In briefBeyond Static Leaderboards: Capability-Focused Evaluation for Time Series Models
SummaryPublic benchmarks for time series foundation models use fixed, narrow test sets, so repeated reuse can lead to tuning for the benchmark rather than true forecasting ability, and a single aggregate score hides differences across temporal structures. The paper proposes CaFE, a refreshable evaluation extension that defines eight structural features, resamples and augments the original data to add structural diversity, and scores models on per-feature capability profiles. Experiments on GIFT-Eval and FEV-Mini20 report stable results and show per-feature strengths and weaknesses that an aggregate score conceals.
In briefRECAP: A training-free method combining sparse attention and caching to speed up video diffusion
SummaryDiffusion transformers produce high-quality videos but are computationally expensive. RECAP is a training-free method that couples sparse attention with hierarchical caching through reciprocal correction, coordinated by online error feedback. On HunyuanVideo at 720p for 5-second generation, it reports a 3.05x speedup and better fidelity and speed than direct combinations of the tested methods.
In briefMachina Pauliana: a neuro-symbolic model of the Dutch actio pauliana
SummaryThis paper presents a neuro-symbolic model of Article 42 of the Dutch Bankruptcy Act, which lets insolvency practitioners avoid transactions that prejudice creditors. The symbolic core is written in Catala to keep statute and formalisation aligned, while neural components handle facts and open-textured terms using case-based reasoning over Dutch judgments. On a gold standard of 100 annotated judgments, segmentation and labelling achieved high precision and recall. On ten annotated cases the model handled most criteria reasonably but struggled with the knowledge requirement, and retrieval depth showed no clear quantitative effect.
In briefCoRS: Faster neural job-shop scheduling by reusing interval-level graph states
SummaryThis paper presents CoRS, a neural solver for the Flexible Job-Shop Scheduling Problem (FJSP). It reuses an interval-level graph representation across several sequential decisions while updating feasibility after every action. Enhanced static features are added to offset the staleness caused by this reuse. Experiments on synthetic and public FJSP benchmarks report significantly reduced inference time with competitive solution quality.
In briefOptimizing a Document Suffix to Make Language Models Follow Context
SummaryThe paper asks whether a language model can be made to follow a user-supplied document, even when it contradicts the model's parametric knowledge, by optimizing only a document suffix through discrete token search, without written instructions or inference-time access. Tested on three instruction-tuned model families with counterfactual passages, the suffix flips nearly all items the model otherwise resists, mostly outperforming a decoding baseline and the strongest hand-written instruction, and matching activation steering. The authors report that the suffix largely works through the passage and that one suffix per model transfers to unseen passages.
Gia-Bao Dinh Ho, Sekh Mainul Islam, Masahiro Kaneko et al.
In briefHeartbeat Thinking: Preparing Voice-Agent Tool Calls While Users Are Still Speaking
SummaryVoice agents using tools often make users wait after they finish speaking. This paper proposes Heartbeat Thinking, which periodically reads the latest partial transcript and shared results of completed reasoning and tool calls during speech, so preparation continues across updates. On FDBench v3, the method increases tool coverage by 16 percentage points and reduces mean final-response latency. The results are reported across SpeechFCEval, FDBench v3, and AudioCRAG.
In briefMapping DSA transparency database categories to platform terms and conditions using AI
SummaryThis paper examines whether the high-level categories of the EU Digital Services Act Transparency Database accurately describe the rules in platforms' terms and conditions. The authors used a few-shot GPT-4.1 classifier to map 447 policy documents from 31 designated services onto the taxonomy. They found that category applicability varies sharply by service type, and that sub-category definitions overlap, making single-category assignment difficult. The authors suggest the catch-all reporting partly stems from the database's schema design and recommend service-specific schemas with more mandatory attributes.
In briefRethinking BPE: choosing how many subword merges to keep as a model selection problem
SummaryThe paper treats the number of merges in Byte-Pair Encoding (BPE), a subword tokenization method, as a statistical model selection problem. It studies information-theoretic criteria for computing optimal merge sets directly from corpus statistics. To improve merge selection when data is limited, the authors augment training corpora with synthetic text generated from the original data. Experiments across multiple corpora and downstream tasks show that principled vocabulary selection reduces model size with little performance loss, and that synthetic data can yield better tokenizers in low-resource settings.
In briefSAR2Net: Learning anchored features to align tissue slides stained differently
SummarySAR2Net is a framework for aligning whole-slide tissue images stained with different dyes, a task complicated by high resolution, tissue fragmentation, and large deformations. Instead of estimating explicit transformations, it learns spatially anchored features that describe each point by its relationship to tissue landmarks, and uses a two-stage retrieval process to find correspondences. Experiments on biopsy-oriented HE-IHC datasets report that it achieves more robust region-level alignment than previous methods under severe tissue distortion.
In briefRoAR-VAE: Making Image Latents Easier for Diffusion Models via Robust Prediction
SummaryThis paper proposes RoAR-VAE, a visual tokenizer that aims to keep images faithful while producing latents that diffusion models can learn efficiently. It introduces a principle of robust prediction, where generation-relevant information should remain recoverable from corrupted latents, implemented through RobustAlign (predicting pretrained semantic features) and RobustRecon (reconstructing the clean image). The authors report that, under the same training budget, RoAR-VAE improves average generation benchmark performance by 2.19 points over the strongest baseline and improves small-text reconstruction accuracy by 16.93 percentage points over FLUX.1-VAE.
In briefFairness or Fallacy? A review of AI bias in disease detection
SummaryThis review paper surveys AI methods for disease detection from 2015 to 2026, with a focus on identifying and mitigating bias. The authors classify solutions by input data type, model architecture, and bias type, including algorithmic, demographic, model, and deployment bias. From 64,389 articles screened across seven digital libraries, 23 were analyzed in detail. The results highlight research gaps and point to explainability, data diversity, and fairness as priorities.
Ghulam Mudassir, Hameed Ullah Khan, Ahmad Sami Al-Shamayleh et al.
In briefExMR: Refining Hand and Foot Motion for Video-Based 3D Human Capture
SummaryThis paper presents ExMR, a system that reconstructs 3D body, hand, and foot motion from monocular video. An initial temporal estimator predicts whole-body motion, and then FootMR and HandMR refiners use local evidence to predict residual joint rotations. Experiments on ARCTIC, MOYO, RICH, and MOOF report that ExMR outperforms state-of-the-art methods in hand and foot reconstruction, reducing hand per-vertex error and ankle joint angle error on MOYO by 19% and 18% relative to the strongest competitor.
In briefLawyer Role Prompts Make Small Local LLMs Refuse More Legal Requests
SummaryThis paper studies over-refusal in small open-weight LLMs deployed on-premises for legal work. Testing four models on benign legal prompts, the authors found that authority-style prefixes such as 'I am a defense lawyer' raised refusal rates by 2 to 20 times over a no-prefix baseline. The authors warn this could introduce bias through uneven access to assistance and call for further investigation.
Anastasiia Kucherenko, Francois Brouchoud, Dimitri Percia David et al.
In briefA system for forecasting laver (Pyropia) bleaching seven days ahead
SummaryThis paper presents a system that forecasts bleaching of cultured Pyropia (laver) seven days in advance for seaweed farm clusters. It combines 32 channels of public marine, weather, ocean-model and satellite data, and uses a 9.2-million-parameter spatiotemporal transformer (ST-MMT) to predict bleaching probability from a 14-day input window. The authors report detection results on a validation set and also describe weaknesses, such as sharply reduced skill when the nutrient survey data becomes older.
In briefFocused Omni-Modal GUI Agents: a paper on helping computer-use agents focus on the most relevant multimodal inputs at each step
SummaryThe paper argues that GUI agents given all streaming inputs, such as screenshots, audio, video and device states, at every step get distracted, and errors compound into trajectory drift. It proposes FocusGUI, a lightweight adapter that compresses the most relevant evidence for each step and injects it into a frozen decoder without changing its parameters. Trained only with action supervision, it reports average gains across four omni-modal backbones of +33.2 Type Match and about 3.0 times the Exact Match of the frozen backbone.
In briefTaQFormer: A Transformer for Multi-Category Coupon Discount Allocation Using Combined Data
SummaryThis paper studies how an online marketplace should allocate coupons across several service categories, choosing a discount for each while accounting for shifts in demand between categories. The authors propose TaQFormer, which treats each discount dimension as a query token and combines observational pretraining with LoRA fine-tuning on randomized controlled trial data. The paper reports that at essentially the same subsidy cost rate, the system improved business value per user over the production baseline and has been rolled out to all marketplace traffic.
WhyA treatment-as-query Transformer with observational pretraining and RCT fine-tuning, validated by a 14-day online A/B test and full production rollout.
In briefGeoEdit: Adversarial image editing to hide photo locations from AI
SummaryMultimodal reasoning models can infer where a photo was taken from subtle visual cues, which enables doxing even by non-expert adversaries. GeoEdit applies minimal edits to high-risk evidence anchors in an image to disrupt these location-reasoning chains while preserving overall scene semantics. It identifies anchors using structured knowledge and adversarially discovered geographic cues, and uses a learnable policy to refine edits. On LifePhotoBench, it protects against 6 mainstream models and exceeds visual perturbation baselines by over 0.59 in image fidelity.
In briefRobust latent-variable graphical models: exact recovery and limits
SummaryThis paper studies robust estimation of a Gaussian precision matrix that splits into a sparse part minus a low-rank part, a structure arising from latent-variable marginalization combined with conditioning on linear measurements. The authors propose a convex estimator that uses a centered Gaussian measure-transformed covariance and penalizes the low-rank part with the nuclear norm of a congruence transform. A primal-dual witness theorem establishes uniqueness, absence of false edges, sign recovery, and exact recovery of rank and inertia under stated conditions, along with information-theoretic limits and robustness under Huber contamination.
WhyGives exact recovery guarantees, information-theoretic lower bounds and near-minimax rates for sparse-plus-low-rank precision estimation.
In briefVEILED-SH: A Benchmark for Indirect Self-Harm Expressions in AI Responses
SummaryThis paper introduces VEILED-SH, a benchmark for evaluating how AI systems respond to indirect expressions of suicide, self-harm, and distress in single-turn exchanges. It contains 10,003 prompts drawn from counselor chats, forum posts, platform searches, and triage, partly grounded in Singapore-based sources and public datasets. An expert-validated multi-label taxonomy and six automated profiles calibrated to practitioners are used for proxy-alignment scoring. Testing five systems, the authors find that proxy alignment varies across the reference policies generated by different profiles, even for identical system responses.
In briefWan-Alpha++: Animating One RGBA Asset into Controllable Transparent Video Effects
SummaryThis paper presents Wan-Alpha++, a unified framework for controllable RGBA video generation, which produces videos with transparency for use in filmmaking, games, and digital content. It extends text-to-video generation to image-to-video settings using a shared RGBA latent space, with a transparency-aware diffusion objective and a transparency-guided sampler for spatial control. To address emergence-disappearance effects where boundary frames are fully transparent, the authors introduce temporal phase reparameterization, which relocates visible content to the temporal boundaries and enables appearance conditioning with limited RGBA data.
In briefCounting graphs that satisfy modal logic with counting
SummaryThis paper studies the theoretical computational complexity of counting trees and graphs that a graph neural network classifies positively. Equivalently, it counts those that satisfy a given modal logic formula with counting. The authors also study convergence.
In briefMoveBench: A Rubric-Based Benchmark for Testing AI Models in Sports and Health Coaching
SummaryThe paper argues that general benchmarks like MMLU and HELM miss sports and health requirements such as exercise safety, physiological appropriateness, and personalization, where unsafe advice can cause harm. The authors introduce MoveBench, containing 809 multiple-choice and open-ended questions paired with 5,012 rubric criteria. Evaluating 15 LLMs, they find substantial variation, especially in safety awareness, structured program design, and personalization.
In briefHP-PnP-Flow: A Faster Plug-and-Play Method for Image Restoration Using Flow Matching
SummaryThis paper addresses image restoration, recovering high-quality images from degraded observations. It proposes HP-PnP-Flow, an improved plug-and-play algorithm built on flow matching, using a different time schedule and an added pixel-space projection operator. The authors claim rigorous convergence guarantees under mild assumptions. Experiments on CelebA and AFHQ-Cat show better results than PnP-Flow on several tasks, with comparable or better quality using fewer iterations.
In briefModeling reading times with pixel-based language models that predict next words from images of text.
SummaryThis paper uses tokenization-free, pixel-based language models to estimate how predictable the next word is during reading. The predictability computed in pixel space strongly predicts gaze duration and word skipping across several languages. Its contribution is complementary to token-level surprisal and entropy, and cross-script comparisons reveal distinct visual processing biases.
In briefLearning goal-conditioned subgoals with a frozen hierarchical world model
SummaryThis paper replaces the costly search over high-level macro-action candidates in hierarchical world models with a lightweight goal-conditioned network. The network outputs a conditioning vector, and a frozen high-level predictor generates the next latent waypoint, trained on offline trajectories using hindsight goals and state-prediction losses. Reported mean success on H-JEPA's HWM checkpoints is 70.4% on Cube, 48.7% on Push-T, 100.0% on FourRoom, and 71.1% on Visual AntMaze. The authors also report 1.28 to 4.39 times end-to-end speedups for batched evaluation on an NVIDIA L40S.
In briefConditional Retrieval for Matching Equivalent Questions Across Longitudinal Surveys
SummaryThis paper tackles finding conceptually equivalent questions across different surveys, where word similarity alone is unreliable. The proposed framework first extracts each question's underlying sociological concept with a fine-tuned generative model, then uses that concept as a conditioning signal for retrieval. On the CLOSER case study, it outperforms baselines by 2-3 points in MAP and Recall. The authors also find that SFT gives the most stable gains, and that a model tuned on the ELSST thesaurus transfers well to task-specific survey data.
In briefVISAGENT: An Agent-Based End-to-End Voice Interaction System
SummaryThis paper presents VISAGENT, an agent-based voice interaction system for large-scale public services that replaces menu-driven workflows with natural speech. A two-level routing mechanism dispatches user requests to a general conversational model or specialized domain agents, and user memory supports preference-aware selection. A voice-agent registration mechanism allows task agents to be onboarded dynamically. The authors report a real-world deployment on China's 114 hotline, with experiments showing low-latency full-duplex interaction, scalable routing, and competitive speech naturalness.
In briefTriTD: A Tri-Partite Distillation Framework for Real-Time Autoregressive Video Generation
SummaryThis paper analyzes three error sources in autoregressive video diffusion generation and derives closed-form upper bounds on per-step and cumulative errors. It proposes TriTD, a distillation framework with three modules: Next-Node Trajectory Distillation, Chunk-Graded Noising, and Orthogonal-Parallel DMD loss, to jointly optimize quality and efficiency. The authors report that TriTD surpasses existing baselines in visual quality and reaches 26.8 FPS on a single H100 GPU.
In briefDynamical Batching: class-cycling training and how it shapes neural network internals
SummaryThis paper examines whether periodically cycling the class composition of training batches changes a neural network's performance and internal organization. The authors propose Dynamical Batching (DB), which emphasizes each class in turn while keeping the loss and optimizer fixed, and compare it with Uniform Batching (UB), where classes are equally represented in every batch. Where standard training fails, DB lets a capacity-limited MLP fit a three-arm spiral dataset. On CIFAR-10, the same protocol modestly improves test accuracy for a narrowed ResNet-10 and a second capacity-limited convolutional network.
Nicol\'as Ratier-Werbin, Samuel Lozano, Miguel Ruiz-Garc\'ia
In briefDeepSEA: A Multilingual Benchmark for Southeast Asian Languages
SummaryThis paper introduces DeepSEA, a multilingual evaluation benchmark covering 12 Southeast Asian languages, built by professionally translating established datasets. It spans five tasks: mathematical reasoning, scientific reasoning, question answering, abstractive summarization, and machine translation. The evaluation finds that frontier models stay robust across resource tiers, while smaller models degrade severely on low-resource languages, and that performance improves when reasoning and summarization are routed through English.
Jiuzhou Han, Abhishek Bijay Mishra, Trevor Cohn et al.
In briefA physics-driven framework for QoS-constrained satellite systems with nonlinear amplifier distortion
SummaryThis paper proposes a physics-driven, self-supervised graph learning framework that directly optimizes a distortion-aware precoder under nonlinear correlated power amplifier distortion, without optimizer-generated labels. It targets a downlink MISO satellite system and minimizes nonlinear PA power consumption under user QoS constraints, jointly designing the precoder and admission control. The authors report that the method closely tracks the consumed-power performance of an SCA-based solution while running faster.
Hamza Ahmed Qureshi, Abuzar Babikir Mohammad Adam, Vu Nguyen Ha et al.
In briefMARC: Mechanism-Aware Reflective Correction for Generating Movable Object Code
SummaryThis paper presents MARC, a method for generating articulated object code from images and text so that the assembled mechanism actually moves as intended. It uses an executable kinematic graph shared across program construction, motion solving, and correction, with a numerical solver tying errors to specific connections. An agent uses these diagnostics to revise the mechanism and retest it. On a 20-object test set, MARC with Kimi-K3 reaches 67.8% valid motion coverage and 60.0% overall success, outperforming six baselines.
In briefFLMNet: A Single-Network CTR Model Trained by Distilling Its Own Dropout Ensemble
SummaryThis paper studies self-distillation in click-through rate prediction, where one network is trained toward the average of its predictions under different dropout patterns. The authors show that as the number of patterns grows, this objective reduces to the label loss of the averaged prediction plus a penalty on disagreement between dropout copies. They propose FLMNet, which approximates this penalty in closed form and trains without random sampling. On five benchmarks against seventeen baselines, FLMNet ranks in the top two for accuracy using a single network at inference.
In briefReinforcement learning controllers for fusion scenario design
SummaryThe paper proposes a single goal-conditioned reinforcement learning controller, trained once in a free-boundary equilibrium and transport simulator, to replace per-scenario controller tuning in tokamak scenario development. Scenarios are specified by target waveforms for plasma current, density, LCFS shape and x-point positions. On EAST plasma-current ramp-up tests across 1024 scenarios, the controller formed a divertor in 86% of cases and reached a diverted state at the episode end in 58%. Failures concentrated in the high current-density region.
Evgeny Adishchev, Maxim Nurgaliev, Dmitry Sorokin et al.
In briefMOSAIC: Toward a Universal Latent Space for Pathology Foundation Models
SummaryThis paper examines whether pathology foundation models (PFMs) learn fundamentally different feature spaces or different coordinate systems of a shared representation. The authors compare 12 patch-level and 6 slide-level PFMs using similarity measures and several alignment strategies. The aligned representations are evaluated through cross-model agreement, retrieval, geometry preservation, and multiple-instance learning on molecular and morphological prediction tasks.
WhySystematically compares latent similarity and alignment across 18 pathology foundation models, evaluated on multiple downstream MIL tasks, a novel combination.
In briefNash-MOPD: Balancing Tasks When Distilling Several Expert Teachers
SummaryThis paper addresses how to allocate a fixed training budget across domains when distilling multiple domain-expert teachers into one student model. It proposes Nash-MOPD, an adaptive scheduler that uses existing reverse-KL statistics in a local minimax game to give more prompts to slower-improving domains. The authors report that it outperforms baselines on mathematics and code metrics without extra teacher queries or rollouts.
In briefA better way to combine several kernels for learning multiple labels from streaming data
SummaryThis paper addresses multi-label classification on streaming data, where examples arrive one at a time. It proposes an online method that combines multiple kernels, using an efficient single-kernel learner and an alternating optimization procedure to learn the best convex combination of kernel classifiers. The authors report asymptotic convergence to the optimal fixed combination of the best single kernels in hindsight, and state that the method outperforms existing online multi-label methods across 11 datasets and seven metrics.
In briefVariance-aware non-stationary bandits: regret bounds for all budgets and the cost of adapting
SummaryThis paper studies bounded non-stationary multi-armed bandits, where reward variance and mean variation are each limited by a budget. It gives minimax dynamic regret bounds within one logarithmic factor for every budget pair. It also proposes calibration and clipping methods for small budgets, and analyzes adaptation costs, showing an extra loss for one policy competing with all budget-based oracle rates.
WhyGives near-minimax dynamic regret bounds under variance and variation budgets, with a clipping-based treatment of the small-budget regime and matching lower bounds.
In briefFrom Retrieval to Recognition: Knowledge-Enhanced Post-training of Time Series Models for Classification
SummaryThis paper proposes Retrieval-to-Recognition (R2R), a plug-and-play post-training method that adapts pretrained time series foundation models to classification. A decoupled query encoder extracts discriminative patterns and retrieves instance-level numerical knowledge from a constructed knowledge base, which is fused with the model's implicit parametric embeddings. The authors report that R2R consistently improves classification performance of state-of-the-art models across 158 time series datasets.
In briefA GPU-based pipeline for multi-robot 3D underwater inspection with limited sensors
SummaryThis paper presents a complete GPU-based pipeline that plans trajectories for a fleet of robots to inspect the 3D surface of underwater structures. Each robot carries a short-range camera with a limited field of view, and the planner jointly covers a target fraction of the mesh while minimizing a combined makespan and total-cost objective. The problem is split into five core stages, with preprocessing and trajectory smoothing around them. Experiments on 14 meshes test the routing cuts and search method, and check collision freedom and waypoint attainment after smoothing.
In briefKeeping Semantic IDs Stable for Generative Retrieval in Dynamic E-commerce Search
SummaryThis paper studies how updating semantic identifier (SID) mappings in generative retrieval affects search quality in dynamic e-commerce settings, using 67 days of large-scale search logs. It finds that remapping SIDs substantially reduces hit rate, with the damage hidden by large beam sizes, and that the fraction of changed SIDs is monotonically linked to lower retrieval quality. The authors propose inserting new items while freezing old-item SIDs and warm-starting from the previous model, which improved old- and new-item hit rates and, in a three-week online A/B test on TmallAPP, raised UCTCVR and GMV.
In briefInteractive system for querying layer-wise explanations of graph neural networks
SummaryThis paper presents SliceGXQ, an end-to-end declarative and interactive system for querying and exploring layer-wise explanations of graph neural networks on large graphs. It offers a graphical interface, a lightweight SPARQL-like query parser, and an execution and visualization workflow built on the SliceGX explanation engine. The demonstration shows how users can diagnose a misclassified node, identify problematic layers and influential messages, and apply the tool to large-scale graph interpretation and fraud-detection analysis.
In briefInferring a bandit learner's internal estimates of how suboptimal each option is, from its choices
SummaryThe paper studies how to recover a no-regret demonstrator's internal beliefs about suboptimality gaps, rather than the true reward parameters that classical inverse bandit methods target. It characterizes a family of randomized policies, such as Thompson sampling, for which gaps can be inverted per step from the arm-selection distribution. The authors identify a stability-coverage trade-off and give conditions under which inversion is reliable, along with theoretical error guarantees. Simulation experiments support the analysis and show how these conditions affect gap recovery in practice.
In briefProvably Detecting Adversarial Images by Comparing Geometry Across Two Spaces
SummaryThis paper studies when adversarial perturbations in vision models can be detected with theoretical guarantees. It compares the geometric structure of samples in a standard CNN embedding space and in a new prime-quantized space that absorbs small perturbations. The authors claim adversarial samples cause unavoidable geometric inconsistencies between the two spaces, and report experiments across many attack types and model families.
Manohar Kaul, Sadbhavana Babar, Aditya Ramesh et al.
In briefReliability First: QuantHarness Governs LLM-Agent Financial Strategy Development
SummaryThis paper introduces QuantHarness, a two-layer agent framework that turns natural-language financial research requests into executable, backtested strategies while enforcing information and action boundaries at each stage. The authors also present FinRelBench, a benchmark that assesses development reliability through execution success, requirement satisfaction, and runtime evidence. The paper reports that QuantHarness achieves the highest task success among the systems compared, and that structured run records support evidence retrieval and decision replay.
In briefMultilingual voicemail detection using pretrained audio embeddings and gradient boosting
SummaryThis paper addresses answering machine detection, determining whether a human or a voicemail picks up a call. The authors propose a lightweight system combining pretrained audio embeddings with gradient boosting machines, trained to prioritize correctly detecting human-answered calls. They also release OpenAMD, the first public AMD dataset, with over 500 recordings in 7 languages.
In briefCEGame-Bench: A Long-Horizon Benchmark for Game-Creating Agents Across Five Engines
SummaryThe paper introduces CEGame-Bench, the first long-horizon benchmark for creating complete games across five engines: Defold, Godot, Roblox Studio, Unity, and Unreal Engine. Agents build games from empty projects and design documents, and a hidden evaluator accepts a game only after real-input play. Agents earn most integration points but usually under half the gameplay points, and only 16% of their games reach an ending. The authors also release Game Completion Boost, an agent skill for playtesting and repair that raises three models' scores by 14.1-24.3 points.
WhyFirst long-horizon benchmark for complete games across five engines, validated by real-input play, revealing a substantial artifact-to-play gap with solid experiments.
In briefPnP-Transformer: A neural network that estimates object pose from 2D-3D matches without iterative fitting
SummaryThe paper presents PnP-Transformer, a differentiable neural solver that estimates 6D object pose from predicted semi-dense 2D-3D correspondences and confidence scores, replacing iterative PnP-RANSAC. It is trained separately from the correspondence module and tested in a zero-shot pipeline. On seven BOP datasets it reaches 41.2% mean AR at 0.46 s per object, outperforming RayPose but not Co-op (58.4%). Swapping in EPnP-RANSAC raises mean AR to 52.2% at an extra 60 ms per object, showing an accuracy-latency trade-off.
In briefHow Robust Are Prognostic Radiomic Features? Comparing Selection Techniques on Glioblastoma Cohorts
SummaryThis study analysed 463 radiomic, anatomical and clinical features from pre-treatment MRI against overall survival in 705 IDH-wildtype glioblastoma patients from three public cohorts. It compared three selection methods: log-rank tests, univariate Cox regression, and XGBoost with SHAP interpretation. Only five features were selected by all three techniques. The authors conclude that analytical pipeline choices, such as ranking method and pre-filtering, can substantially influence which features appear prognostic.
Yannick Sodonougbo, Marlene Schnieders, Alexia Rizoudis et al.
In briefKDHNet: A hybrid model for SSVEP brain signal decoding with little calibration data
SummaryThis paper proposes KDHNet, a hybrid network for decoding steady-state visual evoked potential (SSVEP) brain-computer interface signals. It combines analytical spatial filters (TDCA and TRCA) with a learned SSVEPformer decoder to work with limited subject-specific calibration data. Experiments on three public datasets report the highest mean accuracy and information transfer rate in most intra-subject settings, and the best or second-best results in inter-subject evaluation, with gains generally larger when calibration data are scarce.
In briefGeoNaviDiff: Transferring disentangled StyleSpace edits to diffusion model features for real-image editing
SummaryThis paper presents GeoNaviDiff, an unsupervised framework that transfers disentangled StyleSpace edits from StyleGAN to local Riemannian trajectories in a diffusion model's h-space. It extracts a spherical Log-map tangent from a matched StyleGAN reconstruction-edit pair and parallel-transports it to the real-image diffusion state. A shared, factor-conditioned Navigator predicts this tangent at inference. The authors report better results than compared methods in fidelity, factor selectivity, edit locality, and trajectory smoothness.
In briefRGB-event object detection across operating frequencies via reliability-supervised routing and asymmetric-resolution fusion
SummaryThis paper addresses object detection for autonomous vehicles by combining RGB frames with event camera data, which record brightness changes at microsecond resolution, to make decisions faster than the frame rate. It proposes reliability-supervised routing, which assigns per-location fusion weights supervised by losses from two single-modality classifiers used only during training, and asymmetric-resolution fusion, which encodes frames at native resolution. On the DSEC-Det dataset, routing halves the accuracy loss from 20 to 180 Hz (6.3 to 3.1 mAP). The resulting detector, ERA-Det, reaches 63.2 mAP, 3.7 mAP above FlexEvent.
In briefEvaluating Tibetan Speech Recognition Across Dialects and Adaptation Regimes: A Multi-Granularity Analysis
SummaryTibetan speech recognition is usually judged by character or syllable error rates, since Tibetan script does not mark word boundaries, and these metrics can rank systems differently. The paper combines character, syllable, and segmented word error rates in one framework and compares dictionary-based, neural, and language-model-based segmentation against expert annotations. It evaluates five ASR architectures on Lhasa Tibetan and tests zero-shot transfer and target-corpus adaptation on five other corpus subsets. The results show that lower character error does not always mean lower syllable or segmented-word error, and these ranking differences persist across corpora.
In briefContext-aware graph network distilling EEG and eye movements for emotion recognition
SummaryThis paper proposes CTD-GNet, a model that fuses EEG and eye-movement signals for emotion recognition, using EEG features as nodes of an electrode graph modulated by eye data and preceding windows. Bidirectional distillation transfers knowledge from this multimodal teacher to an EEG-only single-window student. In leave-one-subject-out tests on SEED and SEED-IV, it reports the highest mean accuracy and remains robust when eye data are missing.
In briefDeOverlay: A Benchmark and Method for Removing Video Overlays
SummaryThis paper presents DeOverlay, a method that removes subtitles, titles, and logos from videos while recovering the hidden scene content. The authors also introduce DVTE-Bench, a benchmark of 1,631 paired videos covering eight text-overlay categories. According to the paper, DeOverlay outperforms four baselines on all five reconstruction and perceptual metrics, reducing Mask LPIPS by 58.3% and Mask VFID by 70.3% relative to the strongest baseline.
In briefFacial Deidentification Evaluation Toolkit: An Open-Source Framework for Testing Privacy Protection Methods
SummaryThis paper presents an open-source toolkit for evaluating face deidentification techniques. It provides a complete pipeline covering preprocessing, deidentification, and performance assessment based on data utility, image quality, and identity protection metrics. The toolkit supports a variety of datasets, making it easier for researchers to benchmark new methods. The authors report that some techniques keep privacy protected while preserving image utility, and they aim to offer a standard evaluation framework for facial biometric privacy technologies.
Esteban Leiva-Montenegro, Matthieu Pillot, Manfred Gonzalez-Hernandez et al.
In briefDCR-TSF: Training-Free Zero-Shot Image Retrieval Using Multi-Level Descriptions and Spherical Fusion
SummaryThis paper presents DCR-TSF, a training-free framework for zero-shot composed image retrieval, which finds a target image matching a reference image plus a text modification. It generates fine-grained and coarse-grained descriptions through Dual-level Compositional Reasoning and fuses them with the original composed query using Triadic Spherical Fusion on the unit hypersphere. The authors report state-of-the-art results among training-free methods on the CIRCO, CIRR, and Fashion-IQ datasets.
In briefNot All Similarity Is Leakage: Assessing Identity Leakage in Synthetic Face Datasets
SummaryThis paper examines whether synthetic face datasets reveal identities from a generator's training data, which would weaken their privacy benefits. The authors identify limitations in prior methods, including heavy reliance on classifiers and mislabeling similarity to unused real datasets as leakage. They propose Scaled Overlap (SO), a distribution-aware metric using a multi-model setup to assess identity leakage more reliably. The study finds the strongest leakage in CemiFace, SFace2, and IDNet, and reports that most synthetic images draw from multiple identities rather than a single one.
Pawe{\l} Borsukiewicz, Daniele Lunghi, Wendk\^uuni C. Ou\'edraogo et al.
In briefCUA-Lite: An Open Platform for Computer Use Agents
SummaryThis paper presents CUA-Lite, an open platform for developing computer use agents (CUAs) built on three standardized parts: a unified environment interface, a unified supervised data format, and model harnesses shared across evaluation, RL, and SFT. It integrates 12 benchmarks across desktop, browser, and mobile, along with lightweight sandboxes that provide over 30,000 verifiable tasks. The authors report that RL and SFT run through the same harness improve open-weight models, and they release the dataset, code, and models.
In briefDecomposing risk to reduce physical hallucinations in video sampling
SummaryVideo models produce visually convincing clips but still show physical hallucinations, such as a ball teleporting or bouncing higher without added energy. The paper proposes a two-dimensional risk framework separating low-density sampling risk from high-density drifting risk. Based on it, the authors introduce HA-Sampler, which adds external physical guidance during latent refinement, and HA-Scorer, a likelihood-based evaluator using continuous normalizing flows. Experiments on Physics-IQ and VideoPhy reportedly outperform compared state-of-the-art baselines.
In briefPhysSkin++: Real-time, generalizable physics animation using learned neural skinning
SummaryThis paper presents PhysSkin++, a framework that learns continuous neural skinning fields from static 3D geometry to build compact deformation subspaces for physics-based animation. It uses a Neural Skinning Fields Autoencoder and a self-supervised learning strategy with conflict-aware gradient correction, so no simulation trajectories are needed. Experiments show the learned skinning generalizes across shapes, supports real-time reduced simulation, and enables controllable, collision-aware, and rig-complementary animation.
In briefFinSafeGuard: A Quality-Filtered Synthetic Dataset for Safer Financial AI
SummaryThis paper presents a pipeline that generates large-scale synthetic training data for financial AI from 632,000 scenario specifications. It combines generation parameter research, automatic repair of low-diversity batches, two-layer semantic deduplication, and LLM-based quality filtering. Across Banking, Insurance, and Financial Services, deduplication removed 9.9% to 40.6% of samples, reducing 26.4 million raw samples to 14.3 million.
Joseph Thomas Thacil, Reetu Raj Harsh, Krati Saxena et al.
In briefLearning the Dose Manifold: A Latent Diffusion Framework for 3D Radiotherapy Dose Prediction
SummaryThis paper presents LaDose, a latent diffusion framework that predicts full three-dimensional radiation dose distributions for head-and-neck radiotherapy planning. An autoencoder compresses dose maps 128-fold so 3D diffusion can run on a single GPU, and multiscale classifier-free guidance lets clinical trade-offs be adjusted at inference. On 900 clinical head-and-neck plans from five centers, the median absolute error was 1.18 Gy, and the pipeline produced plans in 15.2 minutes with direct aperture optimization.
In briefProcedural Exposure Bias in Skill-Augmented Language Model Agents: how repeated skill exposure shifts agent action preferences and how visual versus text skill presentation changes this
SummaryThe paper studies how language-model agents use external skills, which are reusable procedural knowledge. It identifies Procedural Exposure Bias, where models increasingly favor an action as its procedural pattern is repeatedly shown in a skill, even when the state and action relevance stay the same. Presenting the same content visually instead of as text substantially reduces this preference shift, and visual skills achieve higher macro-average task success than text skills across ALFWorld, ScienceWorld, and WebShop for all four evaluated models.
In briefHoneyed Words, Hidden Daggers: Testing and Training LLMs to Detect Mock Politeness
SummaryThe paper introduces ZhenhuanBench, a 1,592-item Chinese benchmark built from expert commentary on the drama Empresses in the Palace, testing speaker intent, emotion-language mismatch, and listener uptake. Text-only evaluation of 34 LLMs found GPT-6 Astra highest at 73.48% overall. The authors also propose CPR-GRPO, a training loop that reportedly outperforms SFT and GRPO on several external datasets.
In briefHierDTA: Hierarchical Structure-Aware Learning for Predicting Drug-Target Binding Affinity
SummaryThis paper introduces HierDTA, a framework that predicts how strongly drugs bind to target proteins, helping prioritize compound-protein pairs before lab testing. It represents drugs as atom-motif graphs and kinase domains as residue interaction graphs, integrating them with cross-attention alongside molecular fingerprints and ESM-2 embeddings. The authors report superior results on the Davis dataset across all metrics, an r_m^2 of 0.791 on KIBA, and better generalization under cold-start settings.
In briefORBIT: Autoregressive time series generation by predicting the next frequency band
SummaryORBIT is a unified deep generative framework that synthesizes realistic time series to address data scarcity. It combines a Frequency-Hierarchical VQ-VAE, which splits latent representations into frequency bands and quantizes them residually, with a Frequency-Band Autoregressive Transformer that generates from low to high frequencies. The authors report state-of-the-art overall performance on multiple real-world benchmarks and better inference efficiency than advanced baselines.
In briefCorteXforge: Generating Reproducible Radio Signal Datasets on Real Hardware Testbeds
SummaryThis paper presents CorteXforge, a framework that generates realistic and reproducible radio frequency datasets on programmable software-defined radio testbeds. High-level RF scenarios are automatically turned into synchronized experiments run on the SLICES-RI/CorteXlab platform, producing annotated SIGMF datasets that capture hardware impairments and propagation effects. The authors validate the framework with automatic modulation classification on an over-the-air dataset, finding a synthetic-to-real performance gap while showing that synthetic pretraining still helps adaptation to over-the-air data.
Andr\'ea Joly, Leonardo S Cardoso (CITI), Jean-Marie Gorce
In briefAVTime: Bidirectionally Aligned Reinforcement for Dense Audio-Visual Captions
SummaryThis paper addresses dense, time-stamped audio-visual captioning for video understanding. It identifies gaps in existing work: a lack of suitable datasets, reliance on checklist-based reinforcement learning, and no benchmark for bidirectional evaluation. The authors introduce AVTime-50K, a 50K-scale dataset, the AVTime model trained with a Bidirectional Temporal-Semantic Alignment (BiTSA) reward, and AVTime-Bench for assessing temporal accuracy and semantic correctness.
In briefBalance-aware pivot selection for faster parallel decoding in diffusion language models
SummaryThis paper proposes Balance-Aware Pivot Decoding (BPD), a method for parallel token generation in diffusion language models. It favors central pivot positions within contiguous masked intervals and decodes the selected positions in parallel, combining confidence-based prefilling with confidence-thresholded center selection. The authors evaluate the quality-efficiency tradeoff on LLaDA and Dream across five benchmarks covering mathematical reasoning, code generation, and instruction following.
In briefFinHopper: A step-by-step benchmark for testing financial AI agents on multi-hop investment questions
SummaryThis paper introduces FinHopper, a benchmark that tests financial agents answering investment questions through open search, with reasoning depth from 1 to 6 hops across stocks, funds, and macroeconomic facts. It provides gold intermediate answers so that the reasoning process, not just the final answer, can be evaluated. Experiments with 20 LLM-agent configurations show that performance drops as hop depth increases, and the best model remains far below human experts.
WhyProposes a controlled-depth financial multi-hop benchmark with gold intermediates, validated across 20 agent configurations, a substantive contribution to a subfield.
In briefDecision4Rec: a multi-agent recommender framework using decision models
SummaryThe paper presents Decision4Rec, an open, modular framework that applies decision models, which return a probability for each option, to recommendation. A knowledge graph writes scenario logs as short statements, an agent pool proposes candidates, and a residual ranker keeps the best 30, with a decision model's answers fused into the ranker's scores. With a trained 4B decision model, the framework scores 25 to 36% above the best trained baseline in NDCG@10 on three sparse Amazon catalogues. On the dense Steam dataset it roughly ties with baselines.
Marco Polignano, Het Darshan Mehta, Pasquale Lops et al.
In briefDRA-SAM: Disagreement-driven local re-inference for referring segmentation in low-altitude drone images
SummaryThis paper addresses referring segmentation in low-altitude drone imagery, where small, visually similar targets are hard to pick out. It proposes DRA-SAM, which after global inference selects one image crop for extra local inference based on disagreement among candidate masks and unresolved regions. The authors report state-of-the-art results on most metrics across RIS-LAD and UAVid-RIS, including a mIoU increase from 49.55 to 52.34 on RIS-LAD.
In briefTo Truncate or Not? A Scaling Study of Looped Transformers with Truncated Backpropagation
SummaryThis paper studies whether truncating backpropagation through time (BPTT) is worthwhile when training looped transformers, which reuse shared layers to add depth without adding parameters. Across nearly 600 training runs, the authors fit a unified scaling law and find that full BPTT is favored when forward and backward loop counts are jointly optimized. At matched compute, full BPTT through fewer loops outperforms truncated settings. When model size is also free, two fully trained loops achieve lower validation loss than the compute-optimal non-looped model, with about 27% fewer non-embedding parameters and 8% more training compute.
In briefInstance-dependent regret optimality for episodic MDPs via posterior sampling
SummaryThe paper studies regret minimization in finite-horizon episodic Markov Decision Processes. It introduces π0-PSRL, a variant of PSRL that uses posterior samples to decide when to explore while following a fixed reference policy during exploration episodes. The authors prove instance-dependent regret bounds, identify a logarithmic exploration cost induced by π0, and take a step toward matching asymptotic lower bounds for episodic reinforcement learning.
In briefAfter federated learning: a benchmark for transferring knowledge to newcomers with different architectures
SummaryThe paper introduces FEDNBO, a seven-dataset benchmark for testing whether knowledge left by a federated learning system can be transferred to newcomers using a newly chosen architecture. It compares public-response distillation with strengthened local training, reporting gains of 23.52 and 16.62 percentage points on CelebA and FedEMNIST. On FedEMNIST, the reference method reaches 97.13% accuracy in 12.4 seconds per architecture.
In briefTamedFlow: Trajectory-Aware Distillation for Faster Teacher-Assisted Diffusion Image Generation
SummaryThe paper addresses slow inference in Diffusion Transformers, which need many sequential model evaluations for high-quality image and video generation. It finds that combining model compression with adaptive inference can degrade quality, and proposes TamedFlow, which trains a compressed student with trajectory supervision and uses a learned estimator to accept drafts or fall back to one teacher step. Distilling DiT-XL/2 into DiT-B/2 on ImageNet 256×256 reports an FID of 2.58 with a measured 2.25× speedup over DDIM-50.
In briefMultimodal Decoupling for Diversified Recommendation via Semantic and Structural Shaping
SummaryRecommender systems optimized for accuracy often produce over-concentrated results with limited diversity. The paper proposes MOSS, a framework that decouples multimodal content, selects category-aware semantic features, and uses alternating dual-graph propagation and entropy-based diversity regularization. According to the authors, experiments on real-world datasets show MOSS achieves a better accuracy-diversity trade-off than strong baselines.
In briefSPEM: Structure-aware memory that helps LLM agents reuse skills more efficiently
SummarySPEM is a lightweight memory framework that helps LLM-based agents use reusable skills more efficiently. It stores individual skills as atomic memory and stores successful skill compositions as Skill-DAG templates that can be reused for new tasks. When no high-confidence template fits, it retrieves atomic skills and uses a Skill-GNN to select them with attention to state and structure. Experiments on ALFWorld and SkillsBench report 87.1% ALFWorld success and 38.38% SkillsBench reward, with lower token and runtime costs than strong baselines.
In briefFrameGRPO: Improving Video Reasoning via Multi-Frame Rollouts and Temporal Credit
SummaryThis paper proposes FrameGRPO, a reinforcement learning framework that improves video reasoning in multimodal large language models without auxiliary networks or architecture changes. It uses multi-frame rollouts, which build aligned views from uniform temporal subsampling of one maximal-frame decode to diversify visual evidence and reasoning trajectories, and gated temporal efficiency credit, which reinforces a correct lower-ratio trajectory only when all full-ratio reference rollouts fail. Experiments across diverse video understanding and reasoning benchmarks show overall improvements over standard GRPO.
In briefWhen Reflective Prompt and Skill Optimization Helps Local Reasoning Models
SummaryThis paper tests whether the text-space optimizers GEPA and SkillOpt replicate on two locally served reasoning models, DeepSeek V4 Flash and Qwen3.8-27B. It compares a baseline, a GEPA prompt, a SkillOpt skill, and a manual prompt on HotpotQA and MATH-500. GEPA raises exact match on MATH-500 for both models, while neither optimizer improves the HotpotQA baseline, and validation gains can reverse on held-out evaluation.
Andrea Giummarra (1), Carmelo Baglieri (1), Antonio Di Maria (1) et al.
In briefAerial-R1: Strengthening Reasoning and Segmentation in Aerial Multimodal Models
SummaryAerial-R1 is the first multimodal large language model framework designed to strengthen visual reasoning and segmentation for aerial imagery. It addresses hallucination caused by extreme scale variation and ambiguity by introducing Hybrid-View Group Advantage-Weighted Regression, which uses groups of high- and low-resolution views to enforce consistency on clear inputs and penalize hallucinations on ambiguous ones. Experiments on the VSAI-Ref benchmark show higher segmentation precision than supervised fine-tuning baselines and substantially fewer hallucinations in low-information scenes.
In briefProtGlycanDock: A Dataset, Benchmark, and Model for Protein-Glycan Docking
SummaryThis paper addresses the lack of dedicated datasets, benchmarks, and models for predicting how proteins and glycans dock together. The authors build the ProtGlycanDock benchmark and evaluate ten models across physics-based, machine-learning, and structure-prediction categories. They also propose two finetuning methods that add glycan conformation knowledge to a general molecular structure model, and report that AlphaFold 3 performs best among existing methods while the combined techniques reach state-of-the-art results.
In briefAccelerated Drift in Scientific Writing After LLMs: An Interrupted Time-Series Study
SummaryThis paper examines whether the writing trajectories of scientific preprint abstracts changed after ChatGPT's release in November 2022, comparing arXiv and medRxiv texts before and after that date. Using interrupted time-series regression, the authors find substantial post-2022 changes in vocabulary, readability, and syntax across several domains, such as longer average word length and lower dependency-tree depth. Individual features behave differently, with some lexical markers accelerating persistently and others rising then receding. The authors argue that monitoring scientific writing as a multidimensional, evolving system is more informative than tracking a fixed set of markers.
In briefLarge synthetic pre-training data and evaluation suite for Indic language models
SummaryThis paper introduces Nemotron-Pretraining-Indic-v1, a synthetic pretraining corpus of about 2.83 trillion tokens covering 22 scheduled Indian languages and twelve romanized forms. The authors translate English documents with a GPU pipeline and repair low-quality outputs. Continued-pretraining experiments across twelve languages show that combining the synthetic corpus with natural Indic text improves knowledge and reasoning over a public-corpus baseline. The paper also presents Indic-LLM-Eval-Suite, a multilingual benchmark for chat, instruction following, reasoning, knowledge, code, and translation.
WhyBuilds a large synthetic Indic pretraining corpus and benchmark suite, with continued-pretraining gains over a public-corpus baseline across twelve languages.
In briefTabular Surrogates for Surface Irrigation Water Height Prediction: Same-Sector and Cross-Sector Transfer Study
SummaryThis paper studies tabular surrogate models that predict maximum water height in surface irrigation from drone-derived field morphology, terrain, hydrological, spatial, and scenario features, with each row being a spatial cell under one irrigation scenario. Gradient-boosted trees, neural tabular models, AutoML, and tabular foundation models are compared under same-sector generalization and sector-out transfer. Under a fixed 8/32 same-sector split, the surrogates perform strongly, but transfer to unseen sectors is substantially harder. Sector-frame normalization of spatial coordinates helps transfer between similarly structured sectors but does not solve cross-shape generalization.
Alessandro Cagiano, M. Aybars Seyhan, Fabiola Gangi et al.
In briefCurved Markers for Monocular Pose Measurement of Cylinders
SummaryThis paper proposes a dual-colour-stripe marker for thin cylindrical objects, using the object's known curvature to improve pose measurement with a single camera. It analyses how features at different depths improve estimation of rotation about the cylinder axis (roll). For a marker 15 cm away, the Cramér-Rao lower bounds on roll standard deviation are 2.00° for a flat marker, 0.38° for a single curved marker and 0.21° for a wrap-around marker. Monte Carlo roll RMS errors are within about 10% of these bounds.
In briefTimesJev: Fast time-series reasoning with a System 1 model
SummaryThis paper proposes TimesJev, a System 1 model for time series reasoning that avoids the long chains of thought used by many reasoning models. It compresses raw time series into a representation that keeps key frequency patterns and aligns with the model's embedding space. Experiments on 13 datasets report that TimesJev outperforms reasoning models by over 62% on average, with 62.9% less inference time than their fine-tuned versions.
In briefPnP-CF: A consensus flow-matching solver for general inverse problems that reconstructs data from indirect measurements.
SummaryThe paper presents PnP-CF, a plug-and-play framework that uses a flow-matching generative model as a learned data prior to solve model-based inverse problems. It combines ADMM-like dual variable updates with a time-adaptive pullback onto the intermediate flow path to stabilize the algorithm and promote consensus between solver and sampler. Experiments show it matches or exceeds state-of-the-art baselines on linear inverse problems and extends to nonlinear problems, including those with pretrained neural-network forward operators, where existing methods perform poorly or fail.
In briefTempoMoE: Steering Fast and Slow Thinking Experts to Make Multimodal Reasoning More Efficient
SummaryMultimodal mixture-of-experts reasoning models often produce long chains of thought even for easy inputs, raising token cost and latency. The paper finds that fast and slow thinking activate distinct expert-routing patterns, and that suppressing slow-thinking experts shortens reasoning while suppressing fast-thinking experts lengthens it. It proposes TempoMoE, a training-free method that identifies these experts and reweights routing by sample difficulty. Tested on Kimi-VL, Qwen3-VL, and ERNIE-4.5-VL across six benchmarks, it improves accuracy by up to 2.6 points while cutting output tokens by up to 35.4%.
In briefFolding chain-of-thought into looped transformers to trade tokens for loop iterations
SummaryThe paper builds a family of constant-bit-size looped transformers that interpolates between chain-of-thought, which generates tokens, and latent thought, which recurs in hidden states. For a loop count L, the model simulates a time-t(n), space-s(n) Turing machine using a context window of O(s(n)/L) and O(t(n)s(n)^{1/k}/L) generated tokens. This reduces both resources by a factor of L compared with the unlooped queue-machine simulation. The authors also state that the construction holds with softmax attention at constant bit-size under a specified finite-precision kernel.
WhyConstructs constant-bit looped transformers interpolating CoT and latent thought, cutting context and generated tokens by factor L; a rigorous theoretical advance over unlooped simulation.
In briefFederated Community Detection in Bipartite Networks: a privacy-preserving method that finds groups in networks by running distributed matrix computations without sharing raw data
SummaryThe paper studies how to estimate spectral features of a large matrix when servers hold separate data and cannot share raw data. It proposes FLICKER, which recovers prescribed singular directions using only matrix-vector products on local blocks, and builds BiFLICKER for community detection in bipartite networks. Under a sparse bipartite stochastic block model, the authors give convergence and misclustering results, and on a LastFM user-artist network BiFLICKER reaches an overall ARI of 0.94 against centralized clustering.
In briefProcedural memory for AI language agents using explicit goal structures
SummaryThe paper argues that large language models are unreliable on long-horizon tasks needing procedural reasoning, partly because they lack explicit procedural memory, meaning accessible and modifiable knowledge of how tasks are done. It proposes Goal-Oriented Graphs, which encode goals and prerequisite dependencies as inspectable symbolic structures. The authors examine how these graphs are built, the challenges they raise, and principles for integrating them with neural reasoning.
In briefHow graph-to-table interfaces shape relational generalisation
SummaryThis paper studies whether the encoder, context constructor, and tabular decoder in graph-to-table systems can be assessed independently. Using matched comparisons on CLUTRR, 57 GraphLog worlds, and 57 knowledge graphs with logistic regression, CatBoost, and TabPFN-3, the authors find that conclusions about one component can change when the others change. The results show encoder assessment depends on the decoder and context.
Alessandro Cagiano, Lorenzo Molfetta, Gianluca Moro et al.
In briefMixing specialized route-planning experts for multi-robot path finding
SummaryThis paper studies neural control policies for multi-agent pathfinding, which guide many agents such as warehouse robots to their goals without collisions. The authors combine several small experts, each trained for a different map distribution, using a learned router that weights their outputs per agent. They test this mixture-of-experts approach on a graph neural network policy and a new recursive action-token refinement model. The mixture generally beats the pretrained policy and leading methods, but parameter-matched single models trained well can still outperform it.
In briefClueFlow: Learning Long-Horizon Search from Traceable Research-State Changes
SummaryThe paper introduces ClueFlow, a method that links evidence-gathering browser actions to state updates and the final answer, forming traceable attribution chains for search agents. These chains guide which decisions receive supervised fine-tuning and how credit is assigned during reinforcement learning. Evaluated on two Qwen backbones and four search benchmarks, ClueFlow reports consistent gains from fine-tuning and further improvements from PPO.
In briefGraphEvoResearcher: Training-free graph-based self-reflection for low-cost multi-hop question answering
SummaryThe paper proposes a training-free protocol that organizes intermediate conclusions and evidence into an explicit reasoning graph, using structural signals to decide whether to revise, backtrack, or stop. On GSM8K, MATH, and BBH (50 questions each, using the DeepSeek model), it reaches 98%, 92%, and 82% accuracy, with the largest gain on the hardest set, MATH. Construction cost stays bounded at about 2 steps and 4.5 to 5 nodes.
In briefUncertainty estimation for diatom species classification and optimizing expert re-annotation
SummaryThis paper studies automated diatom identification from microscope images across a 759-class dataset using a ResNet-18 baseline. Temperature scaling reduces Expected Calibration Error from 0.117 to 0.055, enabling selective classification. At 66.2% coverage, the system reaches a Selective Macro-F1 of 99.01%, deferring rare taxa and ambiguous cases to experts.
Kamenan N'Gadi (LORIA, BISCUIT), Martin Laviale (LIEC et al.
In briefAsymmetric target distributions for self-distillation policy optimization, depending on answer correctness
SummaryThis paper proposes APSD, a method for post-training large language models that builds correctness-dependent training targets by mixing student and self-teacher distributions. Interpolation preserves successful behavior on correct rollouts, while extrapolation beyond the self-teacher strengthens correction on incorrect ones. Experiments with reinforcement learning using verifiable rewards and rich feedback report strong overall performance against baselines and SDPO-level results with about 2.6x fewer training steps across Qwen3 model scales.
In briefGLOBE: Stopping Chain-of-Thought Reasoning Early Using Trajectory Shape
SummaryThis paper proposes GLOBE, a training-free online method that stops a language model's chain-of-thought reasoning early. It maps each reasoning segment to a point on a unit sphere and uses the shape of the resulting trajectory, along with short trial answers checked against a confidence threshold, to decide when to stop. Across five benchmarks, including math reasoning, tool calling, and agent memory, GLOBE reduces output tokens by 49.7% and end-to-end time by 51.0% on two models, with a five-task average accuracy of 80.58%, comparable to full chain-of-thought.
