In briefKNOU-Flood: probabilistic machine learning for global river flow forecasts at multiple lead times
SummaryThe paper addresses global streamflow forecasting, where models trained on reanalysis lose accuracy when driven by operational forecasts, especially at longer lead times. Keeping a single-layer LSTM and two-stage training, the authors change the inputs with a longer window, lead-aware forecast inputs, GraphCast as a second source, and a latent-noise ensemble trained with fair CRPS. Compared with a re-implementation of the AIFL recipe under identical data and metrics, KNOU-Flood reports higher KGE' at most lead times and lower CRPS across 2,099 test basins.
In briefForestQuery: A Framework for Unified Forest Point Cloud Segmentation That Handles Tree Boundaries and Spatial Structure
SummaryThis paper presents ForestQuery, a framework for segmenting forest 3D point clouds into both semantic classes and individual instances. It explicitly models boundary uncertainty to build more reliable instance queries and adjusts optimization through adaptive loss reweighting. It also adds spatially anchored semantic queries that use learnable 3D anchors encoding forest vertical stratification. The authors report consistent gains in individual-tree and semantic segmentation across public benchmarks and a self-collected real-world dataset.
In briefRunning neural networks at 16-bit precision on microcontrollers while keeping 8-bit speed and energy use
SummaryThis paper proposes W16A16, a 16-bit quantization method for running neural networks on microcontrollers with high accuracy, speed, and low energy use. The authors analyze the Armv7E-M architecture to explain why 16-bit operations perform well there. They measure quantization error on regression and classification tasks, along with inference time and energy on real microcontroller deployments. The reported result is about 10 times lower quantization error than 8-bit schemes, with similar or better inference time and energy.
In briefBeyond Entropy: Self-Diagnostic Token Optimization for Video Reasoning
SummaryThis paper addresses unclear token-level credit assignment in reinforcement learning for multimodal video reasoning. It proposes DyCPO, which uses a multi-role dependence metric to balance visual exploration with answer-relevant cues and suppresses exploration-only filler tokens and spurious visual noise. Instead of static counterfactual priors, DyCPO derives counterfactual signals dynamically from the model's own successful and failed rollouts. The authors report consistent performance gains on complex video reasoning and general video understanding benchmarks.
In briefA unified Bayesian data assimilation framework using generative models and observation interpolants
SummaryThis paper introduces an observation-interpolant framework that turns pretrained stochastic interpolant, flow matching, and diffusion models into posterior samplers without retraining. Conditioning the interpolant path on observations gives a likelihood-score correction that unifies stochastic and deterministic posterior sampling. A closed-form Gaussian surrogate approximates the score for practical computation. The method is evaluated on linear-Gaussian dynamics, stochastic 2D Navier-Stokes, and urban airflow with up to about 10^4 degrees of freedom.
In briefBidirectional Voronoi-biased curriculum for reinforcement learning: expanding training from both start and goal ends
SummaryThe paper addresses exploration bottlenecks in sparse-reward, long-horizon goal-conditioned reinforcement learning. It proposes BVER, which grows start states from the goal and goals from the initial state distribution simultaneously, biases them toward unexplored space, and trains a single policy on both. On point-mass mazes, quadrupedal box climbing, and robot-arm ring-on-peg transfer, BVER learns faster than all compared reference-free curricula, and is the only one to learn to climb a 0.7 m box.
Juri Pfammatter, Kaixian Qu, Clemens Schwarke et al.
In briefEdgeAgent: On-device LLM inference system for multi-agent workloads on CPU-GPU unified memory
SummaryThis paper presents EdgeAgent, an inference system co-designed for edge devices with unified memory and for multi-agent LLM workflows. It uses zero-copy, unified-memory-aware tensor parallelism and dynamically allocates draft budgets based on sequence predictability. An asynchronous suspend-and-yield mechanism evicts stalled agents during tool calls. On an Apple M4 SoC, the unified-memory-aware execution alone gives a 1.29x speedup over batched speculative decoding, and the full system reaches 1.77x under extreme tool-use latencies.
Yuhai Long (1), Yuanxin Wei (1), Kai Wu (2) et al.
In briefNAVA-WAM: Learning Action Priors Directly from Action-Free Videos for World Action Models
SummaryThe paper proposes NAVA-WAM, which pretrains a robot action policy directly on observation-only videos, reducing reliance on action-annotated robot trajectories. Training has two stages: first, future-video flow-matching supervision learns action-relevant priors, then action-labeled demonstrations post-train the policy through joint video-action flow matching. The authors report that it consistently outperforms prior approaches in both in-distribution and out-of-distribution settings, with strong action-label efficiency and real-robot generalization.
In briefDriftTTS: Distillation-free few-step speech synthesis using distribution-matching drift
SummaryThis paper presents DriftTTS, a few-step mel-spectrogram generator trained without a generative teacher, distillation, or adversarial discrimination. It uses a distribution-matching drift objective in a feature space from raw mels and a frozen masked-autoencoder encoder trained on LJSpeech. At NFE=4 on LJSpeech, it reports 3.87 dB MCD and 3.7% WER, versus 3.85 dB and 3.4% for Matcha-TTS, and a 4.18 MOS in blind listening tests compared with 3.96 for Matcha-TTS and 4.22 for ground truth.
Mohammad Nur Hossain Khan, Subrata Biswas, Bashima Islam
In briefCutting Visual Computation in Vision-Language Models Without Retraining
SummaryVision-language models spend heavy inference cost processing every visual token at every decoder layer. This paper introduces Patch-to-Prune (P2P), a training-free method that replaces visual-token projection outputs in selected layers with fixed neutral activation vectors, keeping sequence length and weights unchanged. On four Qwen2.5-VL and LLaVA models across seven benchmarks, P2P at a 3% tolerance keeps about 94% of dense accuracy while cutting FLOPs by 55%. The layer-wise analysis suggests visual integration concentrates in intermediate layers.
Rahul Chowdhury, Timothy A Rupprecht, Xuan Shen et al.
In briefBenchmark for how rejection policies transfer when candidate sets and tasks change
SummaryThis paper compares Laya, Jev and Qwen2.5-7B-Instruct on public reference labels to test whether rejection policies stay useful when candidate sets and tasks change. Policies calibrated on a source dataset often fail to keep the target operating point; a Jev policy calibrated on DBpedia rejects 69.3% of covered Emotion test inputs. The benchmark provides shared inputs, explicit decision and failure categories, and reproducible scoring.
In briefCVE2AP: LLM-based automatic generation of PDDL-encoded attack paths from vulnerability descriptions
SummaryThis paper presents CVE2AP, an approach that uses large language models to automatically turn natural language CVE vulnerability descriptions into attack paths encoded in PDDL, a formal planning language. It uses structured prompting and an error-feedback loop that refines outputs based on planner-reported syntax and solvability errors. Across multiple LLMs, the authors report up to 86.9% syntax correctness, 78.6% solvability, and 93.1% semantic correctness under LLM-as-expert evaluation. They also state that GPT-5.5 gave the best quality-cost trade-off.
In briefExplainable video deepfake detection using explicit forensic features and temporal modeling
SummaryThis paper presents an interpretable framework for detecting deepfakes in videos. It converts videos into identity-consistent facial trajectories split into fixed-length windows, and represents each frame with 68 descriptors covering photometric, textural, geometric, and compression features. An LSTM network then models temporal irregularities. The method reports F1-scores of 98.0%, 91.0%, 97.6%, and 96.2% on four benchmark datasets and shows good cross-dataset generalization.
In briefInitialization method for Fourier feature PINNs that tailors weights to the PDE being solved
SummaryThis paper studies spectral bias in Physics-Informed Neural Networks, where some frequencies of the target function converge more slowly. Working in the Neural Tangent Kernel regime, the authors derive an evolution equation for frequency-domain residual error. They show that convergence speed for each frequency depends mainly on the product of the differential operator's symbol and the spectral density of the initialization weights. Based on this, they propose an initialization strategy tailored to the specific PDE. Experiments on linear and nonlinear PDEs show improved learning dynamics and accuracy across frequencies compared with standard initialization, at no extra training cost.
Juan Molina, Paris Perdikaris, Mircea Petrache et al.
In briefEVEWorld: Physical Evolution Supervision for Embodied World Models
SummaryThe paper argues that existing embodied world models suffer from "Model Laziness," prioritizing visual fidelity over physical reasoning and lacking supervision of object dynamics over time. It proposes EVEWorld, built from Instance-Guided Restoration (IGR) and Temporal Instance Alignment (TIA) to improve instance and cross-frame consistency. The authors also introduce the Model Laziness Rate (MLR) metric, and report experiments on DreamGenBench, EWMBench, and PBench showing an 87.5% MLR reduction versus GigaWorld-0.
In briefSCAD: A Method for Training Long-Horizon AI Agents
SummaryThis paper introduces SCAD, a training method for agents that carry out long, multi-step tasks. It splits interactions into planning and bounded subtask execution, distills execution within local contexts, and refines planning credit using cross-rollout subtask prefix trees. On text tasks, SCAD improves macro-average accuracy over the strongest training baseline by 4.48 percentage points.
In briefLAS-CLIP: A Lightweight Adapter Steering Method for CLIP's Visual Encoder
SummaryCLIP's visual encoder outputs only global image features, which limits region-level tasks. LAS-CLIP keeps all CLIP parameters frozen and uses a compact MaskAdapter that turns an input mask into attention biases steering the frozen self-attention layers toward the target region. Without a mask, the model behaves like vanilla CLIP. It reports results competitive with or better than Alpha-CLIP on ImageNet-S zero-shot classification and RefCOCO referring expression comprehension, though Alpha-CLIP fine-tunes its full encoder on millions of samples.
Anh-Khoa Dinh-Duc, Duc-Tai Dinh, Tam V. Nguyen et al.
In briefMultilingual GSM-Symbolic: Factors That Determine How Abilities Transfer Between Languages
SummaryThe paper introduces Multilingual GSM-Symbolic, a math dataset with 30,000 item-matched question-answer pairs across 15 languages, built with symbolic templates to reduce overfitting. It jointly estimates how model size, language resource level, reasoning, and typological distance affect performance. The authors find that model size and reasoning narrow the gap between low- and high-resource languages but have little effect on typologically distant languages.
Kenneth Enevoldsen, Riley Herchert, Sofie Mosegaard et al.
In briefGeometry Meets Physics: Data-Efficient Pre-Training for Neural PDE Solvers on Complex 3D Shapes
SummaryThis paper addresses neural surrogate models for partial differential equations on unstructured 3D geometries, which often generalize poorly and need costly training data. It proposes a pre-training framework that needs no stored disk data: a geometry-driven method using intrinsic shape descriptors for steady-state problems, and a physics-driven method that generates synthetic PDE data online for transient problems. Experiments report faster convergence, better data efficiency, and higher fine-tuning accuracy, especially in low-data settings.
Luis Medrano-Navarro, Giacomo Baldan, Qiang Liu et al.
In briefFollow the Winners (FTW): Conservative Policy Improvement with the Cross-Entropy Method for Critic-Free Fine-Tuning
SummaryThis paper proposes Follow the Winners (FTW), a critic-free policy learning algorithm that adapts the cross-entropy method to reinforcement fine-tuning of agentic language models. It replaces GRPO-style repeated group rollouts with an ordinal filter on replay-buffer samples. Derived through a control-as-inference view, the method identifies GRPO as risk-neutral and shows that DPO and FTW share a bounded risk-seeking offset that FTW controls. Scaled to agentic LLM post-training, FTW is reported to match GRPO and PPO on Sokoban and Search-R1 baselines.
Joery Ari\"en de Vries, Neil David Lawrence, Zhenwen Dai
In briefReFract: A Benchmark Testing Whether Language Model Agents Act Appropriately for the User's Role
SummaryThis paper introduces ReFract, a benchmark of 150 expert-validated entries that tests whether language model agents act appropriately for a user's role when given the same query. The entries are built from anonymized domain support conversations and use Text World Models to simulate the agent's operating environment. State-of-the-art LLMs solve at most 69% of tasks, and more than 50% of their trajectories include attempts at perspective-violating actions.
Hainiu Xu, V\'{i}tor N. Louren\c{c}o, Mohnish Dubey et al.
In briefVISTA: An equivariant visual-tactile diffusion policy for contact-rich robot manipulation
SummaryThis paper proposes VISTA, a workspace-level equivariant visuotactile diffusion policy for imitation learning in contact-rich robot manipulation. It projects visual and tactile observations into spherical tokens, fuses tactile contact cues into visual spherical directions, and rotates the fused representation by end-effector orientation to predict spatially consistent actions. According to the authors, experiments in simulation and real-world robots show VISTA substantially improves data efficiency over strong visuotactile imitation learning baselines.
Lik Hang Kenny Wong, Yiyao Ma, Xiu-Shen Wei et al.
In briefA fully automatic method for separating individual 3D dendrites in electron microscope brain images
SummaryThis paper presents a fully automatic pipeline that reconstructs individual 3D dendrites from SBF-SEM images without manual prompting at inference. It combines YOLOv6-guided Segment Anything Model prompting, iterative 2D mask refinement, random forest 3D instance linking, and nnU-Net high-resolution refinement. On hippocampal CA1 data from a control rat and an epileptic rat, it reports Dice scores of 0.93 and 0.91, with instance recognition in dense epileptic regions identified as the main remaining limitation.
In briefCordial Learning: Distributed Training on Correlated Data
SummaryThis paper studies distributed learning where agents hold correlated data, so each agent's label depends on other agents' inputs for the same sample. The authors propose cordial (correlated and distributed) learning, in which agents share only low-dimensional outputs while training local models to extract useful signals from peers. For a linear model, they prove convergence with probability one to a globally optimal solution despite a nonconvex global objective. Experiments on structured multi-digit MNIST tasks show the method remains effective in highly nonlinear settings.
In briefSyntaxBench: A Statistical Framework for Testing Character-Level Reasoning in LLMs
SummaryThis paper introduces SyntaxBench, a diagnostic benchmark and statistical framework for evaluating how well large language models reason at the character level. It covers five core tasks, such as counting characters and detecting palindromes, plus a harder substring-extraction test. Evaluating eight open-weight models from 2B to 32B parameters, the authors find that tokenization affects accuracy, reasoning mode does not always help, and substring extraction remains largely unsolved, with the best four-shot exact-match accuracy at 6.75%.
In briefCompressing diffusion language models while keeping math reasoning intact
SummaryDiffusion language models are usually calibrated on clean, fully visible activations, but inference passes through partially masked states. The paper formulates a trajectory-aware low-rank objective and proposes Traj-MC, which estimates it using Monte Carlo sampling. Under matched compression budgets, trajectory-aware calibration reconstructs the generation trajectory better and preserves considerably more mathematical reasoning than clean calibration.
In briefTo Jev or Not? Evaluating Structured Decision Models for Hate-Speech Moderation
SummaryThis paper introduces HATEDECIDE, which compares six structured decision-model configurations against specialized moderation, zero-shot, commercial, and supervised baselines on four hate-speech datasets. It tests whether supplying dataset definitions or splitting them into multiple questions improves classification, and measures latency and cost. Commercial LLMs significantly outperform all decision models on only one dataset, and adding definitions or decomposition does not reliably improve results. The best hosted decision model comes within 1.6 macro-F1 points of the best commercial LLM at roughly 97% lower inference cost.
Demetris Paschalides, George Pallis, Marios D. Dikaiakos
In briefInformation limits for verifying low-rank approximation error
SummaryThis paper studies how many extra matrix-vector products are needed to verify that a low-rank approximation meets a target error, for both relative matrix error and mean-square output error. For a single candidate it gives an exact dimension-uniform minimax query constant as the failure probability vanishes. It also analyzes reusing validation responses across W nested paths, showing a sqrt(log(W+1)) dependence that is shown to be optimal under stated conditions.
WhyGives exact minimax query constants and matching lower bounds for certification reuse across paths, a clear theoretical advance with impact mainly within a narrow subfield.
In briefRefinement improves intelligibility, while search improves speaker identity in diffusion-based TTS
SummaryThis paper studies masked-diffusion text-to-speech models, comparing how model depth and inference-time refinement steps each contribute to performance. The authors trained 15 models of varying depth and evaluated zero-shot synthesis on 174 held-out speakers. Refinement closes far more of the intelligibility gap than the identity gap, while best-of-K search recovers speaker identity more effectively.
In briefMulti-task evolution with LLMs for zero-shot transfer of heuristics across optimization problems
SummaryThe paper introduces MECo, an LLM-driven multi-task evolutionary framework that transfers and recombines heuristics across combinatorial optimization problems. A compact heuristic set is selected by rewarding coverage of source combinations and is applied to target problems without further search. On 32 vehicle routing and flexible job-shop scheduling variants, MECo reports the lowest mean costs against eight automated heuristic design baselines under the same budgets.
In briefA lightweight, rubric-guided method for evaluating AI agent trajectories in production.
SummaryThis paper presents LiteTrajEval, a lightweight architecture for evaluating the step-by-step traces of LLM-based agents under a fixed budget. It builds rule profiles offline, preprocesses and flags heuristic failure signals in each trajectory, then uses a single rubric-guided LLM judge to produce structured diagnostic reports. On public Magentic-One-style and τ-bench-style datasets, it improves failure-localization agreement with human annotations over AgentRx while cutting cost and evaluation time substantially. The method has also been deployed in the authors' enterprise agentic platform.
In briefDAWIS: Data Assimilation via Multitask Interpolants and Windowed Inverse Sampling
SummaryThe paper proposes DAWIS, a unified data assimilation method that covers filtering, fixed-lag smoothing, and block smoothing. It replaces a single flow time with a multitask stochastic interpolant that assigns a separate flow time to each state in a window of consecutive states, then inverts and regenerates the window under observation guidance. Experiments on nonlinear systems with sparse, noisy observations show that DAWIS outperforms both filtering and smoothing baselines.
Erik Wikingsson, Martin Andrae, Tomas Landelius et al.
In briefSDECast: Continuous-time probabilistic weather forecasting using neural SDEs
SummaryCommon machine learning weather models predict step by step at a fixed time resolution, which can accumulate errors at short time steps. This paper introduces SDECast, a neural stochastic differential equation framework that learns stochastic atmospheric dynamics directly in physical space for continuous-time probabilistic forecasting. On a simulated geophysical flow, the authors show it recovers meaningful drift dynamics, and at hourly resolution for global forecasting it produces skillful probabilistic forecasts up to five days ahead.
Maria Marchenko, Martin Andrae, Fredrik Lindsten et al.
In briefOptimal planning when feasibility changes over time and start time is unknown
SummaryThis paper studies planning problems where feasible states or actions change over time, such as path planning among moving obstacles. The authors define any-start-time planning, where the execution start time is unknown, and introduce a compound arrival time function (cATF) that compactly encodes the optimal plan as a function of start time. Experiments on SIPP show that replanning agents often fail on difficult problems, while cATF-based algorithms can quickly look up the optimal plan once the start time is known.
Devin Wild Thomas (1), Solomon Eyal Shimony (2), Wheeler Ruml (1) et al.
In briefT3lescope: High-fidelity 3D scene mesh reconstruction from images at arbitrary resolution
SummaryThis paper presents T3lescope, a method that reconstructs 3D scene meshes from posed multi-view images without per-scene optimization, for scenes from single objects to large outdoor areas. It uses one fixed-resolution generator in an inference-time coarse-to-fine cascade, where a coarse level sets the layout and finer levels refine geometry within progressively smaller spatial cells. The authors report that it outperforms feed-forward and generative baselines on indoor, outdoor, and city-scale scenes, and recovers glossy and transparent surfaces.
In briefTraining-loss guarantees for the Muon optimizer using finite-step Newton–Schulz orthogonalization
SummaryThe paper studies whether Muon, an optimizer using five tuned Newton–Schulz orthogonalization steps, can reach a prescribed neural network training loss. For full-batch training of a sufficiently wide two-layer ReLU network, it proves that Muon reaches any target squared loss with high probability, with a finite-time hitting-time bound. Experiments at widths below the theoretical threshold report that all runs reached the target loss.
WhyProves finite-step Newton-Schulz Muon reaches target loss on wide two-layer networks, extending prior exact-orthogonalization analyses; experiments cover multiple widths and initializations.
In briefS²-PINN: A Neural Network Method for Random Equations with Uncertainty
SummaryThis paper introduces S²-PINN, a physics-informed neural network for solving partial differential equations with random inputs and quantifying their uncertainty. It represents the solution as a separable combination of a learnable spatial dictionary, temporal features, and a generalized polynomial chaos basis, linked by a low-rank tensor decomposition. The authors report that the method outperforms nine baselines on four manufactured benchmarks using fewer parameters, and also evaluate it on Poisson, Darcy, Navier–Stokes, and inverse problems.
In briefJOVE: Jointly Executing and Verifying LLM Task Graphs Under Resource Limits
SummaryJOVE is an online framework that splits complex reasoning queries into task graphs and assigns each subtask to a suitable language model, while selectively sending some intermediate outputs for paid verification. Verification feedback is used to update estimates of each model's quality over time, and each query is handled by solving a mixed-integer linear program under a long-term budget and per-query latency limit. The authors prove sublinear regret for quality learning under stated assumptions. On four reasoning benchmarks, JOVE reports competitive accuracy with standard baselines while cutting average cost and latency by at least 3.17 times.
In briefWrong Organ, Right Physics: Echo Pretraining for Lung Ultrasound TB Screening
SummaryThis paper studies whether an encoder pretrained on echocardiography, which shares ultrasound physics with lung ultrasound, helps detect tuberculosis when labelled lung data is scarce. It compares seventeen encoders and finds that the choice of pretraining corpus, including echocardiography versus generic video, does not reliably change results. Standardising features between encoder and classifier improved all encoders by a mean of 1.23 points. The authors conclude the limiting factor is the size of the labelled cohort, not the encoder, and note video models were given replicated still images.
Christiaan M. Geldenhuys, Joshua M. Jansen van V\"uren, V\'eronique Suttels et al.
In briefHow vision and text fuse inside two types of multimodal AI models
SummaryThis paper examines how visual and text information are combined layer by layer in two kinds of multimodal large language models: concatenation-based and native multimodal designs. The authors use alignment decoupling, attention routing and entropy, and intrinsic dimensionality analyses, supported by causal intervention experiments. They report two distinct fusion pathways: concatenation models process text first and vision later, while native models show earlier visual-text co-adaptation and reorganization of feature space.
In briefHexVIO: Enabling All-Day Stereo-Inertial Tracking on Commodity DSPs
SummaryThis paper presents HexVIO, which moves the visual frontend of a stereo-inertial odometry system onto the Hexagon DSP found in many smartphones and XR devices, while the backend stays on the main CPU. The authors report a 67% power reduction or an 86% throughput increase compared with CPU-only execution on a commodity smartphone. They state the system sustains real-time 30 fps tracking at 0.83 W, about 18 hours on the test device.
In briefNew Video Saliency Benchmark Emphasizing Motion
SummaryThe paper argues that the popular LEDOV video saliency dataset can be largely explained by static image baselines, so it fails to test temporal modeling well. The authors introduce SalTempto, a benchmark of 224 highly dynamic clips with gaze data from up to 16 subjects and a training split. On SalTempto, the static baseline explains only about 13% of the headroom above centerbias, while a fine-tuned temporal model performs substantially better yet still leaves nearly half unexplained.
WhyIntroduces SalTempto, a more dynamic benchmark; experiments show temporal models capture information LEDOV cannot measure, with headroom analysis.
Susmit Agrawal, Rebecca Wanner, Juliane Verwiebe et al.
In briefEVOL: A simulator-guided evolutionary method for learning path recommendation that needs no deployment-time simulator.
SummaryThis paper tackles learning path recommendation, where reinforcement learning struggles with sparse rewards and the lack of expert learning paths in educational data. The authors borrow a robotics recipe: a knowledge tracing simulator generates per-learner expert demonstrations via evolutionary search, and a deployment-free policy is trained to imitate them. Experiments on ASSIST15, Junyi, and EdNet with path lengths 5, 10, and 20 show EVOL outperforming 8 baselines, and the authors find final performance depends mainly on the quality of the evolutionary experts rather than the imitation objective.
In briefShrome at Touch'e 2026: Soft-Vote Ensembling and Counter-Causal Augmentation for Causality Extraction
SummaryThis paper addresses the Touch'e 2026 causality extraction task, which covers news sentences that look causal on the surface but deny the causal link. The authors build separate models for detection, cause-effect span extraction, and polarity labeling, averaging token scores across three RoBERTa-large taggers for extraction and adding LLM-generated counter-causal training sentences for polarity. On the CCNC test set, the system reports detection F1 of 0.869, polarity macro-F1 of 0.817, and granularity-adjusted extraction F1 of 0.728.
In briefSPEAR: A spectrally disentangled mixture-of-experts model for large-scale PDE pretraining
SummaryThis paper proposes SPEAR, a neural operator for pretraining on large-scale partial differential equations (PDEs). It splits latent features into low- and high-frequency parts, so transferable dynamics are modeled jointly while PDE-specific patterns are learned separately. It also introduces a knowledge-guided strategy that measures expert similarity to merge redundant experts. Experiments on twelve PDE datasets report that cutting the number of experts by 50% maintains or improves prediction accuracy.
In briefConsecutive posterior fusion for recovering image structures the measurements cannot observe
SummaryThe paper addresses severely ill-posed imaging inverse problems and proposes CPF-DDNM, an inference-time method that fuses consecutive measurement-aware estimates to better recover unobservable image structures with diffusion models. It requires no retraining or extra denoiser evaluations. Experiments on sparse-view and simulated low-dose CT and medical image super-resolution show consistent gains over DDNM and competitive results against other diffusion-based solvers.
Elena Morotti, Davide Evangelista, Elena Loli Piccolomini
In briefPaMIR: an open benchmark of public credit-default datasets
SummaryThe paper releases PaMIR, an open benchmark for credit-default prediction when labels are scarce and arrive late. It combines 19 public datasets with binary default labels, about 1.24M loans, firms and card accounts from nine countries, rebuilt with one leakage-audited recipe. Models are evaluated under repeated i.i.d. splits and a label-delayed stream, reporting AUC by label budget.
Mikhail Liashkov, Ilyas Varshavskiy, Shuhratjon Khalilbekov et al.
In briefBenchmarking and improving scalable nonlinear causal discovery methods for accuracy and runtime.
SummaryThis paper empirically compares four major nonlinear causal discovery approaches: differentiable structure learning, amortized structure learning, score-matching, and combinatorial search. It identifies distinct bottlenecks in each, such as an accuracy gap for differentiable and amortized methods and slow redundant scoring in combinatorial search. The authors propose SPADE, a spline-based scoring scheme that reuses sufficient statistics, reducing the Gaussian variant's complexity from O(nd^3) to O(nd^2+d^3) under bounded indegree. The paper reports that SPADE solves 100-variable problems in seconds and 1600-variable problems in minutes while retaining high structural accuracy.
In briefCross-cohort tuberculosis classification using clinical data from Uganda and South Africa
SummaryThis study applies machine learning to clinical and demographic data from patients in Uganda and South Africa, aiming to identify who needs costly molecular TB testing. It compares logistic regression, multilayer perceptrons, and convolutional neural networks with greedy feature selection. Logistic regression performed most consistently on held-out cohorts, narrowly missing WHO minimum sensitivity requirements, and the authors conclude neural-network TB screening appears viable.
Joshua M. Jansen van V\"uren, Devendra S. Parihar, Daphne Naidoo et al.
In briefCOSMI: Composing Single-Object Clips into Multi-Object Human-Object Interaction Data and Models
SummaryThe paper addresses the scarcity of multi-object human-object interaction data by composing contact-consistent single-object clips into multi-object sequences, filtered by a language model and geometric checks. This yields the COSMI dataset, with 222k sequences and 275 hours covering up to five objects. The authors also train COSMI, a text-to-interaction diffusion transformer, which outperforms baselines in text alignment and contact accuracy on a benchmark with unseen objects and interaction combinations.
In briefLearning a Fact Isn't Learning to Retrieve It: How models recall facts depends on the question forms they saw in training.
SummaryThis paper examines whether a language model that learns a fact can retrieve it under different request forms. In a two-stage setup, one model saw facts in five request forms during the first stage while another saw only statements, and both then learned new facts as statements. Both retrieved new facts similarly from statements but differed sharply on other forms. The authors also show that the model's context state before the answer affects retrieval and that changing it during training alters later retrieval.
In briefEmbPASS: A Cross-Embodiment Benchmark and Network for Panoramic Segmentation Across Different Robot and Vehicle Platforms
SummaryThe paper introduces Cross-Embodiment Open Panoramic Segmentation, a task addressing how panoramic perception varies across platforms. It presents EmbPASS, a benchmark covering vehicle, drone, wearable, and quadruped platforms under one semantic taxonomy. The authors also propose EPONet, which uses RAMA and CAST modules, and report 35.82% mIoU on EmbPASS, 1.10% above the strongest baseline.
In briefCollective Bias Mitigation via Model Routing and Collaboration
SummaryThis paper presents Collective Bias Mitigation (CBM), a framework that reduces bias in large language models by learning fine-grained model behavior and sharing knowledge among diverse LLMs. In the top-7 setting, the Committee topology lowers the age bias score from 0.25 to 0.10, substantially outperforming standalone baselines. The Debating and Committee topologies both achieve significant bias reduction, with the Committee approach offering a balance between mitigation effectiveness and inference cost.
In briefWAMpy: Fast Python Tool for Generating Prolog Programs
SummaryThis paper presents WAMpy, a Python framework built for synthesizing Prolog programs. It compiles Prolog clauses into NumPy array-based WAM instructions and supports partial recompilation of hypotheses against fixed background knowledge. Numba just-in-time compilation accelerates performance-critical routines. In a benchmark of repeated compilation-and-evaluation workloads, WAMpy improved end-to-end performance over SWI-Prolog accessed from Python via Janus.
In briefD2K-Bench: Can AI Agents Turn Expert GPU Designs into Fast Code?
SummaryThis paper introduces D2K-Bench, a diagnostic benchmark of 26 tasks and 85 workloads that tests how well LLM agents turn expert design guidance into efficient GPU kernels. The guidance spans high-level algorithmic insights, dataflow design, and low-level optimization tricks. On NVIDIA B200 GPUs across five models, guidance raised correctness from 93.1% to 98.5% and the Performance Score from 1.46 to 1.95. The authors also identify design properties that generated code still fails to implement.
Daifeng Li, Huiqiang Jiang, Chengruidong Zhang et al.
In briefNeuro-Physical Inverter: A Modular Framework Combining Ensemble Conditioning and Residual Learning for Magnetotelluric Inversion
SummaryThis paper presents the Neuro-Physical Inverter (NPI), a modular, uncertainty-aware framework for geophysical inversion, tested on 1D magnetotelluric (MT) data. A first stage conditions an ensemble of prior resistivity models on observed responses to produce a physically admissible reference ensemble. A second stage uses a residual-learning neural network to make targeted corrections, trained on synthetic data and fine-tuned per station. In synthetic tests NPI reduces ensemble-mean error without destabilizing the ensemble. On broadband MT data from Gabbs Valley, Nevada, it lowers the mid-period across-station mean misfit while keeping ensemble spread comparable.
In briefUsing uncertainty to gauge anatomical correctness in diffusion-based CT synthesis
SummaryThis paper studies how diffusion models can synthesise contrast-enhanced CT from non-contrast CT, reducing the need for contrast agents. Using the AortaDiff framework, the authors compare six uncertainty methods at pixel, region and image levels, including on external multi-centre data and out-of-distribution cases. MCDropout performs strongly across scales and can be used on models already trained with dropout. Uncertainty reliably flags severe failures but discriminates poorly among high-quality images.
Yuxuan Ou, Konstantinos Kamnitsas, OxAAA Study et al.
In briefAdaStep: Adaptive step credit weighting to better reward individual decisions in AI agent training
SummaryLong-horizon LLM agents are usually trained with sparse outcome rewards, which makes it hard to tell how much each decision contributed. The paper proposes AdaStep, which adjusts how strongly each step's local advantage affects the overall trajectory signal based on how reliable that estimate is. It needs only lightweight scalar computation, with no critic or extra rollouts. Experiments on ALFWorld, WebShop, and ScienceWorld with three model backbones report consistent gains over baselines at low computational cost.
In briefNear-optimal convex optimization using lazily queried second-order information
SummaryThis paper studies convex optimization where gradients are queried every iteration but Hessians only once every m iterations. The authors prove a lower bound of Ω(m+m^{1/7}ε^{-2/7}) on total iterations needed to reach an ε-solution. They propose a new method with an upper bound of Õ(m+m^{1/7}ε^{-2/7}), improving on the prior Õ(m+m^{13/21}ε^{-2/7}) and matching the lower bound up to logarithmic factors.
In briefVDOT++: Faster Video Generation via Unbalanced Optimal Transport Distillation
SummaryVideo diffusion models are costly because sampling repeatedly runs large backbones. This paper presents VDOT++, a unified distillation framework applied separately to text-to-video, image-to-video, and condition-based generators. It uses an asymmetric unbalanced optimal transport formulation and an ℓ1 ground cost with weighted-median aggregation to handle output diversity. Experiments on several benchmarks report that four-step generators are competitive with many-step teachers and strong few-step baselines.
In briefSame Answer, New Box: Is RLVR's Out-of-Domain Gain Just Answer Format?
SummaryThis paper examines whether gains on out-of-domain multiple-choice benchmarks after reinforcement learning with verifiable rewards come from better answers or from answer formatting that answer extractors detect. Re-scoring outputs with a new extractor that accepts answers in \boxed{} or in text, the authors find that 82% of the reported gain on four 7B model pairs disappears. On a 32B pair, almost none of the reported gain is attributable to format.
WhyShows most RLVR out-of-domain gains stem from answer format, not correctness, verified across multiple extractors and LLM readers; 82% vanishes.
In briefVisionMX: Post-Training Microscaling Quantization for Vision Models
SummaryThis paper systematically studies post-training quantization of vision models using microscaling (MX) formats, which pair low-precision elements with shared block scales. The authors identify three error sources in direct conversion and propose VisionMX, which optimizes bounded weight rounding and applies a foldable affine correction to activations. Evaluated on image classification, object detection, semantic segmentation, and low-light image enhancement, VisionMX outperforms direct conversion and the compared baselines, with the largest gains in architectures most sensitive to MX conversion.
Elad Dror Cohen, Ofir Gordon, Lior Dikstein et al.
In briefKernel SVD Extended to Multiple Data Sources
SummaryThis paper extends Kernel Singular Value Decomposition (KSVD) to multiple data sources, proposing eKSVD for joint nonlinear feature learning on asymmetric kernels. The primal formulation learns projections for each source jointly with pairwise couplings, and a dual formulation generalizes KSVD's shifted eigenvalue problem. A covariance-based framework uses neural networks as explicit feature mappings, and experiments show eKSVD outperforms Mercer-kernel-based methods on multiple data sources.
In briefStanceEval 2026: The Second Stance Detection Shared Task
SummaryStanceEval 2026 is the second edition of a shared task on stance detection in Arabic social media text. Given a tweet and a target, systems must classify the writer's stance as Favor, Against, or None. The task has two tracks: thematically related cross-target transfer and cross-domain transfer to unseen targets. Top systems scored well above the strongest baselines on both tracks.
In briefLLM-enhanced grid-level bike-sharing demand forecasting
SummaryThis paper proposes a short-term forecasting framework for docked bike-sharing demand that uses a large language model to turn urban text, such as event schedules and local news, into spatial-temporal perturbation fields. These fields feed a zero-inflated adaptive spatio-temporal graph convolutional network. Experiments on the Barcelona Bicing dataset report better performance than established neural baselines, with particularly strong gains during high-demand periods.
In briefEvolutionary algorithms that sample rare but useful outputs without retraining models
SummaryThis paper studies generative-model discoveries that are both useful (high reward) and surprising (low probability). The authors interpret various evolutionary algorithms as approximate Markov Chain Monte Carlo, a sampling method that requires no optimization. Based on this view, they propose Distribution Matching Evolutionary Algorithms (DME), which sample from a global target distribution without updating model weights. The paper reports that DME has higher sample efficiency than existing methods on problems that require many samples to find a solution.
In briefSmall language models for checking whether AI agent tool calls match task intent
SummaryThe paper studies using small language models (SLMs) to check, for every tool call an AI agent makes, whether the selected tool is relevant to the assigned task. Standard authorization can only say whether a call is allowed, not whether it serves the task's intent. The authors built a novel dataset of multi-tool tasks whose tools span distinct MCP servers, and optimized SLMs using prompt optimization, supervised fine-tuning, and GRPO reinforcement learning.
Chiara Troiani, Arash Salarian, Majed El Helou et al.
In briefHyperFuse: Fast self-supervised node embeddings for attributed hypergraphs
SummaryExisting self-supervised hypergraph methods often need deep encoders trained for hundreds of epochs, which makes generating node embeddings costly. HyperFuse is a label-free pipeline that computes structural node coordinates via a spectral relaxation of hypergraph modularity, assigns bounded utility weights to hyperedges based on member stability, and trains a lightweight encoder for 100 epochs. On nine public hypergraphs compared with several baselines, it averaged 8.7 seconds per dataset on eight datasets and achieved 13 to 179 times geometric-mean speed-ups.
In briefContextual Flow Matching: a plug-in method that adaptively chooses step counts to speed up flow-based image and video generation
SummaryThis paper addresses the high inference cost of Flow Matching for visual generation by proposing COFLOW, an inference-time method that adaptively selects the number of steps for each generation based on prompt features. Its context-aware component is trained online with an unsupervised reward balancing efficiency and fidelity. The method is plug-and-play, needs no retraining of the base model, and is reported to achieve over 2.5x speedup on image and video generation while preserving perceptual and semantic quality. The authors also provide a theoretical analysis showing an O(1/K) forward-Euler discretization error bound under standard regularity conditions.
In briefPredictively Oriented Gaussian Process Posteriors: a more robust Gaussian process method under model misspecification
SummaryGaussian processes are widely used to model functional relationships and quantify uncertainty, but they require design choices such as the kernel and observation model, and poor choices can misspecify the model. This paper introduces Predictively Oriented Gaussian Processes (PrO-GPs), which treat predictive uncertainty as the primary inferential target. The authors derive a reduced formulation and a practical sampling scheme for efficient computation. In synthetic and real-data experiments, PrO-GPs yield better calibrated predictive distributions under model misspecification than standard GP approaches.
In briefWhere Jagged Competence Comes From: Uneven Learning of Lower-Layer Rules in a Two-Layer Task
SummaryThis paper studies why capable systems can have low average error yet fail on particular inputs. The authors build a two-layer task in which a lower layer computes five per-slot sums and an upper layer predicts the next category using the slot-1 sum and a carried-over mode. Small recurrent networks trained on the upper task acquired the slot-1 information well (97-98% linear-probe readability) but poorly for other slots (5-10%), with category errors concentrated near slot-1 cutoffs. The authors conclude that uneven acquisition, access, and preservation of the lower-layer theory partly explains this jagged competence, though some failures remain unexplained and generalization to larger systems is untested.
In briefPredicting and repairing collapse when merging language models
SummaryThe paper studies merging fine-tuned large language models by averaging their task vectors, where some merges collapse far below the base model. The authors propose a pre-merge score based on the variance of task vectors across specialists as a measure of interference, which predicts collapse. They also introduce PRISM, which averages first and then soft-thresholds each layer according to its interference, keeping destructive merges near base performance without data or tuning.
WhyPredicts merge collapse via task-vector variance and proposes a repair operator, validated on 22 configurations; novel but modest scale.
In briefKV²: A Self-Refining KV Cache Compression Method
SummaryThis paper proposes KV², a query-agnostic compression method for the key-value cache, aimed at reusing one prefilled long context across many queries. It first uses a lightweight proxy scorer to select informative tokens, then reprocesses only that subset to compute final eviction scores. Experiments on RULER, Needle-in-a-Haystack, and LongBench report that its advantage over baselines grows as the KV-cache budget tightens.
WhySelective reconstruction with a lightweight proxy yields query-agnostic KV compression; moderate novelty. Experiments on RULER and LongBench across tight budgets show clear margins over baselines and
In briefHow LLM agents favor items from certain sources, and ways to reduce it
SummaryThis paper studies how LLM agents, which may choose products, hotels, or citations on a user's behalf, prefer items from particular sources such as websites or services. Testing 12 agent models across three domains, the authors find that every model favors some sources and avoids others, and this preference can outweigh how well an item meets the request. The paper examines two routes that produce the preference and reports that supplying missing information or a prompt countering preconceptions reduces it.
In briefBridging Research and Practice: Testing general and skin-specific AI models for classifying skin lesions in clinics
SummaryThis paper benchmarks a range of AI models on recent dermatology datasets that include dermoscopic images and smartphone clinical photos. It compares general-purpose and medical vision-language models and foundation models against dermatology-specific classifiers, such as embedding-based methods and convolutional neural networks. The study evaluates performance under distribution shifts, changes in imaging modality, and demographic variability. It aims to show how far current models are from what clinical deployment requires.
Emanoel dos Santos, Kelvin Cunha, Rodrigo Mota et al.
In briefPocketSplat: Budgeted Gaussian 3D reconstruction for mobile devices
SummaryThis paper presents PocketSplat, a feed-forward method that builds compact Gaussian 3D assets for mobile use under a set output budget. It arranges geometry-aware latent candidates in predicted world space, allocates capacity across local cells, and decodes full Gaussian attributes only for retained candidates. Experiments on DL3DV and out-of-distribution benchmarks show a strong quality-budget trade-off, and on Mip-NeRF 360 it runs directly on an iPhone.
In briefClosing Cases Only When the Evidence Supports It: Training LLM Investigators
SummaryThe paper studies when LLM investigators should close an accident, defect or outage case, judging whether gathered evidence is enough for a conclusion. The authors build Nautil, a set of 731 audited cases, and evaluate closure with three tests: closure accuracy, evidence dependence and conclusion and gap quality. Fine-tuning and reinforcement learning on a 9B model make its closure decisions follow the evidence more closely.
In briefLightweight on-device perception for robotic guide dogs
SummaryThis paper presents on-device perception modules that fuse a 360 camera and 2D LiDAR for collision avoidance, plus moving-object detection and tracking for human-centric guidance. When walking is impossible, a vision-language model provides pathway explanations to reduce user anxiety. The system ran in real time under 55 W and reached 83.8% accuracy on the GuideDogQA benchmark, versus 67.1% for GPT-4o.
In briefMemory-state critic for asymmetric actor-critic, tested on vision-based drone pursuit-evasion
SummaryThis paper studies asymmetric actor-critic learning in partially observable decision problems, where the critic can access the true state during training. The authors propose a memory-state critic that conditions on the state and the policy's own internal memory, giving unbiased policy gradients without a separate recurrent history encoder. In a vision-based pursuit-evasion task between two quadrotors, it outperforms a history-state critic and converges faster, with a slight edge in a wall arena.
Arthur Louette, Alejandro S\'anchez Roncero, Gaspard Lambrechts et al.
In briefOn-policy distillation gains and collapse: a reinforcement learning view
SummaryThis paper examines why on-policy distillation (OPD), a method for post-training language models, can improve performance yet also collapse into overly long, repetitive generation. The authors explain this through reinforcement learning, arguing that the teacher implicitly rewards student behaviors even when it rarely produces them itself. Experiments show OPD helps when this implicit reward is reliable, but when preferences misalign with quality, reward hacking amplifies overlong, repetitive rollouts. Masking unhealthy responses during training and using SFT initialization are reported to each effectively mitigate the collapse.
In briefDistilling Reasoning from Hindsight Teachers for Rare Disease Diagnosis
SummaryThe paper fine-tunes a 1.5B student model on reasoning traces from an 8B teacher that sees the true diagnosis during generation, tested on the ZebraMap rare disease dataset. A filtered student beats the teacher by a small but statistically significant margin (p < 0.001), while the unfiltered student does not, suggesting contamination filtering drives the gain. The authors describe a 'GT hallucination' artifact, where the teacher writes 'ground truth is X' into its reasoning and the student copies it.
In briefGeometry-aligned semantic matching for aligning images across different sensor types
SummaryThis paper proposes CDPM, a method for registering planar images captured by different modalities, such as visible and infrared cameras. It first adapts DINOv3 using geometrically consistent cross-modal patch pairs so that feature similarity better reflects true spatial correspondence, then builds a DINO-centric feature pyramid with a lightweight CNN branch that adds fine structural details. Experiments on three cross-modal datasets show that CDPM performs better than the compared methods.
In briefLower sample-complexity bounds for variance-reduced policy gradient methods
SummarySome variance-reduced REINFORCE methods reach an O(ε^-3) sample complexity, but only under an unrealistic assumption on importance-weight variance. This paper proposes Defensive Policy Gradient, based on defensive importance sampling, which achieves the same rate without that assumption. The authors also establish lower bounds in a generalized black-box model, with optimal rates of Θ(ε^-4) and Θ(ε^-3) under different feedback settings.
WhyProposes defensive policy gradient without importance-weight variance assumptions and matching lower bounds in a generalized oracle model, though direct applicability to standard MDPs remains limited.
Gabor Paczolay, Matteo Papini, Alberto Maria Metelli et al.
In briefPredicting steering vectors and adapter weights for few-shot author-style transfer
SummaryThis paper studies generating scientific abstracts in an author's style from a few example abstracts per author, using large language models. It proposes three methods: contrastive activation steering, a network predicting steering vectors, and a hypernetwork predicting LoRA adapters. The authors report a trade-off between style imitation and output quality, with the hypernetwork achieving the best balance on both seen and unseen authors.
Leonard Popp, Danni Liu, Supriti Sinhamahapatra et al.
In briefBudgeted-GS: real-time large-scale Gaussian splatting via level-of-detail factoring
SummaryThis paper introduces Budgeted-GS, a post-hoc method that converts a trained 3D Gaussian Splatting model into a multi-resolution factoring tree, so one city-scale model can fit GPUs with different memory sizes. It also proposes budget-centered training, which first estimates how many primitives a scene needs and then trains at that size. The authors validate the approach on 13 public scenes and report real-time rendering at 1920x1080 on one consumer GPU, including an official city capture.
In briefOncoNoteBERT: A Language Model for Reading Real-World Outpatient Cancer Notes
SummaryThis paper develops and tests language models built for outpatient oncology notes, which contain specialised cancer terminology, staging, and treatment names. The authors used a governed UK corpus of 290,026 notes from 21,564 lung and head-and-neck cancer patients. A model trained from scratch with an oncology-specific tokenizer, OncoNoteBERT, tokenised clinical text most efficiently and performed better on masked-token probes, though a continued-pretraining variant fit the corpus best. The results suggest that tokenizer and representation design matter for oncology NLP.
Wuraola Oyewusi, Eliana Vasquez Osorio, Goran Nenadic et al.
In briefMultimodal AI reasoning to find broadly neutralizing antibodies from unlabeled human B cell repertoires
SummaryThis paper presents ImmuneAgent, a closed-loop AI system combining multimodal reasoning, continual meta-learning, and wet-lab feedback to screen antibodies from B cell receptor repertoires of vaccinated or infected people. The authors report that 60 of 110 cloned candidates neutralized virus and 12 were broadly neutralizing, with five protecting mice fully against lethal influenza challenge. The system also identified FCRL5+CD27+ atypical memory B cells as a conserved antibody source and extended to discovering antibodies against human metapneumovirus and human papillomavirus.
WhyClosed-loop AI finds bnAbs with in vivo protection versus MEDI8852, with wet-lab validation and cross-viral generalization.
In briefDoes Physics Live in the Activations? Locating Physical Quantities in Video Diffusion Models
SummaryThis paper examines whether video diffusion transformers, which generate realistic video, internally represent physical principles or only reproduce familiar motion. The authors train linear probes on the models' internal activations to read simulator-derived physical quantities covering kinematics and rigid-body dynamics. They find these quantities are decodable early in denoising and localized on object tokens, and that the probes partially generalize to unseen scene variations. The probing directions can also be used as steering vectors to change the model's output.
WhyProbing video diffusion transformers for physical quantities and steering outputs is novel, with solid evidence and notable impact on the subfield.
Jonas Kneifl, Jakub Skalski, Bart{\l}omiej Twardowski et al.
In briefTSGuard: Real-time detection and imputation of missing values in streaming time series
SummaryStreaming sensor data often suffers delays or gaps caused by faults, communication losses, or interference. This paper presents TSGuard, a real-time demo system combining a lightweight graph-aware temporal imputation model with constraint-aware validation, fallback estimation, and operator-facing explanations. It detects problematic observations, imputes missing values, and checks estimates against physical and spatial constraints, then either keeps the original value as a plausible anomaly or replaces it.
In briefCalCErt: Certifying Confidence Calibration per Bin for Medical Image Classifiers
SummaryCalCErt is a post-hoc method that certifies how well a pretrained differentiable classifier's confidence stays calibrated under adversarial perturbations, bin by bin. It combines empirical calibration estimates, statistical concentration bounds, and local Lipschitz estimates of the confidence function to bound worst-case miscalibration within an ℓ2 ball of radius R. Evaluated on 11 medical image classification tasks with multiple adversarial perturbations, it achieves substantially higher certified coverage than baselines while staying competitive in tightness.
Leo Fillioux, Stergios Christodoulidis, Stergios Christodoulidis et al.
In briefBehavior Pack Optimization: a training method that makes video AI models rely on real temporal evidence
SummaryThe paper finds that video multimodal large language models score well on video question answering even when frames are shuffled or key segments are masked, suggesting they rely on appearance and language priors. The authors propose Behavior Pack Optimization (BPO), which scores a set of outputs across counterfactual views jointly and uses the original-view response as a reference for the advantage. On TempCompass, MVBench, and NExT-QA, BPO reports gains over a budget-matched GRPO baseline in macro accuracy, temporal-hard subset accuracy, and abstention F1, with improvements also reported on Video-MME, LongVideoBench, and LLaVA-Video-7B.
In briefMaking JEPA World Models Plannable When Little of the Frame Moves
SummaryThe paper introduces SLIM, a pushing benchmark with several small objects and paired visual and language goals on identical scenes. A LeWM world model that solves PushT succeeds on under 1% of SLIM trials, and probes attribute this to an encoder whose latent barely responds to actions. Adding an inverse-dynamics auxiliary loss raises success from 0.003 to 0.35, and a small language-goal head can then plan from sentences without retraining the world model.
Florian Strohm, Patrick Wagner, Jannik Schwab et al.
In briefWhether reasoning in the same language as the query helps retrieval-augmented generation
SummaryThis paper studies whether forcing a language model to reason in the same language as the query and retrieved documents improves retrieval-augmented generation. The authors built a German-only question-answering testbed based on the tabletop game The Dark Eye. Forced German reasoning outperformed forced French, suggesting the benefit comes from language alignment rather than proficiency. However, forced German only matched, and did not exceed, the model's native English reasoning. The testbed and QA benchmark are publicly released.
Oliver Hauck, Mario Sanz-Guerrero, Katharina von der Wense
In briefPage-EntroKV: a hardware-aligned, entropy-weighted way to evict KV-cache entries in grouped-query attention models
SummaryLong-context language model serving is limited by the key-value (KV) cache, which stores past attention data. The paper proposes Page-EntroKV, which evicts cache entries at the granularity that grouped-query attention actually allocates, pooling heads within a group using entropy-based weights and operating on PagedAttention pages. On a pilot model (Qwen2.5-1.5B-Instruct, r=6), it keeps the union overhead ratio at exactly 1.000, and reaches 100% needle recall at a 20% budget versus 0% for mean pooling.
In briefSafe streaming flow planning that aligns sampling with execution dynamics
SummaryThis paper proposes SafeStreamingFlow, a goal-conditioned planner that generates long-horizon trajectories learned from demonstrations. It sequentially integrates a learned state vector field with hierarchical state prediction, so flow sampling matches the system's execution dynamics. Safety constraints are enforced only for the executed step using high-order control barrier functions. Across navigation, racing, and locomotion benchmarks, it reduces planning latency and improves safety while keeping goal-reaching success competitive.
Seunghwan Jang, Jeongyong Yang, Siddharth Ancha et al.
In briefBenchmark for Literature Retrieval in a Model Organism (Dictyostelium)
SummaryThis paper builds a literature retrieval benchmark for Dictyostelium, a model organism in cell and developmental biology, using curator-generated queries linked to PubMed articles and structured gene annotations from dictyBase. The authors study cross-encoder reranking, gene-aware query expansion, and abstract-only versus full-text retrieval. They report that reranking and gene-aware expansion help only selectively, while full-text chunks substantially improve recall and top-rank performance when abstracts omit supporting evidence.
In briefOptimizing trading strategies with LLM-guided textual gradients (TradeGrad)
SummaryThis paper proposes TradeGrad, a framework that uses LLM-generated textual gradients and accumulated optimization experience to iteratively refine quantitative trading strategies. It introduces a Cross-Period Robust Objective (CPRO) that emphasizes unfavorable historical periods to promote temporal robustness. Experiments on Chinese A-share and U.S. equity markets across cross-sectional and time-series settings report the best in-sample and out-of-sample performance. The Chinese cross-sectional strategy reportedly achieves 27.99% annualized return, 12.19% maximum drawdown, and a Sharpe ratio of 1.63.
In briefOnline calibration that keeps coverage controllable in hardware-aware neural architecture search
SummaryThe paper addresses the high cost of training every candidate in hardware-aware neural architecture search by using conformal prediction to discard architectures whose predicted reward upper bound falls below a threshold. Because reinforcement-learning search violates the exchangeability assumption, the authors replace one-shot quantile estimation with Adaptive Conformal Inference. Across three architecture families, the method tracks requested coverage levels closely while pruning 25-50% of evaluations with no measured accuracy loss.
Pedro Brandimarte, Nerea Aranjuelo, Marcos Nieto et al.
In briefParaGeo: Mapping Speech Tone Variation into a Shared Latent Space
SummaryThis paper introduces ParaGeo, a method that separates paralinguistic speech variation, such as tone and delivery, from linguistic content inside a frozen speech language model. Audio tokens are replayed with a fixed prompt, and pooled representations are projected into a shared low-dimensional space. Tests on GLM-4-Voice across 80 controls report content-held-out accuracy of 9.49% against a 1.25% baseline, with both permutation tests giving p = 0.001. The authors also report attribute-, layer-, and schedule-dependent responses to latent interventions.
In briefDoes Dynamic-Point Filtering Help in Low-Texture Scenes? A Controlled ORB-SLAM2 Study
SummaryThe authors render synthetic indoor sequences that vary surface texture and scene dynamics on a factorial grid, then compare ORB-SLAM2 with no filter, an optical-flow/epipolar filter (FLOW), and a depth-consistency filter (GEOM). Filtering helped mainly in the most dynamic scenes and did not decline monotonically with texture; at the lowest texture level it reduced tracking completeness. Contrary to the hypothesis, GEOM removed more static keypoints than FLOW.
