In briefOnline interval selection with binary predictions: optimal consistency and robustness tradeoff
SummaryThis paper studies online interval selection where each request comes with a binary prediction of whether it belongs to a final optimal solution. Requests arrive in arbitrary order, and accepted requests may be revoked, but rejected or revoked requests cannot be recovered. The authors show that the best consistency under finite strict robustness is (3+√5)/2, and give a deterministic algorithm achieving this consistency with exact robustness φ^5. They also present a continuous family of algorithms with a tunable replacement cost, and show that randomization allows expected consistency approaching one with finite expected robustness.
In briefTimeLineXR: Interactively Exploring How Vision-Language Models Encode Time
SummaryThis paper presents TimeLineXR, an interactive demo that turns the time axis inside vision-language models into a 3D scene. Users can upload images on a web page or in a WebXR environment and see each placed on a color-coded time axis. The system lets users compare where objects land and read their estimated first year of appearance.
Nidham Tekaya, Markus Passecker, Matthias Zeppelzauer
In briefAligning institutional power with operational ability
SummaryThis paper examines how multi-agent AI systems should be designed so that responsibilities match the abilities needed to fulfil them. It proposes a formal framework based on process-oriented modelling and institutional power to analyse normatively relevant processes. The authors also discuss strategies for repairing such processes and explain how the relation between operational processes and institutional states affects the lifecycle of normative positions.
Andrea Marino, Giovanni Sileno, Thomas van Binsbergen et al.
In briefCodeCap-Bench: Can code serve as a good video caption? A benchmark testing whether executable code can describe video content.
SummaryThe paper explores describing videos with executable code, using graphical primitives, layout rules, and time functions instead of natural language. In question answering and video reconstruction tasks, code captions outperform natural language captions. The authors introduce CodeCap-Bench, which contains 323 videos, to test multimodal large language models on generating code captions. Leading proprietary models show promising results but still struggle with complex temporal dynamics, while open-source models lag far behind.
In briefEEG-based depression detection using healthy normative deviation modeling
SummaryThis paper trains structural and distributional reference models on healthy control EEG recordings to capture how an individual's EEG departs from healthy patterns. It proposes two diagnostic approaches: one summarizes deviations into scalar scores, and the other keeps dimension-wise residuals and fuses them with attention-based aggregation. Experiments on two public EEG datasets, under subject-independent within-source and cross-source settings, show the approach is effective.
In briefCoS-VQ: Coordinated Shared-Scale Vector Quantization for Large Language Models
SummaryThis paper proposes CoS-VQ, a vector quantization method for low-bit weight compression of large language models. It keeps codewords low-dimensional while jointly optimizing each position's index and a shared scale, addressing the coupling between scale and codeword combinations. On Qwen3.8-27B, the authors report it outperforms QTIP and LLVQ on the evaluated language modeling, coding, and reasoning benchmarks.
In briefWCC: A Learning Framework for Deciding When to Use Cooperative Coevolution in Large-Scale Optimization
SummaryThis paper proposes WCC, a learning-based framework that decides when to apply Cooperative Coevolution (CC) in large-scale global optimization. It formulates the decision as a Markov decision process and uses an auto-regressive policy to switch dynamically between CC and non-decomposition algorithms while allocating the computational budget. The authors report strong competitiveness and generalization across problems, optimization horizons, and baseline optimizers.
In briefScIntBench: Testing whether generated 3D rooms hold up when objects are physically manipulated
SummaryThis paper introduces ScIntBench, a benchmark for text-conditioned 3D indoor scene generation that focuses on interaction readiness. It runs prescribed actions such as pushing, opening, and surface probing in a physics simulator and measures task success, post-interaction integrity, and dynamic plausibility. The authors evaluate 12 state-of-the-art scene generation systems across 10 room types and find that strong static physical performance does not guarantee interaction readiness, with the best system succeeding on only 36.92% of tasks.
In briefLoss function that keeps fine-scale weather details in kilometre-scale AI forecasts
SummaryThe paper introduces a wavelet-based loss function that constrains atmospheric fields by scale, location, and orientation to preserve fine-scale structure in kilometre-scale machine-learning weather prediction. It is evaluated in a joint global-regional configuration producing 1.5 km forecasts over the UK alongside roughly 31 km global forecasts. Compared with mean-squared error and an FFT-based spectral objective, the wavelet loss substantially improves spectral fidelity while keeping pointwise accuracy similar to MSE training and improving neighbourhood-based verification.
James Chappell, Sophie Luise Arana, Jack Bowyer et al.
In briefUCVE: Graph-Structured Retrieval for Checking Compliance Requirements
SummaryThis paper introduces UCVE, a framework for checking whether a document satisfies, violates, or fails to address every obligation in a predefined set of compliance requirements. It represents the requirements as a graph and uses reinforcement-learning-weighted spectral reranking for retrieval. Across three compliance domains, UCVE averages a 4.3% matched-K recall gain over flat cosine retrieval and outperforms strong baselines.
Emily Chen, Michelle Menachery, Himanshu Kumar et al.
In briefHierarchical sparse graph for linking event-camera tracklets across multiple batches
SummaryExisting event-based trackers link tracklets only between adjacent batches, which fails under long occlusions or missed detections. This paper performs multi-dimensional assignment that associates tracklets across several batches at once, using fuzzy logic to quantify pairwise correlations and a Hierarchical Sparse Graph to model them. The assignment is solved as a graph partition task with a training-free heuristic. On the eTram and Ev-UAV event datasets, the method outperforms existing event-based methods and most image-based association models, reaching over twice OC-SORT's performance in many-to-many association.
In briefAnalogScout: PVT-Aware Analog Circuit Sizing with LLM Agents for Simulation Efficiency
SummaryAnalogScout is a framework for sizing analog devices so that they meet specifications across all process, voltage, and temperature corners. It reuses verified elite designs and applies analytic gm/Id local sizing before global refinement, reducing the number of costly simulations. On three SKY130 OTAs with a budget of 16k simulations, it cuts power by 43.3% relative to CMA-ES and reaches CMA-ES's final power after 7.0% of that budget.
In briefReForest: Combining Interpretability and Reasoning in Isolation Forest for Anomaly Detection
SummaryThis paper proposes ReForest, a framework that combines the statistical efficiency of Isolation Forest with the semantic reasoning of large language models for anomaly detection. It converts Isolation Forest's decision paths into natural-language explanations and uses an uncertainty-aware dynamic ensemble to weigh statistical and semantic signals. Experiments on four real-world healthcare and finance datasets report an average 21% improvement in Recall and a 10% gain in F1-score.
In briefA lightweight attention CNN with saliency maps for explainable skin cancer detection
SummaryThis paper proposes a lightweight attention-based convolutional neural network (CNN-Attention) for binary classification of skin lesions, incorporating saliency maps to visualize regions that influence its decisions. Channel attention is used to focus on informative areas of dermoscopic images while keeping computational cost low. On a public dataset of 2,637 training and 660 test images, the model achieved 87 percent accuracy with 1.95 million parameters and 7.45 MB storage, matching MobileNet while using about half the parameters and storage. Its ROC AUC and PR AUC were 0.9475 and 0.9430.
Sanjana Islam Kasfia, Md. Ismiel Hossen Abir, AKM Bahalul Haque et al.
In briefMvT: Event-camera multi-object tracking using multi-view projection
SummaryEvent cameras output asynchronous data, and common methods convert it into pseudo-frames, which can cause tracking failures when trajectories overlap or are close. This paper proposes MvT, a multi-view tracking pipeline that keeps a 2D format while adding spatio-temporal views to resolve single-view ambiguities, then fuses results in a 3D coordinate system and associates tracklets via an optimization model. The method is training-free and was evaluated on four datasets, including the authors' Small Objects Dataset (SOD), where it outperforms existing trackers.
In briefProbing structured internal representations to anticipate risks in end-to-end self-driving
SummaryThe paper introduces the Structured Representation Probe (SRP), a lightweight monitor that reads actor, ego, and map representations from a frozen driving policy to anticipate failures. It learns a separate linear risk direction for each source and uses event-window supervision to link the readout to upcoming hard failures. On Bench2Drive with DriveTransformer, the probe reaches an ROC-AUC of 0.833 at a 2-second horizon, compared with 0.715 for temporal consistency, at 2.67% readout runtime overhead.
In briefTRUCO: A Scalable Lock-Free Parallel Algorithm for Extracting Contours
SummaryThis paper presents TRUCO, a lock-free parallel algorithm that converts binary masks into geometric contours. It uses a start-point ownership rule and a speculative tracing model so that each contour is processed exactly once. Experiments on synthetic and real-world data, run on consumer and server-grade CPUs, reportedly show significant speedups over existing baselines, over an order of magnitude compared with the industry standard, with good scalability on high-resolution scenes.
Rafael Munoz-Salinas, Francisco J.Romero-Ramirez, Manuel J. Marin-Jimenez
In briefMulti-Label Agreement: A Framework for Measuring Annotator Consistency
SummaryThis paper introduces Multi-Label Agreement (MLA), a flexible framework for measuring how consistently annotators label items in multi-label tasks. It keeps each item's full binary label vector and uses average pairwise Hamming agreement, counting both joint selections and joint omissions. Through a worked example and Monte Carlo simulations, the paper shows how results depend on label prevalence, the declared label inventory, and the chosen expected-agreement model. A Python implementation is also provided.
In briefImproving noise initialization for zero-shot video restoration using image diffusion models
SummaryThis paper proposes a zero-shot video restoration method that uses a pre-trained image diffusion model instead of a costly video diffusion model. It generates initial noise by applying DDIM inversion to the corrupted video frames, preserving temporal correlations between frames. The authors add an adaptive noise blending strategy and measurement-guided trajectory attention to improve temporal consistency. The method is reported to outperform existing zero-shot approaches on the DAVIS, REDS4, and Set8 datasets.
In briefBayesian Inference for Expensive Simulators: A Survey and Comparison of Design Choices
SummaryThis paper surveys Bayesian inference methods that use simulators as forward models, especially when simulations are very costly. It unifies existing Bayesian optimization-based approaches into one framework and systematically compares design choices by the simulation budget they require. The authors find the surrogate model's response variable strongly affects efficiency and propose custom proxy variables, with a Julia package provided.
In briefRPSD: Retrieval-privileged on-policy self-distillation with optimal transport for recommendation
SummaryThis paper addresses sequential recommendation with LLMs, where reinforcement learning shares one sequence-level reward across all tokens and relies on a single user's data. The proposed RPSD method uses on-policy self-distillation, in which a teacher with privileged knowledge re-evaluates the student's responses to assign token-specific credit. It also uses temporally aware optimal transport and diversity-aware selection to draw complementary behavioral evidence from similar users. The authors report state-of-the-art performance across multiple real-world datasets.
In briefUsing directed acyclic graphs to model how user preferences depend on each other in narrative-driven recommendation
SummaryThis paper proposes the Preference Dependency Graph Framework (PDG-F) for recommendations based on free-form user narratives in which preferences qualify one another. The system parses a request into a directed acyclic Preference Dependency Graph, retrieves candidate evidence from the open Web along its paths, and ranks results by topological path ranking. Experiments across three domains in English and Chinese show it outperforms baselines, open-domain retrieval agents, and frontier language models.
In briefWhere to Spend the Evaluation Budget in LLM-Driven Agent Scaffold Evolution
SummaryThis paper studies how to allocate limited evaluation budgets when LLM-driven systems evolve agent workflows through repeated candidate generation and testing. The authors find that sparse feedback can sustain productive evolution, while extra targeted evaluation improves the final choice of scaffold. They propose \method, which combines sparse search-time feedback with targeted final confirmation and reuses accumulated observations. Across four benchmarks and two evolution frameworks, it achieves higher mean performance than full evaluation in most settings while using about half as many evaluation cells.
In briefLow-degree polynomial lower bounds for clustering in low dimensions
SummaryThe paper studies the computational limits of clustering n observations from a mixture of K isotropic distributions in R^d when K is much larger than d. It constructs a mixture of Bernoulli product distributions whose centers come from a dual BCH code, so that moment tensors match those of the uniform distribution up to a given order. For suitable separations, low-degree polynomials require roughly K^2 d^{3/2} observations, and via Gaussianization the hardness extends to isotropic Gaussian mixtures.
WhyUsing dual BCH codes and bipartite-graph polynomial families, it improves the low-degree lower bound from dK² to K²d^{3/2} with full proof.
In briefGradCO: A Behavior-Aware Method for Text-Based Person Anomaly Search That Reduces Gradient Conflicts
SummaryThis paper introduces GradCO, a method for Text-based Person Anomaly Search, which finds a target pedestrian using descriptions of appearance and normal or anomalous behavior. It combines three components: Behavior-Aware Mixture-of-Experts, Conflict Offload Learning to reduce conflicting gradients, and Hierarchical Intent Alignment that aggregates expert similarities with query-conditioned weights. The authors report that GradCO significantly outperforms state-of-the-art methods on five TPAS/TIPR benchmarks.
In briefREGLAT at ArGuard: An Ensemble Model Adding Entity Recognition to Detect Harmful Arabic Content
SummaryThis paper proposes a two-stage method that classifies Arabic user prompts into seven harmful-content categories. GLiNER first extracts harm-related entities, which are appended to the original text and passed to a fine-tuned MARBERT classifier. On the ArGuard 2026 development set it reached a macro-F1 of 0.886 and accuracy of 0.897. On the blind test set it scored 0.793 accuracy and 0.735 macro-F1 for subtask 1, and 0.770 accuracy and 0.754 macro-F1 for subtask 2. The authors note that distinguishing related categories remains difficult.
Ahmed M. Fetouh, Mariam Labib, Nsrin Ashraf et al.
In briefGameVerse: A Minute-Scale Gameplay Video Dataset for Long-Horizon Interactive World Modeling
SummaryThis paper introduces GameVerse, a gameplay video dataset built for training interactive world models over minute-long horizons. It contains 69.7K clips totaling 1,655 hours and 349.6 million frames, including 1,358 hours of continuous two-minute sequences. Each clip has a camera trajectory that stays coherent across minutes and captions aligned to event boundaries. The authors report that training on GameVerse substantially extends the temporal horizon of interactive world models.
In briefYour Policy Is Secretly a Harness: How Agent Interfaces Shape Decisions
SummaryThe paper studies how language model agents depend on their harness, the prompts, tools, and control logic that present tasks to the model. Under KL-regularized control, the authors show that a trained policy's log odds reflect its harness's action preferences, and that deployed return equals the harness ceiling minus the policy's mismatch. In MiniWoB++, reversing observation field order drops a trained model's success from 85.0% to 19.2%.
WhyProposes a theory that policy log odds encode harness preferences under KL control, validated on MiniWoB++ and Terminal-Bench with clear empirical effects.
In briefSafety alignment of large language models in Hausa, an African language
SummaryThis paper examines how well large language models are aligned for safety in Hausa, a low-resource language spoken in West Africa. The authors build HausaSafeLM, a Hausa safety preference dataset of 10,000 prompts with human-corrected and machine-translated examples, and compare supervised fine-tuning with direct preference optimization. They find that refusal-rate metrics alone can misleadingly suggest a base model is safe, and that no configuration jointly maximizes safety, helpfulness, and fluency.
Valdini Douglace Lemofouet, Blessing Ngozi Uzor, Paula Chikaodinaka Anyanwu et al.
In briefTCM-DiagCommit: A Benchmark for Evidence-Based Traditional Chinese Medicine Syndrome Selection and Decision Commitment
SummaryThe paper introduces TCM-DiagCommit, a benchmark built from the public TCM-SD/ZY-BERT resource, converting 1,834 cases into 23,842 task rows over 148 syndrome labels. It uses candidate sets of varying difficulty and staged evidence, plus tongue/pulse perturbations and insufficient-evidence variants. The authors argue that final-answer accuracy alone is insufficient, and that answer selection, evidence use, and commitment should be measured separately.
In briefGrounding LLM agents in business processes using process-aware knowledge graphs
SummaryThe paper integrates OData metadata and BPMN reference processes into a typed property graph, with a hybrid retriever that follows process edges to find subsequent endpoints. On 42 enterprise services, it improves Recall@3 for queries with low lexical overlap with downstream entities. On 81 multi-step queries, an LLM agent completes 38.7% of tasks versus 10.3% for a chunk-based baseline.
In briefGraphRAGProv: Combining GraphRAG with How-Provenance Tracking
SummaryThe paper presents GraphRAGProv, an architecture that runs SPARQL queries over RDF knowledge graphs linked to their source text to compute answers and how-provenance polynomials. These polynomials record which statement occurrences were used and how they were combined. Each statement links back to its source, so users can check whether extracted statements reflect the text and whether the query combines them correctly. An interactive demo examines 400 cases from two multi-hop QA benchmarks and finds extraction and query errors even when the final answer is correct.
In briefHDVS: Semi-Supervised Segmentation with Heterogeneous Dual-Branch Voting
SummaryThis paper proposes HDVS, a semi-supervised semantic segmentation framework that learns from many unlabeled images with fewer human annotations. Built on cross-supervision, it adds a voting mechanism that combines pseudo-labels from heterogeneous branches to improve their reliability. A feature communication module introduces feature-level perturbations to increase prediction diversity between the two branches. The method is evaluated on PASCAL VOC 2012 and Cityscapes and is reported to outperform state-of-the-art approaches.
In briefDecision Models as Judges for Hallucination Detection
SummaryThis paper tests whether cheaper decision models, which output fixed verdicts with probabilities, can replace generative large language models as judges for detecting hallucinations. The authors built binary benchmarks for five hallucination types, with 9,852 items from six datasets, and compared five decision models with the generative judge gpt-6-sol. Performance varies by type: the best decision model matches the generative judge on context inconsistency but trails it on logical and instruction inconsistency.
In briefDoubly Robust Deep Kernel Tests: checks whether two interventions produce the same outcome distribution for high-dimensional structured data like images, text, and graphs.
SummaryThis paper proposes DR-KTE-Deep, a test of whether two interventions induce the same outcome distribution when outcomes are structured and high-dimensional, such as images, text, and graphs. It represents interventional distributions as mean embeddings on learned deep features, with the treatment effect expressed as a doubly robust representation, and optimises the features to maximise estimated test power. Empirically, the method achieves substantially higher power than fixed and bandwidth-optimised kernel baselines across image, text and graph outcomes while keeping Type-I error well calibrated.
Robin Leman, Bariscan Bozkurt, Sofiane Tanji et al.
In briefSubspace Clifford Networks: a neural network design that replaces global cyclic shifts with grouped subspace interactions
SummaryThis paper proposes Subspace Clifford Networks, which replace sampling sparse cyclic diagonals across the whole channel space with interactions inside small grouped subspaces. Features are projected into groups of two to four dimensions, with wedge coordinates and coordinate-wise products computed within each group and cross-group mixing done by learnable 1×1 projections. Experiments on CIFAR-100 report substantially fewer parameters and shorter training time, and the hierarchical variants achieve higher accuracy than the evaluated shift-based hierarchical baselines.
In briefSharpness-Aware Muon: a Muon optimizer variant that seeks flatter minima
SummaryMuon is a matrix-aware optimizer that uses momentum and orthogonalization but does not explicitly control worst-case loss in a neighborhood. The paper introduces MuSAM, which adds SAM-style perturbations to Muon's orthogonalized update, and MoMuSAM, which uses normalized momentum as the perturbation direction to avoid an extra gradient evaluation. The authors prove both methods reach the same asymptotic first-order stationarity order, with MoMuSAM giving a more stable stochastic linearization. Experiments on standard benchmarks show sharpness awareness consistently improves Muon.
In briefSpine: A method that organizes long coding-agent work into subtask trees for shorter, structured context
SummaryThis paper presents Spine, an interface and runtime that lets coding agents declare explicit subtask boundaries and records them in a persistent tree. Before each model call, unfinished subtasks keep their interactions while completed ones are replaced by their results. The authors integrate Spine into Codex and Pi and report score and cost changes on benchmarks such as SWE-Milestone and ProgramBench compared with the native versions.
In briefVision-Language-Action or World-Action? Just Action Models
SummaryThis paper examines whether robot control policies should inherit knowledge from vision-language models or visual generators. The authors separate three choices: backbone initialization, the objective for future observations, and whether future-prediction machinery is kept at inference. They propose the Just Action Model (JAM), a pretrained visual backbone with an action expert trained using an action loss and a foresight loss, where the foresight prediction head is discarded after training. JAM outperforms matched FastWAM and ImageWAM policies on LIBERO and LIBERO-Plus and performs similarly or slightly better on RoboTwin, and with cached foresight targets it reduces estimated training FLOPs per sample by 73% relative to FastWAM.
In briefTopoField: Technical Report for the ECCV 2026 SceneFun3D Challenge
SummaryThis report addresses functional segmentation, which locates small interactive parts such as knobs, buttons, and sockets in indoor point clouds and outputs confidence-ranked instance masks. TopoField, a deterministic decoder built on VOLT, fuses 13 test-time forward passes into a semantic consensus field, decodes instance candidates from its geometry, and ranks them class-aware. It achieves 28.286% AP, 41.383% AP50, and 47.558% AP25 on 561 validation instances across 30 scenes, and placed second in the ECCV 2026 SceneFun3D Challenge.
In briefZhengrong-Spatial-Aware 1.0: Advancing 3D Multimodal AI for Spatial Reasoning
SummaryThis report presents two end-to-end pipelines for adapting multimodal models to answer questions about 3D spatial scenes. One fine-tunes HiSpatial-3B, which adds metric-scale 3D point maps to RGB images, while the other applies QLoRA fine-tuning to Qwen3-VL. On the target spatial benchmark, HiSpatial's overall accuracy rose from 1.73% to 79.89%, and Qwen3-VL-8B and 32B improved from 20.74% to 45.27% and from 30.79% to 47.60%, respectively.
In briefRetrieve to Act: Motion Primitive Graph Retrieval for Stable Vision-Language-Action Execution
SummaryThis paper proposes R2A, a test-time framework that treats robustness of vision-language-action policies to out-of-distribution visual shifts as a geometric motion primitive retrieval problem. Successful training demonstrations are organized into an offline spatiotemporal motion primitive graph, and a graph neural network retrieves structurally compatible future primitives during execution. A learned gating mechanism combines these retrieved priors with the frozen policy's actions, and experiments on three VLA architectures report higher success rates in simulation and real-world tests.
Pham Tri Quang, Bao-Ngoc Dao, Minh Duc Nguyen et al.
In briefRetrospective macromolecular pool fraction mapping from routine clinical MRI using physics-guided self-supervised learning
SummaryThis paper presents a deep learning framework that estimates macromolecular pool fraction (MPF), a biomarker sensitive to myelin, directly from conventional clinical brain MRI sequences (T1-weighted, T2-weighted, and FLAIR) without paired quantitative reference scans. The method embeds a two-pool magnetization transfer signal model and uses adaptive instance normalization to handle heterogeneous clinical protocols. Trained on 7,705 3 T sessions, it reported mean MPF of 0.182 in white matter and 0.144 in gray matter, with lower MPF in demyelinating lesions than in normal-appearing white matter.
Jelmer van Lune, Miha Fuderer, Oscar van der Heide et al.
In briefShortcut amplification in video RLVR: a perceptual uncertainty view
SummaryThis paper studies reinforcement learning with verifiable rewards (RLVR) for improving video reasoning in large multimodal models. The authors find that outcome-level optimization makes models increasingly favor text-prior answers even when visual evidence does not support them. They propose Bifocal Self-distillation Policy Optimization (BSPO), which uses reasoning-sensitive tokens and shortcut-aware visual perturbations to regulate policy updates. Experiments reportedly show improved video reasoning performance and stronger visual faithfulness.
In briefSimplexLoRA: Adjusting LoRA adapter ranks during fine-tuning via simplex-based importance weights
SummaryThis paper proposes SimplexLoRA, a variant of Low-Rank Adaptation (LoRA) that adjusts the rank of each adapter during fine-tuning based on learned importance. After a short warm-up, per-adapter importance weights optimized on an n-scaled simplex are frozen, allocating capacity without increasing the overall parameter budget or inference cost. Evaluated on 12 language understanding, summarization, and reasoning tasks with DeBERTaV3, BART-large, Llama 3.1 8B, and Qwen3-8B, it matches or outperforms standard LoRA and recent adaptive baselines, with the largest gains at low ranks.
Igor Shalygin, Grigorii Davydenko, Andrey Veprikov et al.
In brief3D-SceneQ: Improving 3D Language Models with Query-Guided Pruning and 2D Feature Fusion
SummaryThis paper presents 3D-SceneQ, a 3D large language model for understanding 3D scenes. Its Query-Guided Adaptive Pruning module selects objects relevant to the user's instruction to reduce irrelevant noise, while its Multi-modal Object-level Feature Enhancement module adds features from pre-trained 2D models to the 3D representation. The authors report that the method outperforms current state-of-the-art approaches on several benchmarks, including +5.7 CIDEr on Scan2Cap and +17.0 EM@1 on SQA3D.
In briefDo Generalist Skill Optimizers Really Help AI Agents? A Bipartite Attribution Rethink
SummaryThe paper introduces BiPA, a method for evaluating skill optimizers that rewrite an AI agent's natural-language procedures to improve performance on long-horizon tasks. Across three optimizers, three benchmarks, and two models, learned skills succeed about as often as their hand-written seeds, and development-set gains vanish on the test set. The optimized skills also increase input tokens by about 16%, largely because the added rules cause extra conversation turns. The authors propose four principles, including accepting only edits whose gains exceed what selection alone would report.
WhyIntroduces bipartite attribution and selection baselines showing dev-set gains vanish on test and inputs grow about 16%, with multi-benchmark, multi-model experiments.
In briefHow much can choosing the best of N candidate actions improve a world action model, and how to realize it in closed loop
SummaryThe paper studies how much a world action model can gain at test time by imagining several candidate action sequences and executing the best one. It introduces a denoised counterfactual protocol that measures the selectable headroom a selector can actually realize, and uses these measurements to build a selector that deviates from the policy only when the model predicts failure. On held-out episodes of training tasks, closed-loop success rises from 76.9% to 81.6%, and on unseen tasks from 62.6% to 69.8%.
In briefAction-value calibration through shared continuations for long-horizon agent RL
SummaryThis paper addresses the difficulty of assigning credit to individual actions in reinforcement learning for long-horizon agents. It proposes ConCal (Continuation Calibration), which matches visits from the same task with identical state-action-state transitions, combines shared continuation evidence, and propagates calibrated values backward to refine earlier action-value estimates without extra rollouts. With Qwen2.5-1.5B-Instruct, ConCal improves success rates over GiGPO by 3.39 points on ALFWorld and 5.86 points on WebShop.
In briefVGIQA: Fine-grained image quality scoring that pinpoints local flaws using visual grounding rewards
SummaryThis paper presents VGIQA, an image quality assessment model that evaluates both overall image quality and localized quality within an image. It is trained in two stages, a cold-start phase followed by reinforcement learning, guided by reward functions designed for local quality assessment. The authors also introduce VGSR-2K, a new dataset for local quality evaluation, on which VGIQA outperforms open-source models such as Qwen2.5-VL-72B-Instruct and Qwen3-VL-32B-Instruct.
In briefINTR++: A More Accurate, Interpretable, and Adaptable Interpretable Transformer
SummaryThis paper proposes INTR++, which extends Prompt-CAM with a lightweight Prompt Adapter Module (PAM) that enriches prompts with domain-specific knowledge. The authors aim to keep interpretability consistent across domains with differing image characteristics, such as medical, microscopic, and remote sensing imagery. INTR++ adds only a small number of parameters compared with Prompt-CAM. The paper reports consistent improvements across animal, plant, medical, microscopic, remote sensing, and object-centric datasets.
Dipanjyoti Paul, Arpita Chowdhury, Maneet Chatterjee et al.
In briefMLE-World: Predicting Machine Learning Experiment Results with a Learned World Model
SummaryAI research agents normally must run each candidate program to train and evaluate models, which limits how many options they can explore. This paper introduces MLE-World, a model fine-tuned from Qwen3.5-4B that predicts a program's terminal output, including error traces, training logs, and validation scores, from its code and data description without executing it, along with a confidence score. Five test-time scaling policies guided by these predictions are built, and candidates are actually run only when confidence is low. On 17 MLE-bench Lite tasks, MLE-World matches real terminal output on 77.9% of held-out programs, compared with 41.1% for much larger prompted models.
In briefRényi-Regularized Flow Variational Inference for Bayesian Calibration
SummaryThis paper studies normalizing-flow variational inference for Bayesian model calibration, where posteriors with several modes separated by low-probability regions can cause mode collapse. Using a two-parameter twin experiment, the authors add a regularization term based on Rényi divergence that encourages the approximate posterior to cover the prior support. The near-collapse fraction drops from 50% to 25% in this experiment. The authors note that the regularization alters the Bayesian variational objective, introduces a trade-off in approximation accuracy, and is not a general solution to mode collapse.
In briefLiveSearch: A Live Benchmark for Financial Search Agents Using Executable Verifier Code
SummaryLiveSearch is a live benchmark for financial search agents that stores each question as executable code calling a professional financial data service. The gold answer is recomputed at evaluation time and re-executed daily without human or language model involvement, so memorized answers become outdated. Six tool-calling models reached 70.9% to 76.7% accuracy under general web search. Over three consecutive days, 70% of questions changed their gold value.
In briefBenchmark testing whether retrievers handle evidence that contradicts their prior knowledge
SummaryThis paper introduces CoverageBench-Negate, a benchmark that tests whether retrieval systems find corpus evidence when it contradicts their parametric knowledge. The authors create negated versions of each nugget and edit supporting documents to match the negation while preserving context, yielding about 63,000 original-edited evidence pairs across six datasets. In baseline experiments, BM25 distinguishes original from edited evidence much better than dense retrievers such as Qwen3 and Jina, which perform near chance.
Pasha Abrishamchian, Eugene Yang, Andrew Yates et al.
In briefWorkFactory: Converting Professional Intents into Executable Tasks to Scale Agent Training
SummaryThe paper presents WorkFactory, a framework that builds an information support graph from professional intents and uses probe executions to find and repair missing task information. It constructs 112,686 tasks across 612 occupations and fine-tunes Qwen models on the resulting teacher trajectories, producing the WorkPilot family. The authors report that all three models improve over their base models on four professional-work benchmarks.
WhyInformation support graph and probe-based repair for constructing professional tasks offers substantive improvement; multi-benchmark, multi-scale experiments with ablations give solid evidence, with m
In briefPhysics-Guided Flow Matching: A Generative AI Model for Forecasting Global Electron Content
SummaryThis paper proposes a physics-guided generative AI framework for day-ahead forecasting of global total electron content (TEC), which matters for space weather monitoring and satellite communication and navigation. The model builds on flow matching and incorporates ionospheric dynamics described by a partial differential equation into the transformation process, combining them with statistical patterns learned from data and adaptively weighting each location. Trained and evaluated on one year of global TEC data from the TIEGCM model, it outperforms representative benchmarks at all lead times and across latitudes.
In briefWhat survives when GPT-2 is distilled into a smaller model: a mechanistic interpretability study.
SummaryThis paper compares GPT-2 with its six-layer distilled student, DistilGPT-2, using CKA, logit lens, attention analysis, sparse autoencoders, and activation patching. It examines two circuits: Indirect Object Identification, which collapses after distillation, and Greater-Than, which largely survives by remapping a layer-9 head to layer 4. The authors conclude that circuit compressibility determines mechanistic survival under distillation.
In briefFastBEVFusion: A Simple Fast Fusion Method for 3D Object Detection at 50 FPS
SummaryThe paper proposes FastBEVFusion, a simple, fast, and robust system that fuses LiDAR and camera data for 3D object detection. It projects modality-specific features into semi-aligned BEV spaces without depending on depth estimation, and combines them with a lightweight, dynamic cross-attention fusion module. According to the paper, experiments on nuScenes show it outperforms existing real-time systems, runs at over 50 FPS without deployment libraries, and is robust to calibration errors.
Tom Sanitz, Emre G\"ulsoylu, Simone Frintrop et al.
In briefGCSR-OCC: Geometry Calibration and Semantic Routing for Open-Vocabulary Occupancy Prediction
SummaryThis paper presents GCSR-OCC, a method for predicting indoor 3D occupancy from a single RGB image, recognizing objects beyond predefined categories. It calibrates Gaussian-based geometry to suppress scale inflation and excessive overlap, and uses text-conditioned semantic routing to reduce semantic mixing within Gaussians. On Occ-ScanNet, without 3D semantic annotations, it reports 63.31% geometric IoU and 27.44% semantic mIoU, and lowers geometric aggregation latency from 19.76 ms to 12.88 ms.
In briefVariational Optimal Control for Diffusion Guidance
SummaryThis paper proposes a variational control framework that links sampling-time guidance and reward-based model adaptation through a shared Jacobian pullback. The proposed Variational Optimal Guidance (VOG) acts in state space through the denoising Jacobian, while Distilled Variational Optimal Control (DVOC) amortizes locally improved endpoints into model parameters. On ImageNet, the authors report that VOG lowers DiT-XL/2 FID from 2.27 to 2.04 at 250 steps and from 7.58 to 4.44 at 25 steps.
In briefFisher-Shaped EWC: a replay-free continual learning method that shapes solutions for future retention
SummaryThis paper proposes Fisher-Shaped EWC (FS-EWC), which extends the Fisher information matrix's role from protecting parameters to forming solutions that are less sensitive to future updates. During task learning it minimizes the worst-case empirical Fisher trace within a neighborhood, and at task completion it builds the retained solution by trajectory-based model merging. Experiments across benchmarks show improvements over EWC, with a better stability-plasticity balance and greater tolerance to parameter drift.
In briefGenerative replay for training Hopfield networks without forgetting past tasks
SummaryThis paper addresses catastrophic forgetting in Hopfield networks trained with Equilibrium Propagation during continual learning. It proposes Energy-Based Generative Replay (EB-Gen), where a frozen copy of the network generates pseudo-samples for rehearsal, so no external data buffer is needed. Evaluated on Split MNIST and Split CIFAR10 in the class-incremental setting, the method reports 75% and 64% average accuracy, respectively.
Victor Cook (LIRMM | ADAC), Gilles Sassatelli (LIRMM | ADAC), Thierry Gil (LIRMM) et al.
In briefMixSIS: Instance segmentation that mixes annotation types
SummaryThis paper proposes Mixed-Supervised Instance Segmentation (MixSIS), a task where a dataset can combine image, point, box, and mask annotations. The authors introduce AnnotationsMixer (AnnsMixer) as a baseline that generates pseudo mask annotations from all these annotation levels, even within a single image. The paper reports that AnnsMixer largely outperforms state-of-the-art methods on weakly-supervised and MixSIS tasks, and shows in three case studies how it can reduce annotation time or improve results.
WhyIntroduces a new mixed-annotation instance segmentation task and baseline, largely outperforming prior methods on WSIS and MixSIS with three supporting case studies.
Christian Wilms, Emre G\"ulsoylu, Sven Schreiber et al.
In briefWhat the Primary Lund Plane Omits: A Diagnosis of a Jet-Tagger Representation Gap
SummaryThis paper uses model stitching to test whether the performance gap between Lund-plane jet taggers and Particle Transformer (ParT) comes from a weaker representation or a weaker downstream computation. Using only the primary declustering branch, the representation shows lower background rejection than ParT, indicating a representational deficit. Rebuilding the representation on the full Cambridge/Aachen declustering tree closes 72% of the gap.
In briefPhysics-guided diffusion model for exposure interpolation in HDR imaging
SummaryThis paper addresses brightness order reversal artifacts that arise when fusing two large-exposure-ratio standard dynamic range images captured in a single shot. It proposes a physics-guided diffusion model based exposure interpolation algorithm, where a physics-driven component produces an initial interpolated image and a diffusion model component generates residual high-frequency details. Experimental results are reported to validate the proposed algorithm.
In briefMRI-to-CT Spinal Image Translation Using Anatomy-Guided Pseudo-Pairing and Adversarial Diffusion
SummaryThis paper proposes a three-stage framework that synthesizes CT-like images from spinal MRI to support surgical planning without extra radiation. It matches unpaired MRI and CT slices using vertebral segmentations and rigid registration to create pseudo-paired training data, then trains an adversarial diffusion model for translation. A 3D refinement network improves continuity between slices. On 83 MRI volumes, the approach raised the binary 3D vertebral Dice score from 0.666 to 0.739 compared with diffusion alone.
In briefMCP-Net: A Coronary Perception Network for Robotic Guidewire Guidance
SummaryThis paper presents a simulation-driven framework for robotic guidewire navigation centered on MCP-Net, a multi-task network. From a single two-dimensional fluoroscopic frame, it jointly performs coronary vessel segmentation, guidewire-tip localization, and stenosis localization. The predictions are converted into a structured vascular graph with centerlines, bifurcations, endpoints, and branch angles, from which a roadmap toward the target lesion is generated.
Sahar Nasirihaghighi, Tianyuan Wang, Kenan Niu et al.
In briefHidden Internal Damage from Pruning Transformer Language Models
SummaryThis paper studies how weight pruning affects GPT-2 by testing combinations of importance criteria, component types, and sparsity levels against the indirect object identification (IOI) circuit. Using activation patching and related causal methods, it identifies two distinct failure modes: magnitude pruning of attention heads destroys the circuit, leaving the model near chance, while magnitude pruning of 30% of MLP neurons keeps the circuit intact but flips the sign of Name Mover contributions, yielding 17.7% accuracy. The authors note that perplexity rates the neuron-pruned model as less damaged than random head pruning. They conclude that output metrics and structural checks can miss such mechanistic damage, and recommend circuit-level causal analysis for evaluating compressed language models.
In briefSFE-O: A Knowledge Ontology for Supercritical CO2 Biomass Extraction
SummaryThis paper presents SFE-O, a modular OWL 2 ontology that describes supercritical CO2 biomass extraction processes. It links biomass, process, and product dimensions and explicitly represents process topology, parameter provenance, mass-transfer phenomena, and constant and falling extraction rate mechanisms. The ontology is evaluated through a pilot-scale extraction batch example, a reasoner-based consistency check, and a competency-question SPARQL query.
In briefProspector: a compact model that learns to traverse knowledge graphs for scientific reasoning
SummaryThe paper presents Prospector, a compact model trained to reason over a domain knowledge graph by traversing it step by step. It calls a small set of tools one at a time to build context and answers with an evidence chain that can be checked edge by edge against the retrieved graph content. Trained only on a semiconductor-manufacturing graph, it is reported to transfer to unseen graphs from three other fields.
In briefMultibranch ensemble head for efficient cucumber growth-stage classification with vision foundation models
SummaryThis paper proposes a regularizing multibranch ensemble head for vision foundation models such as DINOv2, which samples one branch stochastically during training and averages both branches' probabilities at inference. On the NALEF Food Safety dataset of five cucumber growth stages, the DINOv2 ViT-S model with this head reaches a Macro F1 of 0.978 and Cohen's Kappa of 0.971. It requires 12.2 GFlops and 22.3M parameters, which the authors state makes it deployable on smartphones and edge devices.
In briefSelf-Regularized Flow Language Models: a method that uses the denoising trajectory to improve continuous diffusion language model training
SummaryThe paper introduces SERF, a self-supervised framework that improves continuous diffusion language models through noisy-to-clean self-distillation. An EMA teacher receives a cleaner version of the student's noisy input along the same interpolation path and provides additional supervision. Two variants, SERF-L and SERF-P, are tested across three continuous language-model paradigms, showing gains in training efficiency, generation quality, and few-step generation.
In briefRestoring Diversity in Distilled Diffusion Models via a First-Step Bias
SummaryThis paper addresses the loss of output diversity in diffusion models distilled to generate images in a few steps. The authors find that most of this diversity loss occurs in the first, high-noise denoising step. They propose ReSeed, which adds a seed-response bias adapter only to the first student step, aligning its velocity with the teacher's, while all other weights and steps stay frozen. Experiments on several text-to-image and text-to-video models show improved diversity with competitive visual quality and very few additional trainable parameters.
In briefGenerating Text in Low-Resource Languages with Human-Curated Vocabularies
SummaryThis paper examines whether large language models can generate coherent text in extremely low-resource and unseen languages by using human-curated vocabularies. The framework combines these vocabularies with a small number of human-written exemplars that supply syntactic signals, sampling candidate words from a Zipfian distribution. Generated text is evaluated for lexical diversity, off-target ratio, and human fluency judgments. Reasoning-enabled models produced fluent sentence-level text and showed promising paragraph-level performance.
Michael Christlambert Sinanta (1), Muhammad Dehan Al Kautsar (2), Amanda Nurul Izzah (1) et al.
In briefHardware-matched spiking networks for real-time Purkinje cell spike sorting
SummaryThis paper presents two compact spiking neural networks that separate Purkinje cell simple spikes from complex spikes on untethered recording headstages. ConvSNN targets a microcontroller's matrix vector processor, while StreamSNN targets a dedicated SKY130 circuit. On cells unseen during training, complex-spike F1 scores are 0.873 for ConvSNN and 0.837 for StreamSNN, both with eight-bit weights.
Arooba Farhan, Afsah Syed, Muhammad Fahd Ibrahim Bhatti et al.
In briefSplatScout: Query-driven part-level referring segmentation in 3D Gaussian scenes
SummaryExisting referring segmentation methods for Gaussian scenes mainly target whole objects, choosing among fixed candidates produced by a segmenter. SplatScout is a closed-loop approach in which a vision-language model actively selects views, directs segmentation, and inspects rendered 3D results to refine its predictions. The authors introduce PartRef-ScanNet++, a part-level benchmark with 261 manually verified queries across 12 ScanNet++ scenes, where SplatScout reaches 0.538 mIoU. It also scores higher than existing methods on four other referring benchmarks.
In briefTriDifNet: A Unified Diffusion Network for Joint Infrared-Visible Alignment, Fusion, and Pedestrian Detection
SummaryThis paper proposes TriDifNet, a diffusion-based network that jointly performs registration, fusion, and detection on misaligned infrared and visible image pairs. It uses three diffusion modules for alignment, fusion, and detection, linked by a closed-loop optimization scheme. Experiments on the KAIST multispectral pedestrian benchmark report superior detection performance under misaligned RGB-T settings.
In briefLearning to Look Beyond Screens: Transferable Visual Observation from GUI Grounding
SummaryThis paper introduces LBS (Look Beyond Screens), a vision-language model first initialized with supervised training on teacher-generated GUI grounding trajectories. It is then trained with reinforcement learning to make multi-turn crop-and-inspect decisions. LBS consistently outperforms its baselines on Qwen3.5-35B-A3B and Qwen3.8-27B under fixed inference settings. The gains also transfer to PerceptionBench for both backbones and to HR-Bench and ChartQAPro for Qwen3.5-35B-A3B.
In briefPaper overview: why Direct Preference Optimization yields sparse weight updates
SummaryThis paper theoretically explains why Direct Preference Optimization (DPO) produces sparse parameter updates. Analyzing DPO under a neural tangent kernel linearization, the authors show its logistic loss induces an implicit maximum-margin bias and introduce the notion of dual sparsity. Experiments on GPT-2 and several 7–9B instruction-tuned models confirm both sparsity sources, and on GPT-2 DPO reduces cross-task interference by up to 35 times compared with supervised fine-tuning.
In briefOpenGray: A Benchmark for Assessing Decisions of Radiotherapy Planning AI Agents
SummaryThis paper introduces OpenGray, a benchmark for testing whether large language model-based radiotherapy planning agents make correct decisions, such as when to submit a plan or request review. It adds tasks with known correct answers to the OpenKBP-Opt head-and-neck dataset, including contradictory dose goals and a feasible control task. The authors find that systems with favorable overall scores still made specific decision errors, particularly on contradictory cases. They conclude that planning agents should be evaluated on their decisions as well as dose and should have an explicit action for stopping and requesting review.
In briefSYMBA: Self-supervised brain MRI autoencoder for detecting neurodegenerative diseases
SummaryThis paper introduces SYMBA, a self-supervised pretraining framework for structural brain MRI that uses anatomically constrained masking, an MS-SSIM reconstruction objective, and a curriculum moving from local detail to global morphology. It was pretrained on 6,953 T1-weighted scans from nine cohorts and evaluated with frozen linear probing on two unseen datasets. SYMBA reached 0.930 AUROC for Alzheimer's disease classification and 0.772 for mild cognitive impairment conversion, and about 0.90 AUROC for AD/CN classification using only six labeled subjects.
Nuria Balb\'as-Esteban, Albert Belenguer-Llorens, Elaheh Moradi et al.
In briefCorrectness Filtering as Selection Bias in Chain-of-Thought Faithfulness Evaluation
SummaryThis paper argues that restricting chain-of-thought faithfulness evaluation to correctly answered instances changes the population being measured. Using a seven-variable structural causal model, the authors derive the exact observed gap and review 18 papers' protocols. Positive selection gaps appear across four arithmetic deployments, with the primary run showing a 7.4 percentage-point gap. The authors note that when counterfactual compliance is observable for every instance, all traces can be evaluated directly without correctness filtering.
In briefHierarchical low-rank reparameterization to compress trainable parameters in parameter-efficient fine-tuning.
SummaryThis paper addresses redundancy in the trainable parameters of parameter-efficient fine-tuning (PEFT) methods. It proposes a hierarchical low-rank reparameterization that reorganizes these parameters into an approximately square matrix and applies low-rank decomposition recursively to form a multi-level, progressively compressed structure. The method is combined with LoRA, DoRA, MoSA, and RaSA and evaluated on commonsense reasoning with LLaMA-3.1-8B and Qwen2.5-7B, and on GSM8K math reasoning with LLaMA-3.1-8B. The authors report better performance than baselines under comparable parameter budgets, and competitive or higher average accuracy across eight commonsense benchmarks when parameter counts are compressed by about an order of magnitude.
In briefEnergyCal: Reducing Loss Spikes in Muon-Family Optimizers for Sparse MoE Training
SummaryThe paper introduces EnergyCal, a lightweight rule that scales down routed-expert updates based on the normalized squared Frobenius norm of their finite-step Newton-Schulz directions. It leaves the spectral transform unchanged and adds no persistent optimizer state. On a 4B-parameter MoE model, EnergyCal-MuonH cuts mean thresholded loss-spike burden by 99.51% and final geometric-mean perplexity by 4.06% compared with a lower-learning-rate MuonH recipe, at a modest throughput cost.
In briefLimits of the Delta Learning Hypothesis: Generator Biases in Single-Pair DPO Preference Data
SummaryThe paper examines Direct Preference Optimization (DPO) when preference data comes from a single pair of models, treating the larger model's response as preferred. The authors show theoretically and empirically that fixed model pairs introduce generator-dependent biases that predict the preference label regardless of response quality. As a result, DPO trained on single-pair data consistently underperforms DPO trained on multi-model preference data.
Giuseppe Curci, Stefano Simonazzi, Andrea Molinari et al.
In briefQuantized edge inference of a self-supervised SAR-optical encoder on ordinary x86 computers
SummaryThis paper deploys a pretrained SAR-optical masked-autoencoder encoder on three commodity x86 machines and compares FP32, FP16 and INT8 precision for latency, energy and linear-probe accuracy. INT8 quantization leaves accuracy essentially unchanged. It lowers CPU latency on every machine and reduces CPU energy where measured.
In briefLearning to Code the World: Evolving Harnesses for Compositional 3D Reconstruction
SummaryThis paper presents SceneEvolve, a framework that represents 3D scenes as executable code and improves reconstruction by evolving the harness around a small open-source agent. A stronger teacher agent iteratively refines the harness using feedback from development-set reconstructions, while model parameters, action semantics, and the execution backend stay fixed. On two in-domain datasets and one out-of-domain dataset, the method improves all four reconstruction metrics for the Qwen3.8-27B model and cuts mean inference time by 47.0% relative to the base model.
In briefAI-assisted discovery of CNOT decompositions for quantum permutations on constrained qubit topologies
SummaryLimited qubit connectivity in quantum computers adds overhead when implementing logical operations. Using the AlphaCNOT tool, the authors find a decomposition for remote transpositions that uses at most 6d-4 CNOT gates for two qubits at graph distance d, compared with 6d-3 for the SWAP-based construction. They formalize this as Global Transpositions, a deterministic synthesis method for whole permutations, and identify an intermediate CNOT fusion mechanism.
Daniele Lizzio Bosco, Jacopo Cossio, Simone Faro et al.
In briefGuided-RLVR: Guided Exploration in Reinforcement Learning with Verifiable Rewards
SummaryThis paper proposes Guided-RLVR, an extension of reinforcement learning with verifiable rewards (RLVR) for improving LLM reasoning. Standard RLVR rollouts tend to collapse onto a narrow set of reasoning trajectories. The method adds a lightweight controller that injects semantically meaningful steering tokens into the sampling distribution, so the model approaches problems from multiple reasoning perspectives. Experiments on math reasoning and code generation benchmarks show it consistently outperforms existing RLVR methods and produces more diverse reasoning trajectories.
In briefZooming without Zooming v2: Connecting Visual Evidence for More Generalizable Multimodal Perception
SummaryThe paper introduces EvidenceConnect, a framework that synthesizes training questions from zoomed views of related regions within one image, requiring comparison, aggregation, or chaining of visual evidence. Reference answers are kept only when independent VLMs agree, and region-ablation tests remove questions whose answers survive region removal. Base VLMs trained on this data with RLVR, yielding ZwZ2, show gains in fine-grained and spatial perception and in several other tasks, even though the data were not tailored to any benchmark.
In briefGraftRay: Physically Consistent Mesh Insertion into Gaussian Splatting Scenes via Path Tracing
SummaryGraftRay is a method for inserting mesh objects into indoor scenes represented by 2D or 3D Gaussian Splatting while keeping lighting physically consistent. It estimates scene materials and surface normals using diffusion-based priors and models spatially extended light sources in closed form. A hybrid renderer path-traces the scene Gaussians together with the inserted meshes to produce view-consistent shading, interreflections, and soft shadows. On the FIPT benchmark, it outperforms baselines such as ComGS and EAG-PT in rendering fidelity and shadow accuracy while remaining highly efficient.
In briefA study comparing explainable deep learning models for breast cancer diagnosis from medical images
SummaryThe paper benchmarks ten ImageNet-pretrained deep learning models for classifying benign versus malignant breast tumors on the BreakHis histopathology dataset, using a patient-level split. Predictions are explained with Grad-CAM, LIME, Integrated Gradients, and LRP. DenseNet121 performed best, reaching 99.4% test accuracy and an AUC-ROC of 0.999, while Transformer-based models trailed the CNNs.
Sami Alnawali, Muhammad Atif Saeed, Abdul Hadi et al.
In briefComparing Regression Models for Predicting Crop Yield
SummaryThis study compares five regression models for predicting crop yield from environmental, crop, and farm management variables. The dataset contained about one million observations, and models were tested with 80:20 and 70:30 train-test splits. Performance was measured by R-squared, mean squared error, and root mean squared error. Linear-type models performed competitively and similarly, while random forest showed a large gap between training and testing performance.
In briefGround-to-Satellite Knowledge Distillation for Crop Classification Under Shift
SummaryThis paper studies whether ground-level field photographs can improve crop classification based on pretrained satellite embeddings. A teacher model using frozen CLIP features supervises classifiers trained on AlphaEarth and TESSERA representations, with photographs used only during training. On a paired dataset of 17 crop classes, distillation improves over satellite-only classification, with gains of up to +5.5 macro-F1 points under temporal shift and larger gains in geographic holdouts.
Miro Miranda, Rocco Sedona, Valerio Marsocci et al.
In briefV-LUMEN: Visual Lookup Memory for Scaling Embeddings in Vision-Language Models
SummaryThis paper introduces V-LUMEN, a visual lookup memory for vision-language models that adds external memory instead of increasing model size. It discretizes visual features with a frozen codebook, builds spatially aware memory keys by hashing, modulates them with text-conditioned relevance signals, and fuses retrieved embeddings into the backbone through a gated residual pathway. On five reasoning-intensive benchmarks it improves a frozen Qwen3-VL-2B backbone and transfers zero-shot to three benchmarks whose images are disjoint from the training data.
WhyApplies lookup-based external memory to VLMs with spatial hashing and gated fusion, validated on five benchmarks with moderate experimental scope.
In briefTesting Priority Dispatch for Premium Members in Shared On-demand Taxis
SummaryThis paper examines whether premium membership can be added to an on-demand taxi-sharing service at the dispatch level, using a sequential optimal insertion method. Two approaches are proposed: shortening deadlines for premium users, and shortening deadlines while using dedicated premium vehicles. Both are compared with a baseline dispatch method through SUMO simulations under steady, peak-period, and urban-concentrated demand. The results indicate that the suitable priority mechanism depends on operating conditions and the premium ratio, so switching methods by condition may benefit premium users while limiting the burden on non-premium users.
In briefFrom Ethernet to Wi-Fi: A Method for Choosing Features to Classify Network Traffic
SummaryTraffic classifiers trained on Ethernet data can lose accuracy on Wi-Fi traffic. The paper proposes a five-stage procedure for selecting and constructing flow features so that models trained only on Ethernet data can classify Wi-Fi flows. Ethernet and Wi-Fi traffic are simulated in ns-3, with Wi-Fi using the Best Effort access category. Labelled Wi-Fi data from six development scenarios guide feature design, and a separate scenario is reserved for final testing.
Francisco Rau, Petia Georgieva, Carlos Herranz et al.
In briefCompote: A Collapse-Resistant Multi-Hypothesis Method for Sequential Recommendation
SummaryThis paper addresses the ambiguity of next-item prediction in sequential recommendation by proposing COMPOTE, a multi-hypothesis transformer trained with an annealed objective that shifts from uniform supervision to winner-takes-all assignment, so hypotheses specialize in distinct catalogue regions. Using item IDs alone, it outperforms strong ID-based, generative, and multi-interest baselines on three Amazon Reviews 2023 benchmarks. Matched-capacity ablations attribute the gains to the annealed objective rather than extra capacity, ensembling, or routing.
Louis Castres Saint Martin, Clement Calauzenes, Alain Rakotomamonjy et al.
In briefMeta-models: models that predict how other models learn
SummaryThis paper studies meta-models, which predict the training behavior of other models. Treating language-model training records as numeric data, the authors find that the off-the-shelf tabular in-context learner TabPFN matches or outperforms a fine-tuned LLM meta-model in several evaluated settings, while needing only seconds of fit-and-predict time. Augmenting TabPFN with Kolmogorov-Arnold Networks (KANs) further improves accuracy on selected tasks, and the authors pre-registered a forecast for a large public training run.
