In briefSmartSalesBench: A Multi-Turn Benchmark Testing AI Sales Agents Against Resistant Buyers
SummaryThis paper introduces SmartSalesBench, a multi-turn conversational benchmark for testing large language model sales agents. Unlike existing benchmarks that assume cooperative users, it uses buyer personas programmed to resist and challenge the agent across conversations of up to twenty turns. The benchmark covers five agent profiles, twelve buyer personas, fifteen smartphones, and 13,680 sessions, with nine evaluation metrics in three tiers.
P N Bhargav Teja, Velpula Lokesh Reddy, Madhu Basumata et al.
In briefDo Fast Calibrated Models Transfer to Indonesian Regional Languages?
SummaryThis paper evaluates two calibrated "System One" models, which output a label and probability in one pass, on Indonesian regional languages using NusaX and NusaTranslation sentiment and emotion test sets. Presenting the Indonesian translation raises accuracy substantially, but the models' ability to rank their own errors degrades across all three datasets. The authors also compare a confidence-gated cascade and several commercial LLMs on accuracy and API cost.
Romi Fadli, Rusdah, Wahyu Ramadani Putri Utami et al.
In briefMinions: A Dynamic Distributed Virtual GPU Cloud Using Idle Apple Silicon Macs
SummaryCentralized GPU infrastructure for machine learning inference is often costly, inflexible, and slow to access under resource limits. This paper introduces Minions, a system that pools idle GPUs on Apple Silicon macOS devices into a distributed virtual GPU cloud for large-scale inference workloads. It offers flexible task execution, fault tolerance, and efficient resource use without buying new GPUs. The authors report that idle macOS devices can form a practical distributed GPU inference platform where centralized GPUs are scarce.
In briefMotionCoach: Multimodal Guidance for Precise Professional Motion Editing in Video
SummaryThis paper addresses the problem that video editing models often fail to generate professional motions, such as sports, martial arts, or performance techniques, when given only motion names. It introduces MotionCoach, a trained multimodal agent that uses multi-turn tools and self-verification to produce a grounded prompt and reference clip that inject professional motion knowledge into existing video editing models without retraining them. The authors also present VPMEBench, a benchmark of 314 professional motion editing tasks, and report that overall scores rose from 55.1 to 65.6 on Seedance 2.0 and from 51.2 to 64.5 on MiniMax-H3.
In briefMVGraph: Testing how vision-language models combine evidence across several linked visual graphs
SummaryThis paper introduces MVGraph, a benchmark for testing large vision-language models on reasoning where the needed evidence is spread across several related visual graphs. It contains 8k instances across 8 graph-theoretic tasks, plus 29 diagnostic tasks with 64k instances that probe perception, alignment, composition, and reasoning separately. Evaluating 15 models, the authors find that aligning edges across views and composing relations are the main bottlenecks.
In briefDataCupid: Stable matching repair for training data acquisition with delayed feedback
SummaryThis paper studies how to repair a matching after utility changes by reusing valid observations while respecting capacity limits and handling out-of-order, possibly delayed evaluations. The authors propose Residual-load repair (RLR) and a monitored variant, RLR-P, with theoretical bounds on issue rounds, delay, and evidence lifetime. Paired simulations and a measured-feedback replay show RLR-P has lower mean capped update-certification time than a width-based baseline, at higher oracle CPU cost.
In briefTraining-free video token compression guided by codec information for video LLMs
SummaryVideo large language models need long sequences of visual tokens, so compressing them efficiently matters. This paper proposes a training-free method that uses information already present in compressed video, such as reference-frame dependencies, motion vectors, and residuals, as a low-cost search guide. The method works without modifying the base model. Experiments on three VLLMs and four benchmarks show performance comparable to the FlashVID baseline, sometimes better, with far fewer feature comparisons.
In briefJoyInside-VLM-Agent: A multimodal AI agent for robots and smart devices that learns to act under product constraints
SummaryThis paper presents JoyInside-VLM-Agent, a multimodal agent that turns images and user requests into executable actions for robots and smart hardware. It is trained with a closed-loop sandbox that generates interaction trajectories including goal changes and failure recovery, followed by supervised fine-tuning, multi-teacher distillation, and reinforcement learning with DPER. The authors report a BFCL v4 overall score of 77.44, mixed results on other public benchmarks, and leading performance on their own JoyInside-Devices and JoyInside-Robotics sets. Two real-world quadruped robot case studies are also presented.
In briefMathVis: Can AI Solve Math by Drawing? A Benchmark for Math Diagrams
SummaryThis paper introduces MathVis, a benchmark of 512 tasks that tests whether large language models can visualize advanced math problems and reason with diagrams. Each task includes a math statement, a reference diagram, and executable TikZ or SVG code. Three models scored 31.25%, 26.56%, and 22.85% pass@1, and their outputs were mostly formal language and code rather than diagram-based reasoning, often visually confusing to humans.
In briefIn-context learning helps vision-language models overcome training priors
SummaryVision-language models learn statistical priors during training, which can make them ignore visual evidence when images conflict with those priors. The paper uses controlled visual in-context learning to steer models toward visual evidence, improving accuracy on prior-conflicting counterfactual examples (up to +27.2% for Qwen3-VL 235B) across five vision-centric tasks while leaving prior-aligned accuracy nearly unchanged. The gains appear across five VLM families, but are limited when the underlying visual capability is weak or absent.
In briefUniDeblur: One Motion Deblurring Model via Fourier Convolution Learning
SummaryThis paper presents UniDeblur, a unified framework for removing motion blur across different cameras and image content. The authors build a multi-source dataset and a data pruning and quality control pipeline. Instead of pixel-level mapping, the method learns blur-adaptive global filters in Fourier space. The paper reports competitive results on both in-domain and out-of-distribution benchmarks.
In briefSelecting LLM hidden-state layers for speculative decoding using information and geometry measures
SummaryThis paper studies how to choose which layers' hidden states from a large language model to use in speculative decoding. It proposes a selection framework that scores layer sets by informativeness, measured via matrix entropy, and diversity, measured via centered kernel alignment (CKA), and adds token-entropy filtering. In EAGLE3-style speculative decoding, the selected sets reportedly outperform fixed baselines, including the default [2, N//2, N-3] rule.
Raghavv Goel, Sudhanshu Agrawal, Mukul Gagrani et al.
In briefLearning More by Training Less: Uncertainty-Guided Progressive Contraction for Efficient Reasoning Post-Training
SummaryThis paper examines how far supervision for reasoning post-training of large language models can be compressed without losing capability. The method uses Monte Carlo reasoning trees to profile process-level uncertainty in teacher chain-of-thought, shrinking the corpus into a compact teacher-informed pool, then routes optimization online during SFT according to the student's learning state. The authors report that it uses as little as 0.22% of the optimization budget of full-data SFT while maintaining or improving reasoning performance across benchmarks and model families.
WhyIntroduces uncertainty-guided progressive contraction using Monte Carlo reasoning trees and learner-aware routing, cutting training budget sharply. Broad benchmark and model-family experiments support
In briefSEEDS: Preview-Guided Evidence Admission for Search Agents
SummaryThis paper argues that search agents usually add all retrieved documents to their reasoning by fixed or rank-based rules, even though a document's usefulness depends on the agent's current reasoning state. The authors propose SEEDS, which first shows lightweight previews of retrieved candidates, then lets the agent decide which candidates and how many to examine in full. Across nine knowledge-intensive QA benchmarks, using both local corpora and live web search, SEEDS improves search effectiveness while substantially reducing full-evidence consumption.
In briefCollaborative Learning over Model Slices: Small, Diverse Models Cross-Check Each Other's Reasoning
SummaryThe paper addresses the problem that a single large language model's reasoning errors can propagate through later steps, making them hard to reassess. It proposes collaborative learning, using model slicing to replace one large model with smaller heterogeneous models of comparable parameter budget, and pyramid supervision that combines protocol distillation with step-level supervision. Evaluated on multi-hop question answering, mathematical reasoning, code generation, and constrained open-ended generation, the trained heterogeneous models match or exceed the original large model and multi-agent baselines while achieving higher inference efficiency.
In briefWhat egocentric video embeddings encode and how to fuse them with inertial sensors for activity recognition
SummaryThe paper examines what pretrained embeddings of head-mounted camera video capture, finding that their leading directions are dominated by camera orientation and scene layout. The authors propagate inertial-sensor predictions over a within-participant video graph. On the WEAR dataset this raises subject-wise cross-validated macro-F1 from 0.648 to 0.852.
In briefSparse dictionary learning for probability-measure data via linearized optimal transport, with sample complexity bounds
SummaryThe paper studies dictionary learning for probability measure data by embedding measures into Hilbert spaces using features derived from optimal transport geometry. Sparse autoencoders learn data-adapted dictionaries that represent each measure as a sparse finite-dimensional code. The method is demonstrated on hyperspectral images, geometric point clouds, and document embeddings from large language model latent spaces. The authors also prove generalization bounds for sparse dictionary learning objectives and covering number bounds for optimal and entropic transport map dictionaries.
In briefTeaching neural decoders with language: using text class descriptions to guide brain signal decoding
SummaryThis paper proposes Language as Supervision (LaS), which aligns neural features with natural-language class descriptions and predicts classes using candidate text prototypes. The authors evaluate seven encoder architectures for visual, motor, and speech decoding from scalp and intracranial recordings. LaS improves speech-decoding accuracy over matched class-label baselines and enables conditional decoding of unseen frequencies in steady-state visual evoked potentials.
In briefBeyond Second-Order Information: Paid Logistic Regression with Noisy Features
SummaryThe paper studies online binary prediction where a learner can pay to obtain more precise observations of latent stochastic features. Unlike the linear regression case, covariance alone is insufficient, since noise distributions with equal covariance can yield different optimal logistic losses. Under a structural ordering where larger payments remove independent noise, regret is Õ(√T) when the noise structure is known and Õ(T^{2/3}) when unknown. Matching lower bounds hold up to logarithmic factors.
In briefControlled-Minimax: Using Control Theory to Stabilize Adversarial Optimization
SummaryThis paper studies instability in minimax optimization, which is central to adversarial learning, by modeling training as a continuous-time dynamical system. It identifies the eigenstructure of the loss Hessian near equilibria as the root cause of non-convergence. The authors propose Controlled-Minimax, a feedback controller claimed to give local exponential stability, and report experiments on Dirac-GAN, 2D Gaussian mixtures, Swiss Roll, and StyleGAN2-ada, showing suppressed oscillations and faster, more robust convergence.
In briefOGPO: Offline Goal-Conditioned Policy Optimization for Robots That Recover from Errors
SummaryVision-language-action models for robot manipulation are usually trained by imitating expert trajectories, which gives limited guidance when execution drifts off the demonstrated path. The paper proposes OGPO, which relabels offline demonstrations with semantic key states and uses goal-conditioned process rewards to teach policies to reach task-progress states. Experiments on LIBERO and MetaWorld show gains over standard supervised fine-tuning, and evaluations on LIBERO-Plus and the authors' LIBERO-ReAct benchmark indicate better robustness and recoverability.
In briefBLEND: Balancing personalization and generalization in federated vision-language models
SummaryThis paper introduces BLEND, a framework for fine-tuning vision-language models in federated learning to improve both personalization and generalization. It shares global vision and text adapters across clients and gives each client a personalized vision projection adapter, with a new loss function blending their outputs. Experiments reported in the paper show BLEND consistently outperforms state-of-the-art baselines across diverse system settings.
Payam Abdisarabshali, Fardis Nadimi, Kasra Borazjani et al.
In briefLatLM: Parallel Language Model That Delays Output Commitment
SummaryThis paper introduces the Lattice Language Model (LatLM), which decides adaptively during decoding when to commit each token while generating multiple text lines in parallel. Confident candidates are committed, while those whose prerequisites are not ready are deferred. At 482M parameters, the authors report near-autoregressive accuracy with 25–41% of the decoding rounds on independent-line tasks, and 96.5% versus 78.5% accuracy for LCLM on mixed dependency chains.
In briefGIST: Using Generator States to Improve Training of Tabular Foundation Models
SummaryTabular foundation models usually train on sampled hard labels, which adds avoidable randomness. This paper proposes GIST, which uses intermediate generator states to average query losses over conditional label distributions, reducing this variability. With matched pretraining, GIST improves mean rank across three architectures on TabArena and TALENT, and wins 62.1% of paired dataset-level comparisons against hard-label baselines.
In briefCompaction-aware reinforcement learning for generating Ascend Triton kernels
SummaryThe paper proposes ATRL (Ascend Triton Reinforcement Learning), a post-training method that adapts GRPO into a compaction-aware, segment-level objective for long-horizon kernel generation. It also introduces a lightweight agent harness giving cascaded compilation, correctness, and profiling feedback, plus an Ascend-specific knowledge base. On NPUKernelBench and KernelBench, run on Ascend hardware, ATRL consistently outperforms all baselines, including GRPO, in correctness and speedup.
In briefJMT-YOLO: Multi-Task Domain Adaptation for Object Detection via Segmentation-Guided Consistency
SummaryThis paper addresses the drop in object detection accuracy when training and test data follow different distributions. It proposes JMT-YOLO, a joint multi-task framework built on the Mean Teacher architecture that adds an auxiliary semantic segmentation task and a masking mechanism. On three benchmark adaptation tasks it reports detection accuracy of 46.9%, 57.3%, and 61.0%, which the authors state exceeds existing state-of-the-art and baseline methods.
In briefA lightweight add-on that helps fast text-to-image models follow complex instructions better
SummaryFew-step text-to-image generators often mishandle complex prompts involving multiple subjects, spatial relations, counting, and attribute binding. This paper trains a small corrector of under 20M parameters that adjusts the generator's velocities at selected sampling steps, while the frozen 6B-parameter generator and its forward-pass count stay unchanged. On the target complex-scene distribution, it reaches a +17.60-point NetWin over frozen Z-Image-Turbo in pairwise instruction-following evaluation, adding 26.9 ms per image.
In briefRethinking Diffusion and Decoding Roles: Predictive-State Flows for Parallel Text Generation
SummaryThis paper proposes Predictive-State Flows (PSF), a method for continuous diffusion language models that splits text into blocks and assigns each block a predictive state computed from preceding context. A rectified-flow prior generates these states, and a block-causal decoder then produces all blocks in parallel, autoregressively only within each block. On unconditional OpenWebText generation, a 197M-parameter model with K=8 reports better generative perplexity and MAUVE scores than ELF-B with 32 sampling steps, at lower computational cost.
In briefDS-Net: A Dynamic Split Network for Edge-Cloud Collaborative Inference
SummaryThis paper proposes DS-Net, a model that lets edge devices and cloud servers share deep neural network inference by choosing among several split points within a single network. Compared with existing methods that load a separate model for each split point, it aims to reduce storage, memory, and switching costs. A hierarchical compensation mechanism is introduced to correct feature distribution shifts and recover task-relevant information lost during splitting and compression. The reported experiments show a 74.6% memory reduction and millisecond-level switching latency.
In briefAD-SRF: attack-defense method that strengthens meaning in fused images
SummaryMulti-modal image fusion often trades visual quality against semantic content. This paper proposes AD-SRF, which first builds a transformer-based fusion network for visual quality, then uses an attack-defense game in which a defender restores missing semantics and reintegrates them via an imperceptible semantic map. The authors report that this improves downstream segmentation without hurting visual quality. Code is publicly available.
In briefMeshBind: Point-to-Mesh Binding for Automatic Scene Flow Annotation of Non-Rigid Objects
SummaryThis paper presents MeshBind, a method that uses mesh models to automatically generate per-point scene flow labels for non-rigid objects such as pedestrians, where ground truth is hard to obtain. It reconstructs articulated meshes from six surround cameras, registers them to LiDAR frames, binds points via per-triangle local frames, and combines ego-motion, box, and mesh binding in a three-layer pipeline. The authors also release SFB, a 41k-frame benchmark built on SUScape, and report that fine-tuning on it reduces three-way endpoint error by 2 to 3 times.
In briefModels that capture both attraction and repulsion in selecting item subsets
SummaryDeterminantal point processes usually model repulsion when selecting subsets, but many applications also show attraction from pairwise or shared latent relations. The paper proposes the determinant-ratio point process (DRPP), which uses a ratio of determinants of two symmetric positive-definite kernels to specify repulsion and attraction separately. An auxiliary-variable representation enables efficient MCMC sampling and tractable expectations for maximum-likelihood learning. Experiments show DRPPs fit transitive attraction with substantially fewer parameters than nonsymmetric DPPs.
In briefWin-Win Fusion: Image fusion that combines visual quality and semantic understanding
SummaryThis paper proposes W2F, a method for multi-modal image fusion that couples visual fusion with semantic fusion to close the gap between the two. It uses segmentation-guided loss, interactive attention modules, a feature search module, and a gradient correction strategy to make the two tasks reinforce each other. The authors report that W2F outperforms state-of-the-art methods in both visual appeal and decision accuracy, and the code is public.
WhyProposes a visual-semantic collaborative fusion paradigm with segmentation-guided loss, attention modules, and gradient correction, reported to surpass state-of-the-art methods.
In briefPersonalized federated preference tuning using LoRA adapters
SummaryThis paper addresses a limitation of federated direct preference optimization (FedDPO), which assumes all clients converge to one shared policy despite heterogeneous preferences. It proposes pFedDPO, where each client keeps a private LoRA adapter on top of a shared global model that is synchronized via FedAvg. The authors report that experiments on preference datasets show pFedDPO attains a lower gradient norm than vanilla FedDPO while remaining competitive in training loss and per-client reward accuracy.
In briefCoGameReg: Cooperative-Game Semi-Supervised Regression for Seismic Reservoir Prediction
SummaryThis paper addresses estimating reservoir properties from seismic data when well-log labels are sparse. It proposes CoGameReg, which builds shared pseudo-labels from contribution-weighted predictions of several regressors, with weights estimated by a tri-game Shapley mechanism. The method was tested on 10 public regression datasets, a synthetic P-wave velocity task, and a field water-saturation application. Its advantage is largest at the 2.5% labeled ratio, and performance varies across metrics and well locations.
In briefHART: A hybrid attention and RWKV model for time series forecasting
SummaryThe paper finds that RWKV-7 degrades over long horizons in non-autoregressive multi-step forecasting and attributes this to a mechanism called Placeholder Idling Decay. The authors propose HART, which alternates attention and RWKV blocks to combine direct access to history with recurrent modeling. On GIFT-Eval and TIME-Benchmark, HART improves over pure RWKV at medium and long horizons and beats a similarly sized Transformer overall. With 36M parameters, it matches Toto-2.0-1B on long-horizon MASE (0.752 vs. 0.754).
In briefMind the Spatial Gap: A Benchmark for Geospatial Reasoning and Constrained Spatial Editing
SummaryThis paper introduces GeoSpatialBench, a benchmark that tests multimodal language models on spatial understanding, geometric reasoning, and constrained spatial editing. It uses real urban data from six cities and covers 28 spatial relations, with answers verified through GIS operations. The results show models perform better with GeoJSON than map images, struggle more with quantitative metric relations, and most failed edits violate at least one spatial constraint.
In briefOne-Step Text Generation: The Seq-Drifting Method
SummaryThis paper proposes Seq-Drifting, a framework that aims to generate an entire text sequence in a single forward pass instead of the multiple sampling steps used by diffusion language models. It applies drifting to sequences of continuous token representations and learns a direct map from Gaussian noise to text, with position-wise targets that adapt to the generated context. The authors report a generative perplexity of 36.01 on LM1B in one step, and experiments on unconditional and conditional generation, translation, summarization, and reasoning.
WhyExtends drifting to token sequences for one-step generation, a novel direction, with experiments on LM1B and several tasks but limited comparison depth.
In briefWhen solvers get moody: measuring and predicting how much MIP solver performance varies between runs
SummaryMixed-integer programming solvers can vary widely in performance even under minor perturbations, which complicates benchmarking. The authors introduce the Weighted Variability Score (WVS), which combines dispersion in runtime, optimality gap, LP iterations, and branch-and-bound nodes across repeated runs. On a large MIPLIB 2017 subset, most instances are stable while some show large, not fully explained variability. A tree-based predictor using early root-level features forecasts high variability with moderate accuracy, offering a tool for tuning and fairer evaluation.
Mourchid Adegbindin, Brenda Dietrich, Jamol Pender et al.
In briefT-Mem: Temporal Evidence Forests for Long-Horizon Multi-Party Dialogue Memory
SummaryThis paper addresses how LLM-based agents can keep track of long-term memories in multi-party conversations, where information and states change over time. It proposes T-Mem, a memory framework that uses temporal structure to recover locally concentrated evidence, model state transitions over time, and compress low-information spans into an evidence forest. The forest is then linearized for answer generation. The authors report that T-Mem consistently outperforms competitive memory frameworks on two long-term multi-party memory benchmarks.
In briefTeaching AI to Predict User Replies and Revise Its Own Answers
SummaryThis paper proposes Collaborative Theory of Mind (CToM), a framework that trains large language models to predict how users will respond to a draft and use that prediction to revise the draft before answering. During training, the user's actual reply serves as the prediction target, combined with group-relative policy optimization (GRPO) to optimize task performance and interaction efficiency. Across mathematical reasoning, coding, and writing tasks, CToM achieves higher collaborative performance than all baselines while reducing the number of user replies needed.
Fanzhi Zeng, Anubrata Das, Maytal Saar-Tsechansky et al.
In briefROSETTA: Training-free alignment of feature spaces across sensors and encoders
SummaryEncoders trained separately on different sensors or models produce feature spaces with different coordinate systems, making cross-modal and cross-device matching difficult. The paper shows that the relative transformation between such systems is a rotation in whitened feature space, and that low-rank cross-covariance makes local alignment tractable. ROSETTA estimates this local rotation with an expectation-maximization procedure, requiring no training or co-observation at query time, and is validated on three public driving datasets.
In briefDiffusion-based generative design of ship hulls, checked with high-fidelity CFD
SummaryThe paper proposes a reward-guided diffusion framework that refines only selected hull regions, such as the bulbous bow and forebody, while keeping the full hull continuous. Surrogate-assisted sequential Monte Carlo sampling targets lower predicted calm-water resistance. A validated viscous free-surface CFD pipeline shows the generated hull has higher resistance at higher Froude numbers, while a naval-architect-refined hull performs better at lower speeds. The authors attribute these gaps to limits in the hull representation and the reward function.
Hadi Keramati, Mohammed A. Hannan, Patrick Kirchen et al.
In briefSelective Spatio-Temporal Fruit Inspector (S²FI): A Mamba-based Framework for Checking Fruit Quality While It Rotates
SummaryThis paper addresses the high cost of large automated sorting systems and the inefficiency of manual sorting for small and medium fruit growers. It proposes S²FI, a lightweight deep learning architecture combining a 3D convolutional encoder with a Mamba-based backbone, which weights informative frames during fruit rotation. The authors also built a conveyor-based fruit quality video dataset of 6,460 samples and report the best trade-off among accuracy, efficiency, and practicality on a held-out test set.
In briefCaptchaAgent: A Verifiable Visual Reasoning Framework for Solving CAPTCHAs
SummaryThis paper presents CaptchaAgent, a structured visual reasoning framework for solving interactive CAPTCHAs. It first builds a visual intermediate representation that records entities, attributes, and spatial relations, then refines ambiguous regions and checks whether actions stay stable under task-preserving transformations. On the Open CaptchaWorld and Oedipus-data benchmarks, the authors report higher Pass@1 scores than baseline agents. They also identify exact counting, continuous geometry, and ordered execution as major difficulties for current agents.
In briefRobust sampling method for noisy, heavy-tailed data, applied to fine-tuning language models
SummaryThis paper studies a sampling method (stochastic gradient Langevin Monte Carlo) for complex distributions where standard assumptions about gradients and noise variance may not hold. The authors introduce a heavy-tailed noise model and propose three methods, R-LMC, A-LMC, and RA-LMC, with non-asymptotic convergence guarantees in Fisher information. Experiments include synthetic problems and Bayesian fine-tuning of large language and vision-language models.
WhyExtends Langevin sampling to state-dependent heavy-tailed noise with non-asymptotic Fisher convergence guarantees, validated on LLM fine-tuning experiments.
In briefTime Delay System for Sequence Modeling: A New Framework That Estimates Delays Dynamically From Semantic Similarity
SummaryThe paper argues that existing sequence models lack an explicit, interpretable way to recall past content. It proposes a framework based on nonlinear time-delay systems, which dynamically estimates the effective delay from the semantic similarity between current and historical inputs. The method combines a differentiable semantic lag estimator, a diagonal state space with closed-form cumulative-sum solution, and an interpretable fuzzy semantic modulation mechanism. According to the authors, experiments on text classification and generation outperform state-of-the-art methods.
In briefLow-cost Satire Explanation of Australian and American News Using Unreliable Narrator Framing
SummaryThis pilot study explores generating explanations for satirical news articles from Australia and the USA. The authors propose framing explanations using unreliable narrator theory with open-weights models, presenting it as a feasible, cost-effective alternative to frontier models. Models identified satirical tropes more accurately on American articles than Australian ones. The work also contributes a novel dataset and evaluation benchmark for satirical explanation.
In briefChecking whether a physics-informed neural network's equation parameter was set wrong
SummaryPhysics-informed neural networks are trained to satisfy a given equation, so a low loss does not reveal a wrong coefficient in that equation. The paper proposes a closed-form least-squares estimator that recovers the parameter from a single evaluation of the trained network when the residual is linear in the parameter. Its accuracy is tied to a normalized residual measure, and the paper reports test results on Burgers' equation and a blowup-rate case, along with cases where errors were not detected.
In briefPreventing Lost Skills When LLMs Train Themselves
SummaryThis paper studies on-policy self-distillation (OPSD), where a language model acts as both teacher and student to improve itself. The authors find that OPSD can cause previously correct answers to become wrong during training, even while overall performance keeps rising. This degradation coincides with worsening geometry of hidden representations, with clusters becoming less compact and less separated. The authors propose Representation Geometry Regularization (RGR), which encourages compact clusters and separation between them, and can be added to existing OPSD methods; experiments reportedly show it reduces degradation while keeping gains.
In briefUniversal Bounds for Self-Normalized KL Divergence Estimation
SummaryThis paper studies an entropy-based estimator of KL divergence built from normalized importance weights. The authors prove that its expectation exceeds the target-proposal KL divergence by at most log 2 for any sample size. They also show that the KL divergence between the importance-weighted resampling distribution and the proposal is bounded by both the estimator's expectation and the target-proposal KL divergence. These bounds need neither bounded weights nor finite higher moments, and are applied to Soft Best-of-N sampling and relative entropy coding.
In briefTwo-way Guidance: Jointly Learning Summary Statistics for Generative Bayesian Inference
SummaryThis paper studies simulation-based Bayesian inference (SBI), which can struggle to recover increasingly concentrated posteriors as observation size grows. The authors propose learning sufficient summary statistics jointly with the posterior, while imposing a structural restriction on how the posterior depends on the learned summary. They describe a "two-way guidance" mechanism and provide theoretical guarantees on posterior estimation error and the summary's departure from sufficiency, including conditions on the simulation budget.
Percy S. Zhai, Max Pachman, Peverel D. Shipley et al.
In briefChanging App Code and Deployment Setup Together: A Study of AI Agent Pull Requests
SummaryThis paper examines about 27,000 pull requests from six AI coding agents and a human baseline, focusing on PRs that modify both application code and deployment infrastructure such as CI/CD pipelines and containers. Such dual-scope changes make up only about 4.5% of agent PRs, but their integration outcomes are bimodal. Some are reviewed and take much longer to merge, while many merge silently with little or no inspection. The authors conclude that deployment-aware benchmarks and strict path-based access controls are needed.
In briefTraining flow policies in RL without data via fixed-point flow matching
SummaryThe paper addresses sampling from the Boltzmann-form optimal policies in maximum entropy and behavior-constrained reinforcement learning, which are often complex and multimodal. It proposes fixed-point flow matching (FFM), a data-free method that formulates the actor update as a fixed-point iteration using a flow-matching-style regression with critic action gradients as targets. The authors report that FFM fine-tunes pretrained flow policies in offline-to-online settings and trains flow policies from scratch efficiently in online RL.
In briefWhen Generative Models Act as Simulators: Whether Their Internal States Can Be Edited Like a Simulator
SummaryThis paper studies when a generative model trained only on observations can be edited like a simulator, meaning changing its latent representation yields the same next-step outputs as editing the simulator's state. The authors compare two mappings learned afterward, probes that decode and steer the state and an inverse map, across the board game Othello and their continuous world Rayworld. They find the successful mapping depends on the environment: in Rayworld, probes fail to steer along the continuous state, yet an inverse map still sets the model from that state.
In briefDefining and measuring creativity in AI applications through newness, value, and surprise
SummaryThe paper proposes formal postulates of creativity grounded in psychology, centered on Boden's criteria of newness, value, and surprise, and offers a conceptual mathematical definition aligned with them. The authors validate the formula on chess moves and on paintings by artists such as Picasso and Renoir. Experiments show it captures creative patterns in both domains.
In briefAPEX: Asynchronous Prefix Denoising for Extreme Latent Compression
SummaryThis paper addresses the trade-off in diffusion models where compressing latents improves cheap generation but hurts reconstruction quality. It introduces APEX, which combines a hierarchical latent space with asynchronous denoising over channel groups. An autoencoder orders channel prefixes from structure to detail, and one shared denoiser generates them at staggered local times without adding tokens. On ImageNet-512, the authors report gains over existing methods at 64x and 128x compression and over a 256-token model at matched compute.
David Dinkevich, Nisan Chiprut, Yoav HaCohen et al.
In briefNeural networks show more modularity at subspace level than neuron level
SummaryThe paper proposes a nested hierarchy of modularity, spanning unit-level, orthogonal subspace, and linear subspace levels, evaluated causally through selective ablation. Across vision models, language models, and multi-task recurrent neural networks, it finds that networks lacking neuron-level modularity often show clear functional modularity within subspaces. These subspaces also dissociate in ways that align with brain activity data and allow targeted intervention on specific computations.
WhyProposes a nested hierarchy of modularity and validates it across vision, language, and RNN models via selective ablation and brain alignment, showing strong novelty and solid evidence.
In briefSame emotion label, different meaning: how “fear” and “surprise” vary across datasets
SummaryThis paper tests whether five emotion labels shared by GoEmotions, XED, and BRIGHTER mean the same thing across these datasets. On identical texts, one fear classifier flagged 71% as fearful and another 8%, though their rankings largely agreed. When models were transferred across corpora, they lost 55% of native discrimination overall, and for surprise the resources disagreed on which texts express it. The authors conclude that a shared emotion label name does not make a resource an interchangeable measurement target.
In briefCOFFEQA: Counterfactual-Guided Few-Shot Distillation for Question Answering
SummaryThis paper proposes COFFEQA, a method that improves few-shot knowledge distillation for open-book question answering by adding counterfactual training pairs. The authors find that using training samples near the teacher model's decision boundary can significantly hurt distillation performance. Counterfactuals are minimally modified contexts that make the teacher give a different answer, and a new metric, Context Following Difficulty (CFD), is used to validate them before distillation. Experiments on five in-domain and three out-of-domain QA benchmarks with T5 and Qwen2.5 models show COFFEQA outperforms factual-only baselines under the same few-shot budget.
In briefStructure-aware matching for locating drones by matching thermal images to satellite maps
SummaryThis paper proposes SAMF, a framework that matches real-time UAV thermal images to georeferenced satellite images by exploiting stable structural cues shared across the two modalities. It combines a Directional Structure Enhancement Module (DSEM) and a Modality-Aware Affine Remapping Module (MARM), and introduces the Thermal-UAV-Paired benchmark. The authors report state-of-the-art results on the Boson-nighttime and Thermal-UAV-Paired datasets.
In briefUnified training from demonstrations and rollouts via a shared regression target
SummaryThis paper proposes Cross-Behavior Policy Optimization (CBPO), which uses one regression loss with a fixed reference target to learn from offline demonstrations, current rollouts, and historical rollouts in a shared replay buffer. It avoids separate SFT and RL objectives and does not require stored behavior probabilities, importance weights, or a critic. Across two model scales and four math reasoning benchmarks, CBPO-Unified improves average avg@32 and pass@32 over SFT followed by DAPO, by 2.33 and 4.44 points on Qwen3-1.7B.
In briefScalable learning of determinantal point processes using kernel-structured interactions
SummaryThis paper studies determinantal point processes (DPPs), probability models over a finite set in which items repel each other according to a positive definite kernel matrix. Scaling DPP learning to large datasets usually requires structural constraints on the kernel, which trade expressive power against computational scalability. The authors model full-rank interactions with a nonlinear kernel and latent features, design a surrogate for the maximum likelihood estimator, and use stochastic optimization with random features of the nonlinear kernel. Experiments on real-world datasets show the method improves this trade-off, achieving better generalization with high parameter efficiency.
In briefTreeHCA: Hindsight-guided exploration and credit assignment for multi-turn LLM agents
SummaryThis paper introduces TreeHCA, a training method for multi-turn LLM agents that couples hindsight-guided exploration with hindsight credit assignment. It revisits intermediate histories based on changes in reference-answer likelihood and samples alternative continuations, forming a tree of reasoning and tool interactions. A normalized tree backup then propagates terminal outcomes into turn-level advantages. Across seven question-answering benchmarks and three Qwen backbones, TreeHCA achieves the highest aggregate exact match among all baselines.
Debajoy Mukherjee, Jeremy Carleton, Sushil Vemuri et al.
In briefWhen Charts Deceive Web Agents: A Benchmark on Misleading Visualizations
SummaryThis paper introduces MisVisAgentBench, a paired executable benchmark that tests how misleading charts affect web agents' task execution. It contains 280 task instances, each with a clean and a misleading chart that differ only in the chart evidence. Across thirteen vision-language web agents, switching from clean to misleading charts lowered average task success by more than 34 percentage points, and failures mostly followed the misleading action branch. The authors conclude that visualization robustness is an important but understudied part of web agent safety.}
In briefGenerating acceleration waveforms matching a target shock response spectrum using an ensemble of conditional VAEs
SummaryThis paper addresses inverse synthesis: generating acceleration waveforms whose shock response spectrum matches a given target. The authors extend a prior conditional variational autoencoder with a recurrent decoder that predicts wavelet coefficients, reconstructed via an inverse wavelet transform, and use an ensemble of models with different mother wavelets and decomposition levels to produce distinct candidates. On four held-out datasets, the method shows lower mean and median spectral errors than the earlier CVAE and a sum-of-decayed-sinusoids baseline. Single-model inference takes about 0.1 to 0.6 ms per waveform, versus 5 to 8 s for the baseline, but the physical suitability of the candidates beyond spectral agreement still requires assessment.
In briefLeWAM: End-to-end JEPA for world and action modeling in control
SummaryThe paper introduces LeWAM, an end-to-end joint-embedding predictive architecture (JEPA) for control that learns latent world models without pixel reconstruction. It trains next latent state prediction and action flow matching jointly and applies an isotropic Gaussian constraint on the latent state, so one model supports direct policy execution and planning. According to the paper, success rates on contact-rich and long-horizon visual planning tasks rose from 28.6% to 89.7% and from 10.9% to 46.2%, with planning 32.3 times faster.
In briefDeploying RDMA at scale for distributed ML training and inference
SummaryThis paper reports ByteDance's experience deploying ByteExpress, a software communication layer over hardware RDMA, in parameter-server-based inference and training systems. The authors share three lessons: buffer management must span application and transport layers, connections and receive capacity need coordinated management, and availability requires recovery above the transport layer. In production inference, compared with kernel TCP, ByteExpress reduces mean and P99 model parameter fetch latency by 35.7% and 43.7%, and total CPU usage by 10.69%. In training, it reduces P99 parameter push latency by up to 74.0%, with no PFC pauses observed.
In briefBudget-Constrained Audio-Visual Alignment via Adaptive Particle Search and MH Refinement
SummaryThis paper proposes SWIM, a training-free, reward-agnostic two-stage inference-time alignment framework that combines particle search with Metropolis-Hastings refinement under a fixed compute budget. The first stage adaptively allocates particles across denoising checkpoints using online variance estimates, and the second stage applies local noise-and-denoise proposals to completed outputs. The authors report relative gains over BoN and EvoSearch on JavisBench-Mini, and improvements over uniform allocation on SD1.5 image generation.
Shivanshu Shekhar, Raghavendra Addanki, Tong Zhang et al.
In briefHow well random-feature message-passing networks generalize across graph sizes
SummaryThe paper studies how random node features affect the expressive power of message-passing neural networks (MPNNs) when approximating functions across graphs of different sizes. The authors show that continuous random features can be quantized to finitely many colors with arbitrarily small uniform error in expected output. From this they prove that these MPNNs are not size-uniformly universal and analyze gains from colors and independent trials. They also give finite-color constructions for tree subgraph isomorphism and a local MST task.
WhyIntroduces a population-level framework proving size-uniform non-universality and incomparability with higher-order WL, a clear theoretical advance with rigorous support.
Robert R. Nerem, Mitchell Black, Antonis Vasileiou et al.
In briefPeople Making Conversation: How Naturally Do Speech Models Converse?
SummaryThis paper evaluates whether real-time conversational speech models respond appropriately at each pause and turn end. The authors tested six systems at 427 pauses and 482 turn ends from 176 unscripted conversations in Seamless Interaction, comparing their outputs with what the person actually did. Hosted API systems respond at nearly every pause, while full-duplex models more often stay silent and can interrupt within human reaction time, and hosted systems tend to reply longer than people. The authors also measured how well systems recall conversation facts, and they release the labels, harness, and all outputs.
WhyProposes a new evaluation of whether speech models speak at real conversational pauses, comparing six systems on 427 pauses and 482 turn ends; solid evidence, useful for the subfield.
In briefNeural network model with guaranteed convexity for predicting how anisotropic materials deform under stress
SummaryThe paper builds a neural-network energy model for the effective anisotropic hyperelastic response of a representative volume element, using learned paired directional features. Its convex cores guarantee objectivity, a stress-free reference state, two-dimensional polyconvexity and growth conditions by construction. Compared with an unconstrained model and three reduced-order references on periodic cells, the constrained energies stay accurate within the data and keep their guarantees beyond it, while the unconstrained energy loses rank-one convexity.
S. Ares de Parga, A. Cornejo, J. A. Hern\'andez et al.
In briefEfficient audio-visual speech enhancement found by reinforcement-learning architecture search
SummaryThis paper proposes a reinforcement learning-driven neural architecture search framework to find compact audio-visual speech enhancement models suitable for edge devices. The search evaluates only 35 candidate architectures and yields a model with 2.9 million parameters, compared with 22 million in the baseline. The model achieves a 14.5 dB SI-SNR improvement, 0.7 dB below the baseline, with about nine-fold smaller size and a 4.3-fold lower real-time factor on mobile CPUs.
In briefMeta-Reflective Evolution: Improving Production AI Systems
SummaryThis paper introduces CoCoEvolve, an evolutionary optimizer built around "meta-reflection," in which an agentic mutator retrieves evidence from the whole evolutionary history and probes the live system before proposing changes. It is applied to optimize data agents, LLM data pipelines, and data engineering workflows. On DABStep, the authors report 92.8% held-out hard accuracy, compared with 78.4% for Meta-Harness and 30.4% for GEPA.
Seung Won Wilson Yoo, Chang Zeng, Aditi Somayajula et al.
In briefHardware-aware LLM-VLM integration for zero-shot robot task planning and grasping
SummaryThis paper presents a robot manipulation framework combining a language model and a vision-language model for a simulated Franka Research 3 robot in NVIDIA Isaac Sim. The language model produces only symbolic task information, such as the target object and relations, while coordinates, grasp poses, and trajectories are handled by deterministic modules including MoveIt 2 collision checking and inverse kinematics. The framework is evaluated in digital-twin trials covering pickup, occluded-source access, keepout reasoning, placement, and sequential manipulation.
Lonnie Tankersley, Noorbakhsh Amiri Golilarz, Soroush Korivand
In briefLarge Language Models and Machine-Assisted Cultural Analytics: A Chapter on the Agentic Turn
SummaryThis chapter examines how large language models can turn cultural material into analyzable variables, and how LLM-driven agentic systems extend machine assistance to other stages of research. It is not a how-to guide or an exhaustive review, but a discussion of key aspects of this paradigm and its broader picture. The author argues that these technologies expand scale, methodological range, and what can be delegated, while the basic logic by which measurements support cultural claims remains largely unchanged.
In briefEvoBelief: Falsifiable World Models for Self-Improving LLM Agents
SummaryThe paper introduces EvoBelief, a self-evolving world model that lets LLM agents refine their beliefs about environment dynamics during deployment by treating them as falsifiable hypotheses. Unexpected outcomes trigger new hypotheses, which are validated by later observations and experiments, while counterexamples narrow or retire rules. The language model stays frozen. Across four benchmarks, the method improves task performance and reuses learned rules across tasks, including on ALFWorld, MiniGrid, and DiscoveryWorld.
WhyTreats environment knowledge as falsifiable hypotheses revised during deployment, with comparisons across four benchmarks supporting gains over frozen world models.
In briefAdapting SAM 2.1 to segment building parts for 3D reconstruction from images
SummaryThis study fine-tunes SAM 2.1 on the HBD dataset of 2,679 annotated images from 42 buildings to segment 11 building-component classes. Multi-class mIoU rose from 49.9% to 69.6%, and it was compared against a SegFormer baseline. In a neural radiance field, the fine-tuned model grouped building components better than the pretrained model at a user-selected grouping scale, supporting 3D building reconstruction.
Ehsan Agha Ebrahimi, Yong-Cheol Lee, Keith G. Mills
In briefA method that refines latent features to keep tracking the referred object in audio-visual video segmentation
SummaryThis paper addresses Referring Audio-Visual Segmentation (Ref-AVS), which locates and segments a specific object in a video using a language description grounded in both visual and audio cues. It notes that existing models often confuse objects when occlusion, fading sound, or identity ambiguity causes target shift. The authors propose the SAIL framework, which uses a dual-token strategy, a shift-aware refinement module, and a decoder with continuous memory updates. They report significant gains over existing AVS, Ref-VOS, and Ref-AVS methods on two benchmarks.
In briefAligning two encoders from few pairs via distilled optimal transport
SummaryThis paper studies aligning two independently pretrained unimodal encoders when only about 100 to 500 paired examples are available, while unpaired data is abundant. Each representation is expressed by its similarities to the paired anchors, and a single entropic optimal transport problem matches the two point clouds, then distills the plan into two lightweight encoders. On image-text and molecule-text alignment, the method outperforms existing approaches at the same anchor budget.
In briefTemporal Diffusion Bridge Model for Vessel Extraction from Physics-Informed Synthetic Data
SummaryThe paper proposes a Temporal Diffusion Bridge Model (TDBM) that treats vessel extraction from X-ray coronary angiography as image translation, trained on physics-based synthetic image pairs without manual vessel annotations. It uses adjacent frames for temporal consistency and a vessel-aware contrastive loss. The authors evaluate it on two XCA segmentation datasets, a reader study with two real-patient sequence datasets, and zero-shot tests on retinal and fluorescence microscopy data, reporting consistent gains over existing methods.
WhyPhysics-based synthetic data and a temporal diffusion bridge enable annotation-free vessel extraction, validated on multiple datasets, a reader study, and zero-shot tests; impact is mainly within medi
Jinkui Hao, Donald R. Cantrell, Ramez Abdalla et al.
In briefCoordinate-Free Automatic Differentiation: A Reusable Derivative Operator for Geometric Calculus
SummaryStandard automatic differentiation frameworks fix a Euclidean basis to compute gradients and Jacobians, which makes geometric objects such as rotations, volume forms, and curvatures awkward to express. This paper introduces a single basis-independent derivative operator that can represent many geometric differential operations. The authors show it can be built entirely from existing AD primitives and composed with JAX batching, compilation, and hardware acceleration. Experiments demonstrate reuse across different manifold operators, use in a spacetime inverse-problem pipeline, and optimization over different rotor subgroups.
In briefCrediting Soccer Players for Goal Buildup Using Hawkes Processes
SummaryTraditional soccer statistics credit only the final pass before a goal, overlooking players who contributed earlier in the attack. This paper models timestamped sequences of passes and shots as continuous-time networks using multivariate Hawkes processes, which quantify how much each pass contributes to later passes or shots. Building on this, the authors propose expected goal involvement (xGI), a metric measuring a player's involvement in creating scoring chances through their passes, including players who did not make the final pass.
In briefREDaCT: Unlearning in Reasoning Models Without Breaking Their Reasoning
SummaryThis paper addresses unlearning in large reasoning models, where existing methods often remove sensitive information but cause the model's chain-of-thought to become incoherent or fail to terminate. The authors propose REDaCT, an inference-time framework with a binary detector that flags reasoning sentences leaking forget-set information and an offline-learned corrective embedding injected at the first detected leak. Experiments reported in the paper show REDaCT avoids reasoning collapse entirely, better preserves reasoning than prior methods, and maintains strong forgetting of answers and chain-of-thought content.
In briefSemViews: Certified Reuse of LLM Labels Across Semantic Queries
SummarySemantic query engines use a large language model to evaluate natural-language filters, and each evaluation is a costly call. The paper presents SemViews, which orders tuples using signals that need no model calls, such as labels from earlier queries, local embeddings, and a helper model's score, then refines the order with paid labels and certifies the answer with fresh stratified samples. On four replayed workloads it meets both precision and recall targets on 99.63% of queries while saving 28%-45% of oracle calls.
WhyReuses cached oracle labels as certified semantic views with stratified sampling guarantees, replaying 725,000 labels to cut oracle calls versus prior guaranteed methods.
In briefA provenance-grouped protocol for testing machine learning on perovskite solar cells
SummaryThe paper finds that random data splits in perovskite solar cell machine learning leak information, because nearly two-thirds of devices in the Perovskite Database share an exact recipe with another device. The authors build grouped splits that isolate duplicate devices, publications, and institutions, then benchmark thirteen models. Tree ensembles lead on random splits but lose their ranking on unseen institutions and future discoveries, where composition-aware neural networks perform better.
In briefAutomatically building faithful, complete hardware knowledge graphs from datasheets with LLMs
SummaryThis paper presents an LLM-driven pipeline that converts hardware datasheets into hierarchical knowledge graphs using chunk-based property extraction, hierarchical routing, and iterative refinement while keeping source provenance. On 20 datasheets, it reaches 98.8% property faithfulness versus 76.1% for a one-shot baseline. It also fully covers 83.6% of the faithful baseline properties.
In briefDistressBench: Testing How Well AI Chatbots Support People in Suicidal or Self-Harm Crises
SummaryThis paper introduces DistressBench, a benchmark of 718 English conversations with clinician-written user messages covering 23 suicide and self-harm subcategories, each scored against weighted criteria. The best of 25 frontier models met 88.4% of the weighted criteria, the median model 71.4%, and clinician reference responses 98.7%. The authors argue that crisis safety should be judged by the care a model actually delivers, not by whether it refuses to respond.
WhyProposes rubric-based crisis-support evaluation beyond refusal, with 718 clinician-authored conversations, 25 models, and clinician-reference comparison, offering solid evidence and influencing safety
In briefMADS: A multi-agent simulator that generates challenging dialogues to train sturdier chatbots
SummaryTask-oriented conversational agents often fail when users are uncooperative, and existing function-calling benchmarks contain few adversarial conversations. The paper presents MADS, a synthetic pipeline that generates structured, validated dialogue trajectories, including realistic unhappy paths. Using it, the authors build MADS-Bench, covering six domains, 55 personas and 1095 conversations, with about 66% containing failure-prone scenarios. They evaluate robustness of models trained with supervised fine-tuning followed by reinforcement learning with verifiable rewards.
Arkajyoti Chakraborty, Adam Zernik, Ishika Kulkarni et al.
In briefHow Hardware Faults Show Up Differently in LLM Agents
SummaryThis paper is the first systematic study of how large language model (LLM) agents withstand random hardware faults. Across three agent benchmarks, the authors ran 6,814 executions, including clean runs and runs with injected random two-bit weight faults. They find that agent execution turns model-level corruption into different system-level outcomes, such as agents recovering and still completing tasks at higher cost. For some models and scenarios, faults can also alter security and alignment behavior.
In briefNV-Tesseract-AD: curriculum-enhanced diffusion model for detecting anomalies in multivariate time series
SummaryThe paper proposes NV-Tesseract-AD, a diffusion-based framework for detecting anomalies in multivariate time series. It uses a Transformer backbone and a three-stage curriculum that increases masking ratio and diffusion noise level together. Trained on TSB-UAD-M and tested zero-shot on ten EasyTSAD benchmark families, it reports about 3.7-fold higher F1 and 56% lower MSE than a non-curriculum baseline. The authors identify unsupervised threshold calibration, rather than reconstruction quality, as the main remaining bottleneck.
Aditi Gautam, Manikandan Ravikiran, Saira Qureshi et al.
In briefPredicting Self-Driving Car Paths Through Road Work Zones Using Human Driving Data
SummaryThis paper studies how autonomous vehicles can predict their own trajectories in urban road construction zones, an area with little prior research. The authors collected WZDrive-Data, a multimodal dataset of 1,183 work-zone driving segments from 33 drivers, using a human-in-the-loop driving simulator. Their proposed predictor, WZDrive, reports a minADE (K=6) of 1.09 m and minFDE (K=6) of 2.12 m, improving on the best baseline for each metric by 7.18% and 13.18%, respectively.
In briefRandom masks beat gradient masks for limiting forgetting in LoRA B-matrix updates
SummaryThis paper studies which LoRA adapter parameters to update during continual learning. The authors note that earlier comparisons confounded selection rules with candidate pools, since gradients with respect to A are zero under standard initialization. Restricting both methods to the same B-matrix pool, a uniform random mask gave consistently lower loss-based backward transfer than gradient attribution across seven seeds on Qwen2.5-3B, mainly driven by the medical task.
In briefArabic Relation Extraction with Ontology-Constrained Encoder and LLM Systems (U4RASD paper)
SummaryThis paper describes a submission to KnowledgeGraphEval 2026 Subtask 2 on directed Arabic relation extraction. Given an Arabic sentence and a subject-object pair, the systems predict one of 40 relations or no relation. The best system, an ensemble of Yehia-RE, Qwen-RE, and TypedPair, reached a micro-F1 of 94.06 and ranked first in the subtask. The authors highlight the value of constrained prediction and complementary modeling approaches.
Ahmad Abbas, Hussein Mazeh, Chadi Abou Chakra et al.
In briefNeural Expectation-Maximization and Meta-Models: Reading Transformers as Probabilistic Inference
SummaryThe paper interprets Transformers as learned approximations to a hybrid inference procedure combining max-product Loopy Belief Propagation with Expectation-Maximization. It argues that self-attention approximates max-product LBP on an input-conditioned, dynamically structured Markov Random Field, and that parameter transitions across layers act as an implicit M-step. Building on this view, the authors introduce meta-models, where inputs generate instance-specific parameters, unifying self-attention, mixture-of-experts, and gated units, and report empirical results supporting the perspective.
In briefMinimax rates for estimating powered sliced Wasserstein functionals
SummaryThis paper studies two-sample estimation of the powered sliced Wasserstein functional under projected r-th moment assumptions. It establishes minimax rates in the noncritical regimes, where the plug-in estimator is rate-optimal, with order N^{-(r-p)/r} for p<r<2p and N^{-1/2} for r>2p. At r=p, uniformly consistent estimation is impossible, and at the critical case r=2p a logarithmic gap remains between the lower and upper bounds.
In briefKITE: A Knowledge-Injected Temporal Encoder for Time-Series Event Forecasting
SummaryThis paper addresses forecasting rare, costly events in multivariate time series, where precursors are seldom observed and purely data-driven models have little to learn from. The authors propose KITE, which injects engineers' semantic prior knowledge into a self-supervised event forecaster via two switchable mechanisms, a covariate mechanism and an attention-bias mechanism. Results indicate that informative priors help when the context is short and that priors already recoverable from the observed window add nothing, while the attention-bias mechanism showed no measurable gain.
Jonas Petersen, Pratik Jawahar, Riccardo Maggioni et al.
In briefKernelized Parameter Sensitivity Atlases for Faster Input-Dependent Model Interpretation
SummaryThis paper introduces KPSA, a method that estimates input-dependent parameter sensitivities with a lightweight kernel instead of repeatedly running backpropagation for each sample. On vision, forecasting, and protein-modeling tasks, its selected parameter groups reach 39-85% of the perturbation susceptibility of a first-order backpropagation oracle. For activation attribution of transformer MLP features, it increases query-time throughput and reduces peak GPU memory compared with the backpropagation oracle.
Davide Wiest, Ahanaf Hasan Ariq, Abdul Monaf Chowdhury
In briefReference-free food volume estimation from wearable stereo video
SummaryExisting stereo-vision methods for estimating food portions require users to take multiple phone photos or place a reference object for scale calibration. This paper proposes a method that uses stereo video from a wearable egocentric device to reconstruct food in true physical scale and estimate its volume. It combines zero-shot stereo matching, food segmentation, and keyframe TSDF fusion to produce a 3D mesh from a single recording. On five food items, the method achieves a mean absolute percentage error of 9.1%.
In briefStereoEngine: Scaling Stereo Bimanual Robot Data Without Stereo Cameras
SummaryStereoEngine synthesizes stereo data for robotic manipulation to address the scarcity of real-world stereo datasets. One method warps monocular video into a right view using estimated depth and inpaints disoccluded regions with a video diffusion model. The other renders stereo pairs in simulation and jointly generates photorealistic views. The authors report that the generated stereo quality beats existing baselines and that using the synthetic data improves downstream bimanual manipulation.
Jason Chen, I-Chun Arthur Liu, Gaurav Sukhatme et al.
