceeds the baseline by 0.95 points.
In briefMemRole: Turning Memories into Decisions for Long-Horizon Role-Playing AI
SummaryThis paper presents MemRole, a framework for LLM-based role-playing agents that makes consistent decisions over long narratives by reasoning from past memories. It introduces Long-Horizon Memory Thinking (LHMT), which breaks reasoning into context understanding, memory activation, and response strategy planning, trained on a novel-derived dataset with timeline-constrained memory pools. The authors also propose Self-Refining On-Policy Self-Distillation (SR-OPSD) for token-level guidance from self-critique. The paper reports that MemRole-32B improves over its backbone by 15.2 points on memory-grounded role-playing and achieves the best character-motivation recognition accuracy on CroSS-MR.
In briefSpectral Signatures of Memorization Basins in Diffusion Trajectories
SummaryThis paper studies how diffusion models reproduce individual training examples, a phenomenon called memorization. The authors give a geometric characterization showing that when a denoising trajectory enters a memorization basin, the local spectrum shows a dominant singular mode followed by a sharp recovery. This signature appears both during inference trajectories and over training. The authors use these geometric signals to suppress memorization during training and to steer trajectories away from basins at inference, and report that the approach extends to large text-to-image diffusion models.
Achref Jaziri, Martin Mundt, Ahmed Imtiaz Humayun et al.
In briefDiagnostic checklist for silent reward and environment defects in EV charging simulation benchmarks
SummaryThe paper audits a five-agent reinforcement-learning benchmark for electric vehicle charging assignment, built on SUMO and OpenStreetMap road networks. It identifies five silent defects that do not crash the simulation but bias comparisons, and quantifies three of them. It also proposes an eight-item checklist as a low-cost first-pass audit.
In briefFinishStrong: Ending Video Clip Localization Well After a Flawed Start
SummaryThis paper addresses a weakness in autoregressive video temporal grounding, where models predict clip endpoints much worse than start points. It proposes FinishStrong, a two-stage post-training framework: Online-start SFT trains endpoint prediction on the model's own generated contexts, and Conditional-tree GRPO computes end rewards only within the same-start branch to separate end quality from the start penalty. The authors report 61.7% overall IoU on TimeLens-Bench for FinishStrong-9B and a 36.5% reduction in subsequent end prediction error when continuing from imperfect starts.
In briefA VR-based remote control system for service robots that maps human body pose to robot movement.
SummaryThis paper presents a teleoperation system for service robots that uses virtual reality and human pose estimation for remote control. Mediapipe detects 2D poses, which are extended to 3D using RGB-D data from an Intel RealSense D455 camera and mapped to the robot's kinematics. Real-world experiments on the CHARMIE service robot tested object manipulation tasks. The authors note that accurately capturing some joint angles, such as wrist yaw, remains challenging.
Eduardo Fernandes, Tiago Ribeiro, Antonio Ribeiro et al.
In briefA bounded protocol for testing whether neural networks internally represent parts systematically
SummaryThe paper proposes Bounded-RS, an evaluation protocol for checking whether a neural network's internal representations, not just its outputs, are systematic. For an explanation attributing systematic behavior to a declared representation, it tests on matched behavioral contrasts whether the representation is recoverable and causally used, then assigns one of four bounded verdicts. On 96 semantic-parsing sets released by the authors, a role-filler binding explanation for fine-tuned T5 models receives the Causal-use verdict, while a distributed geometry explanation is Unsupported within its declared search space.
In briefOptimization Frameworks and Learning Algorithms for Discrete-Time Mean Field Games: a monograph on computing equilibria in large multi-agent games.
SummaryThis monograph studies discrete-time mean field games with finite state and action spaces, starting from Markov decision processes and finite-player stochastic games. It develops mean field Nash equilibria and establishes existence and uniqueness under suitable assumptions. Using linear programming, occupation measures and primal-dual formulations, it connects equilibrium computation to optimization and presents fixed-point iterations, fictitious play and online mirror descent with their convergence properties. Numerical examples cover crowd motion, traffic routing, epidemic management and multiple equilibria.
In briefReCoG: Relative Node Contexts for Pre-training Graph Models Across Domains
SummaryThis paper proposes ReCoG, a method for pre-training a single graph model on data from multiple domains by defining node contexts with criteria that are consistent across different graphs. The authors observe that removing many edges from source graphs during pre-training improves cross-domain transfer, and they attribute this to conflicts arising from the raw graph structure. ReCoG builds attribute, structural, and semantic contexts and trains one graph encoder by reconstructing node attributes over them. On 18 target datasets, it improves mean 1-shot cross-domain node classification accuracy by 7.61 percentage points over the strongest baseline.
In briefA deep reinforcement learning system for managing real-time electric vehicle charging in smart cities
SummaryThis paper proposes an edge-intelligent architecture in which electric vehicles pick charging stations by querying a shared Deep Q-Network policy over an MQTT publish-subscribe backbone. The authors simulate the approach in SUMO on two real road networks from Morocco (Rabat and Tangier) and compare four decision policies: an adaptive heuristic, tabular Q-learning, DQN, and Double DQN. Under evening peak demand, the DQN agent improves charge rate by 6.8% and delivered energy by 5.4% over the adaptive heuristic, but shows no waiting-time advantage, and the ranking is not stable across the two networks.
In briefOnline Meta-Control for Self-Evolving LLM Agents Using Experience Memory
SummaryThis paper introduces OMCE, a framework that updates an experience memory and learns a meta-control policy from task feedback during test-time adaptation. A lightweight contextual-bandit meta-controller selects among four actions (DIRECT, EPISODE, GUIDE, DISTILL) to decide whether to use experience, which memory format to apply, and when to consolidate new guidance. Experiments on four online adaptation benchmarks across four backbones report up to 20.5 percentage points higher first-attempt success than the strongest baseline and 5.7 to 62.5 times lower per-task token use than ACE with GPT-5.2.
In briefLarge-Scale Convex Regression: Sparsifying Constraints While Preserving Geometry
SummaryConvex regression is a nonparametric method that uses shape information, but its standard least-squares form has O(n²) supporting-hyperplane constraints, which becomes computationally costly as sample size grows. This paper proposes a K-nearest-neighbor localized relaxation that keeps only the K nearest hyperplane constraints for each sample. The authors establish statistical guarantees in every fixed dimension d≥2. Experiments on synthetic convex benchmarks show that the KNN method's prediction risk approaches that of the full estimator at neighborhood sizes matching the theoretical scale.
In briefScaling laws of partner feedback and convention rigidity in meta-learned multi-agent reinforcement learning
SummaryThis paper studies whether meta-learned reinforcement learning algorithms, discovered automatically, keep improving when trained alongside teammates that also adapt, or whether they get trapped in brittle conventions. Over 1,800 runs in continuous robotics (VMAS) and discrete coordination (Overcooked), the authors report five laws of multi-agent adaptation. These cover evaluation illusions, feedback coupling, training length, cross-partner performance, and diagnostic signals.
WhyThe authors report scaling laws for partner feedback and an evaluation illusion, tested over 1,800 runs in VMAS and Overcooked.
In briefHow Chain-of-Thought Training Teaches Models to Reason: A Theory
SummaryThis paper studies how training with chain-of-thought (CoT) supervision shapes a model's reasoning at inference time. The authors introduce an in-context weight prediction model trained on finite gradient descent trajectories for linear regression, and use high-dimensional asymptotic analysis to examine when long-horizon reasoning succeeds or fails. They distinguish two mechanisms of overthinking, error accumulation and error amplification, and report experiments with linear and multi-head softmax attention models that show these behaviors persist.
WhyThe authors introduce an in-context weight prediction model and use high-dimensional asymptotics to show phase transitions in long-horizon reasoning, with experiments on attention models.
In briefRSI-Env: An Open-Ended Environment for Agents That Improve Themselves Recursively
SummaryThis paper introduces RSI-Env, a persistent environment where agents autonomously improve their models and harnesses, with each accepted version carrying out the next improvement round. Three settings vary goal specificity and benchmark feedback, and the Self approach is compared with two fixed external Teachers on seven tasks starting from Qwen3.8-27B. Without benchmark feedback, the Open-General Self setting raises GPQA Diamond accuracy from 65.2% to 73.2% and AIME 2026 accuracy from 53.3% to 56.7%.
In briefCacheWeave: Learning Which Tokens to Recompute for Faster Position-Independent KV Caching
SummaryThis paper addresses a problem in position-independent KV caching, where encoding chunks separately drops cross-chunk interactions and can reduce accuracy. It proposes CacheWeave, which adapts the backbone model to this caching scheme, freezes it, and trains a block selector on the adapted model's oracles to choose which contiguous token blocks to recompute. The companion runtime, CacheWeave-Serve, supports private and shared KV caches and keeps selection and recomputation on the GPU. At a 5% recomputation budget, the method outperforms CacheBlend and ProphetKV on average.
In briefLess but Sufficient: Relation-Guided Compact Knowledge-Graph Contexts and Limits of External Checks
SummaryThis paper studies how large language models read knowledge-graph facts. The authors use relation knowledge, either hand-written or learned from training questions labelled with relation paths, to select a compact set of facts and derive an answer. They add a derivability check, a consistency check, and a routing gate that decides when the derived answer replaces the LLM's answer. Across seven open LLMs, the compact context gives better exact-match results than relevance-ranked facts, while using about a quarter of the prompt tokens on ComplexWebQuestions.
In briefRobust EEG emotion recognition from coarse trial-level labels despite emotional shifts within trials
SummaryThis paper addresses EEG emotion recognition, where each long trial has only one global emotion label that is copied to all its short windows, even though emotion can change within a trial. The authors propose VARY, a framework combining a two-stage sample selection strategy to identify windows with reliable labels and MixAug, an EEG-specific data augmentation method. Tested on two public datasets with three network architectures and several label-noise settings, VARY achieves accuracy gains of up to 11.54% on DREAMER and 5.39% on SEED over the best baseline.
In briefA review of deep learning methods for automated cervical cancer cell screening.
SummaryThis review surveys deep learning research for cervical cytology screening, organizing the field by data scale, task and supervision. It compares performance across 16 public cytology datasets and assesses commercial systems, regulatory status and low-resource deployment. The authors argue that single-cell benchmarks have plateaued and that progress requires multi-center benchmarks, external validation and pragmatic clinical trials.
In briefMAC-Bench: Testing How AI Agent Teams Collaborate on Linked Tasks
SummaryThis paper introduces MAC-Bench, a benchmark for evaluating how well teams of language-model agents collaborate. Instead of fixing agent roles or communication rules in advance, it builds collaboration requirements into task dependencies and shared-resource constraints. Reusable local collaboration patterns and global task structures are mined from selected tasks and combined into task graphs. The tasks run in executable environments that automatically verify partial progress and full completion.
In briefUncertainty Estimation for Data with Missing Values, Without Imputation
SummaryThis paper proposes a conformal prediction method for quantifying uncertainty on data with missing values, without imputing the covariates. Each input is represented by a tree embedding together with its raw missingness mask, and a regularized quantile score function is learned. The method provides guarantees at several validity levels across general missingness mechanisms (MCAR, MAR, MNAR). Experiments on synthetic and real datasets show reduced miscoverage risk while keeping intervals informative, especially in non-MCAR settings.
Jiarong Fan, Juhyun Park, Thi Phuong Thuy Vo et al.
In briefA sensor-driven physics-informed neural network for real-time, occupant-centered space heating control
SummaryThis paper proposes a control-oriented physics-informed neural network (PINN) that predicts indoor temperature fields from spatial coordinates and limited sensor measurements. Unlike conventional PINNs that rely on explicit time inputs, the model infers temporal dynamics implicitly to handle variable initial conditions and heating sequences. The method was validated on nine horizontal sensing planes across three room-scale configurations, with a maximum relative error below 1.85% against a validated COMSOL simulation. The predictions were used to support occupant-centered heating control.
Yunfei Bai, Chenghao Li, Sunku Prasad Jenne et al.
In briefDelivered but Unused: Assessing Agent Memory at the Moment of Use
SummaryThis paper introduces MemUse, a method that evaluates an LLM agent's memory at the point it is delivered, by examining recorded trajectories and the agent's later actions and their consequences. Across 8,000 SWE-bench Verified episodes from three memory systems, resolve rates stayed within 0.4 percentage points of a no-memory baseline. MemUse found that most delivered entries were irrelevant and most relevant entries went unused.
In briefEarlySelectKV: Precomputing Sparse KV Selection Using Proxy Queries
SummaryThis paper introduces EarlySelectKV, a training-free method that uses proxy queries from earlier hidden states to precompute selected KV indices, while the real target-layer query is still used for final attention. Tests with RocketKV, Quest, and Loki show no consistent quality degradation across the evaluated settings. In a TensorRT-LLM deployment, ES-RocketKV improves end-to-end output throughput by up to 12.4% and reduces GPU-resident memory by 11.1%.
In briefRoute-guided CBF-QP repair to make vision-language-action robot policies safer
SummaryVision-language-action (VLA) models can produce task-oriented robot actions but do not guarantee safe execution in cluttered scenes. The paper proposes Route-Guided CBF-QP Repair, a post-hoc module that uses the VLA's short-horizon action chunk to detect blocked trajectories and generate entry-exit-rejoin routes around obstacles. The route is added to the CBF-QP objective as a soft guidance term while the CBF safety constraint stays hard. On SafeLIBERO, the method improves collision avoidance, task success, and SafeSuccess compared with AEGIS.
In briefWhen Labeling Disagreements Make F1 Rankings Uncertain
SummaryThis paper examines whether two models' F1 rankings can be determined after auditing labeling disagreements. Using two agreement counts, the authors derive exact bounds that certify a ranking, or a choice within a specified F1 tolerance, under every possible completion. Across 3,200 frozen model pairs, shared completions tighten the bounds, and a 0.01 tolerance lowers labeling costs. The guarantees apply only to the evaluated batch, not to future data.
Roberto Fern\'andez-Barrios, Iker Pastor-L\'opez, Asier Gonz\'alez-Santocildes et al.
In briefCST: Lightweight Risk Triage with Source-Grounded Evidence for Web Discussions
SummaryCST is a lightweight pipeline that helps moderators find risky messages in web discussions by combining character patterns with multilingual sentence meaning, running on an ordinary CPU. It groups related messages, selects representatives from each group, and links every pick to its original thread. The paper reports that across six tasks, CST ranks harmful messages more accurately than a fully fine-tuned multilingual model and finds more under the same reading budget.
In briefComparing Methods for Imbalanced Drug-Activity Regression, Tested on hERG Prediction
SummaryThis paper examines data imbalance in QSAR regression, a area less studied than classification. The authors compare data-level and algorithm-level strategies across GCN and fingerprint-based machine learning models. They divide the target range into subregions using domain knowledge and data distribution, and focus evaluation on the region of greatest practical interest. They report that selecting models by this region-based metric is more reliable than using the full range, and that strategy effects differ by model type.
In briefSMART-AIMC: A Sensitivity-Aware Mixed ADC Resolution Framework for Analog In-Memory Computing
SummaryThe paper proposes SMART-AIMC, a framework that allocates mixed ADC resolutions across analog in-memory computing units. It introduces an ADC Channel Sensitivity metric that links task loss to ADC quantization noise, directing high-resolution ADC resources to critical computations. A sensitivity-based channel reordering scheme addresses hardware granularity constraints. The authors report a better accuracy-energy trade-off across workloads and, near digital-level accuracy, 6.99× modeled system-level energy efficiency for linear projections over an INT8 digital baseline.
In briefFréchet Oral Distance: A Metric and Benchmark for Judging Speech-Driven Mouth Animation Without Coordinates
SummaryThe paper proposes the Fréchet Oral Distance (FOD), a metric for evaluating speech-driven oral animation quality. It uses embeddings from an Oral Motion Encoder and a TIMIT-Vis reference distribution instead of coordinate measurements. In experiments, FOD produced model rankings similar to Lip Vertex Error and showed trends consistent with user preferences. The framework enables objective, reproducible benchmarking and can be extended to compare both 2D and 3D models.
In briefMatMoE: Matryoshka Mixture-of-Experts with Dynamic Mixed-Precision Quantization for Efficient Inference
SummaryThis paper presents MatMoE, a dynamic mixed-precision quantization framework that makes Mixture-of-Experts language models more efficient to run. It uses a Matryoshka-style quantization that lets experts operate at multiple precisions, a floor-and-compensation method to avoid redundant memory accesses, and a bit-width allocation based on routing scores to meet changing latency needs. The authors report 4.0% to 18.3% higher accuracy than a state-of-the-art static mixed-precision method across three MoE models, and 1.2x to 3.7x speedups over llama.cpp on Raspberry Pi 5, NVIDIA RTX 4090, and Jetson AGX Orin.
In briefSparseLift: Automatically Optimizing Opaque Sparse Operations in Large-Scale Recommendation Model Training
SummaryThis paper presents SparseLift, a compiler that optimizes the sparse path of large-scale recommendation model training. It turns sparse operations and communication, which are usually opaque nodes in the computation graph, into analyzable subgraphs so that whole-graph optimization can be applied. On an industrial video-platform task with 100B sparse parameters, it raises steady-state training throughput by 16.5% over the native baseline. On the DLRM benchmark, it achieves 1.60× and 1.26× execution speedups over DLRM-Baseline and TorchRec.
In briefDynamic skill selection for LLM agents via long-horizon state tracking
SummaryThis paper addresses how language model agents should choose skills during long-horizon tasks, where needed procedures emerge as execution proceeds. It proposes just-in-time skill retrieval, selecting skills dynamically rather than fixing them before execution. The authors introduce JitSkill, a framework that continuously tracks task state to link new requirements with accumulated progress. Experiments on three agent benchmarks are reported to show improved long-horizon task completion.
In briefMemHarmBench: Testing whether AI agents with memory can be secretly manipulated across sessions
SummaryThis paper introduces MemHarmBench, a benchmark for testing cross-session prompt injection in AI agents that keep persistent memory. The method plants malicious instructions, checks whether they are retained in memory, removes the attack content, and then runs a new legitimate task while the memory is kept. Evaluation covers whether the agent follows the attacker's intent, completes the task, and shows observable anomalies. The authors report that such attacks can stay behaviorally stealthy, with agents completing tasks normally while still making critical decisions aligned with the attacker's intent.
In briefTraining-free high-resolution image generation that restores semantic fidelity in diffusion transformers
SummaryDiffusion Transformer text-to-image models often lose meaningful content when generating images beyond their training resolution. This paper proposes a training-free framework with two parts: a text-visual scale-decoupled attention module that keeps text tokens' global receptive field while aligning image tokens to the training resolution, and a structure-detail guidance strategy that models structure and fine detail across scales and time steps. The authors report that the method improves structural coherence and detail fidelity over existing training-free approaches.
In briefSPAR: Experience-Driven Strategy Evolution in Multi-Session B2B Sales
SummaryThis paper presents SPAR, a framework that treats sales strategies as testable hypotheses and refines them without updating model parameters. It uses a double-loop design: an inner loop adapts to individual merchants through episodic memory, while an outer loop evolves a shared strategy playbook via causal attribution. In a live A/B test on the 1688 platform with over 200,000 merchants, SPAR raised conversion by 15.8% and high-intent merchant growth by 35.8%. Domain expert evaluation rated its overall sales quality at 4.62 out of 5, the highest among compared methods.
In briefDirect covariance-to-beamformer mapping for low-latency ISAC beamforming with later DoA estimation
SummaryThis paper addresses the latency of two-stage ISAC beamforming, where target directions are estimated first and beamforming weights are then computed by solving non-convex optimization. It proposes BFDNN, a compact regression network that maps covariance-matrix features directly to hybrid beamforming weights, removing the DoA estimation stage during inference. On the evaluated platform, it achieves beam patterns comparable to MUSIC and DNN-based pipelines with higher effective spectral efficiency, while reducing prediction latency by 25 to 30 times. Target directions can also be inferred indirectly from the generated radiation pattern, with competitive DoA accuracy.
In briefPhased Fine-Tuning of Multimodal LLMs for WebUI Code Across Multiple Frameworks
SummaryThis paper studies how to fine-tune multimodal large language models so they can turn WebUI screenshots into code for several frameworks (React, Vue, Angular) without the compilation errors caused by mixing their syntax. The method normalizes code to keep each framework's native syntax, estimates compatibility and heterogeneity between frameworks offline, and orders training groups in phases. It then applies phase-wise adapter tuning to reduce interference between frameworks. Experiments on multiple benchmarks report better generation quality and compilation success rates than baselines.
In briefSEALECT: Plausible Maritime Vessel Retrieval Using Scene-Based Pose Priors
SummaryThis paper addresses retrieving vessels for compositing into maritime images, where geometric constraints make appearance-only retrieval produce physically implausible insertions. The authors propose SEALECT, which combines visual similarity, vessel pose, scale, and scene geometry into one score using a nonparametric kernel for pose and scale priors. Across seven maritime benchmarks, its trained retriever improves recall over the SigLIP 2 encoder, and the training-free prior further improves geometric plausibility. In a blind study of end-to-end composites, SEALECT ranked first among six methods.
In briefGaitFM: A cross-modal AI model that assesses walking patterns using wearable sensors.
SummaryThis paper presents GaitFM, a foundation model for gait assessment using inertial measurement unit (IMU) sensors. It transfers knowledge from large vision models and is pretrained on unlabeled IMU data from about 4,000 participants, comprising 1.23 million gait windows. The model performs strongly across 18 gait-related tasks and generalizes across different sensor configurations and body locations.
