In briefRLSplat: Resource-linear budget control for fast 3D scene reconstruction
SummaryFeed-forward 3D Gaussian Splatting builds 3D scenes from images in one forward pass. Earlier budget-controllable methods predict many Gaussians and then prune them, so inference cost stays nearly constant regardless of budget. RLSplat instead learns to allocate a Gaussian budget across local scene regions, generating exactly the requested number, and refines them with neighbor-limited decoding so that size, latency, and peak memory fall linearly. In zero-shot tests it beats prior methods at every budget, running up to 4.1 times faster with 9.6 times less peak memory at low budgets.
In briefHow fine-tuning objectives shape layer-wise information in LLM hallucination detection
SummaryThis paper compares two fine-tuning approaches for binary hallucination detection: classifier-head fine-tuning (CLS) and LLM-as-Judge sequence-to-sequence fine-tuning (S2S). CLS performs better, producing hidden states with a clearer two-cluster structure and stronger class separation. The authors also find that upper layers become largely redundant and truncate the network above the identified information window, keeping accuracy unchanged on Qwen3-4B and 8B while cutting latency by 1.4×.
In briefVEOcc++: Persistent Voxel Feature Maps for Online and Offline Semantic Occupancy Mapping
SummaryVEOcc++ is a voxel-based framework for robots that builds semantic occupancy maps online during exploration and refines them offline afterward. It links both stages through a persistent sparse voxel feature map that grows with new observations and reuses cached local features. On EmbodiedOcc-ScanNet it reports 67.12% IoU and 58.40% mIoU for online mapping, rising to 67.79% IoU and 59.37% mIoU after offline refinement.
In briefA multi-objective evolutionary method for semi-supervised smart city classification that balances accuracy and feature use.
SummarySmart-city applications often have little labeled data but plenty of unlabeled data. This paper presents MOEA-SSL, a multi-objective evolutionary approach for lightweight semi-supervised tabular classification that jointly optimizes predictive quality and feature usage. Tested on three Smart City datasets, it achieves competitive performance against established lightweight SSL methods while significantly reducing feature usage compared with EA-SSL in all conditions. The resulting Pareto fronts also offer lower-feature alternatives with limited performance loss.
In briefBeyond Calibration: Checking How Reliable LLM Confidence Is Across Languages and Local Knowledge
SummaryThe paper examines whether an LLM's confidence before solving a problem (pre-solve) and after solving it (post-solve) reliably signals its uncertainty, testing English, translated Marathi benchmarks, and a new set of local Marathi factual questions. Pre-solve confidence is weakly calibrated, while post-solve confidence generally aligns better with correctness despite higher inference cost. Calibration largely transfers under translation but degrades substantially on local Marathi knowledge, and hallucinated answers keep elevated confidence after generation.
In briefTheoretical Analysis of Data Augmentation When Labels Are Noisy
SummaryThis paper theoretically analyzes when data augmentation (DA) helps and when it hurts under label noise, giving sufficient conditions that separate these regimes in both i.i.d. learning and domain generalization. It reveals that DA acts as an uncertainty amplifier, reducing model stability and generalization performance, and supports this with proposed learning and generalization bounds. The analysis does not depend on specific DA techniques and requires only mild assumptions, providing a unified theoretical foundation for understanding and designing DA methods under label noise.
In briefMemory as Code: Declarative Retrieval over Codified Long-Term Memory
SummaryThis paper proposes Memory as Code (MaC), a framework that unifies how LLM agents build and retrieve long-term memory using a shared code abstraction. Memory is built as a three-layer executable structure, and queries are translated into structured retrieval programs. On LoCoMo and LongMemEval, the authors report relative gains of up to 15.91% in LLM-judge accuracy over the strongest baseline, along with substantially lower token consumption.
In briefE8-QAT: Efficient 2-Bit Training Using Implicit Lattice Codebooks
SummaryThis paper proposes E8-QAT, a 2-bit quantization-aware training framework that uses the E8 lattice as an implicit eight-dimensional codebook, quantizing weight groups jointly through arithmetic projection without storing or searching an explicit codebook. On Qwen3.6-35B-A3B, E8-QAT matches full-precision model performance in average score across four tasks and outperforms the state-of-the-art baseline on IFEval and LiveCodeBench under the same data, schedule, and bit width. The 2-bit representation reduces peak memory and supports a larger batch size.
In briefVLA-Retain: Benchmark on How Vision Tuning Affects Other Conditions in Robot Models
SummaryThis paper introduces VLA-Retain, a benchmark for measuring how fine-tuning the vision module of vision-language-action (VLA) robot models affects performance outside the target condition. It compares five open VLA models with their base checkpoints across seven perturbation dimensions on the LIBERO-Plus dataset. The results show that target-domain gains do not predict retention, and off-target performance often declines, so cross-dimensional retention is proposed as a key criterion for adaptation.
In briefPanoGraphGS: Combining Panoptic Segmentation and Hierarchical 3D Scene Graphs via Gaussian Splatting
SummaryPanoGraphGS is a framework that combines panoptic 3D Gaussian representation learning with hierarchical 3D scene graph construction. It learns semantic and instance features on shared Gaussian geometry, producing panoptic predictions that stay consistent across viewpoints. The method supports novel-view panoptic segmentation and recovers a floor-room-stuff-object hierarchy with containment and spatial relations. The authors also release NaviGS, a real-world dataset for panoramic scene understanding. The paper reports better 3D scene graph object-node accuracy than the strongest evaluated baseline and state-of-the-art 2D panoptic segmentation among evaluated methods.
In briefOptimal transport for choosing key frames in long video understanding
SummaryMultimodal LLMs have limited visual context, so long videos must be reduced to a subset of frames before reasoning. This paper treats frame selection as a structured evidence-allocation problem and proposes OTFS, an optimal transport framework that maps multiple evidence sources onto frames. It uses a semi-unbalanced entropic objective with an efficient solver, OTFS-Sinkhorn, and a coverage-aware dynamic program to extract the subset in a training-free two-stage pipeline. Experiments on three long-video benchmarks show it consistently outperforms strong training-free baselines.
In briefEASformer: An Environment-Action-State Aligned Model for Industrial Generative Ranking
SummaryEASformer is a Transformer-based ranking architecture for an industrial app marketplace that separates environment context, user actions, and state. It builds an aligned context space shared by historical and candidate non-sequence tokens, and supports a masked-action self-supervised objective. On an offline benchmark it reaches an AUC of 0.89411, a 0.88% relative gain over the deployed baseline. Online A/B tests show relative eCPM lifts of 1.26% and 3.94% across two ranking scenarios.
In briefNeoWorld-3: Building Physically Interactive 3D Worlds from Code
SummaryThis paper presents NeoWorld-3, a coding-agent framework that reconstructs interactive 3D worlds from visual observations using tunable, executable programs. It provides a computational engine for geometry, part composition, differentiable refinement, and physics validation, and exposes these as tool-callable APIs for vision-language agents. The authors report state-of-the-art geometric fidelity and physical readiness on object- and scene-level benchmarks, and say manipulation plans made in the reconstructed twins ran zero-shot on a physical dual-arm robot.
In briefInfluence equals the target-response Jacobian times residuals
SummaryThe paper introduces the target-response Jacobian J, which measures how model predictions change when individual training targets are altered, complementing training-data attribution methods. It shows that sample-weight influence and J differ only by a per-example multiplier, which equals the training residual for canonical-link losses such as squared error and logistic loss. Across several datasets, J achieves a per-prediction R² of 0.92 to 1.00 against refitting, while influence's R² is negative.
WhyIntroduces target-response Jacobian and proves it differs from sample-weight influence only by residual multipliers; validated against refitting on several datasets, with impact mainly within attribut
In briefPattern Intelligence: Unsupervised Discovery of Systematic Underpayment and Denial Patterns in Healthcare Billing
SummaryThis paper presents Pattern Intelligence, a deployed system that finds recurring, correctable underpayment and denial patterns in healthcare claims without supervision. Claims are clustered by adjudication behaviour using a 130-dimensional representation, with payer and contract identity excluded from the distance metric and used only afterward for segmentation. A large language model describes each cluster for analysts. The authors evaluate cluster coherence and recoverable dollar value on two production tenants.
In briefAllocating limited retrieval evidence across patient timelines
SummaryClinical records repeat stable diagnoses across years, so similarity-based retrieval can miss single-entry events such as recurrences or therapy changes. The paper argues that longitudinal clinical retrieval should allocate a fixed evidence budget across time rather than rank purely by relevance. In a study of 13 queries over 20 patient records, a temporal allocation configuration achieved higher milestone slot recall than a global baseline, but it also retrieved far more evidence. Without expert reference timelines, the measurements describe retrieval behavior rather than clinical accuracy.
In briefPhonoBridge: Error-aware phoneme-based decoding of muscle signals into text
SummaryPhonoBridge is a two-stage system that decodes silently mouthed speech from electromyography (EMG) signals into text. A frontend converts EMG to phonemes, and a large language model reconstructs text from them, with extra training data created from LibriSpeech transcripts. A new augmentation method, Error-Profile-Guided Augmentation, makes the language model robust to frontend errors. On the Open Vocabulary silent-speech dataset it reaches 24.77% word error rate, beating the previous best of 28.27% on the same split.
In briefCovNav: Conformal Prediction for Deciding When Robots Reason Slowly in Object Navigation
SummaryThe paper addresses object goal navigation, where a robot must find a target object such as a bed in an unseen environment. It proposes CovNav, a training-free framework that uses conformal prediction to decide whether a fast on-device language model's judgment is reliable or whether slower chain-of-thought reasoning is needed. On Gibson and HM3D, CovNav improves success rate over the strongest zero-shot baseline by 0.8 and 1.7 points on HM3D v1 and v2, invoking slow reasoning for only 12.4% of queries, and was further validated on a TurtleBot 4 robot.
In briefHybrid momentum methods that combine long- and short-term gradient history for training deep learning models
SummaryThe paper introduces Hybrid Momentum (HyM), which blends a linearly weighted average of past gradients with a fast geometrically discounted momentum to retain long-term history while adapting to recent changes. It also proposes HyMUON, which applies Newton–Schulz orthogonalization to the hybrid direction. The authors report that HyM and HyMUON match or improve on strong baselines across image classification, machine translation, and language modeling. For LLaMA-1.3B pretraining on C4, HyMUON reaches a final validation loss of 2.777, versus 2.903 for Muon.
In briefReusing Assurance Evidence in AI Pipelines: Validity Rules and a Fail-Closed Release Gate
SummaryThis paper examines when governance check results can safely be reused in AI pipelines instead of rerunning every control after each change. It defines conditions under which signed evidence is valid for a release candidate and implements a reference gate using Ed25519 signatures that blocks eight of nine tested manipulation scenarios. Replaying 38,083 commits from 714 public DVC pipelines, declared-dependency reuse avoided 83.4% of stage executions but reused stale stages in 3.3% of commits. Adding static import closures reduced stale reuse to 0.4%.
In briefLearning fluid flow in complex shapes by splitting time steps into learnable parts
SummaryThis paper proposes Neural Operator Splitting, a framework that learns fluid dynamics while keeping the structure of fractional-step numerical schemes. Each time step is decomposed into physically interpretable substeps, with each operator learned separately from the solver's intermediate states, using a conditional-linear Fourier neural operator. The authors report lower rollout error than a single full-step neural operator in 2D and 3D flows, and inference roughly 10³ times faster per step than a Monte Carlo solver at comparable error levels.
In briefReusing Multimodal Agents: Efficient Evaluation of Sound Designs in Games
SummaryThis paper proposes a framework that uses multimodal learning and reinforcement learning to efficiently evaluate in-game sound designs. A multimodal encoder is pre-trained to map visual and auditory information into a shared latent space, and a reinforcement learning agent is trained on that representation. For each new sound design, only the encoder is fine-tuned while the policy stays fixed, isolating the effect of sound changes on agent behavior. Experiments on games from multiple genres suggest the method detects performance differences caused by degraded or altered sound and significantly reduces computation compared with retraining agents from scratch.
Thai Nguyen Van, Ibrahim Khan, Mury Fajar Dewantoro et al.
In briefA paper showing pairwise ranking comparisons can make neural networks learn regression and classification from less labeled data
SummaryThe paper proposes a pairwise comparison framework that recasts regression as aggregated relative rankings. On the UTKFace dataset (23k images), it generates over one billion training comparisons without extra supervision, and a lightweight ResNet10 trained from scratch reaches a 3.64-year MAE, surpassing prior models including large pretrained ones. On STL10 classification with 500 to 2000 labeled examples, pairwise supervision improves accuracy by up to 5.2% over a 10-way crossentropy baseline. The authors suggest it is especially useful when labeled data are scarce, such as in clinical settings.
In briefPredicting cognitive traits from brain MRI using a pairwise ranking neural network
SummaryThe paper introduces a pairwise convolutional ranking framework that learns from whole-brain T1-weighted MRI by making comparative predictions, and reports good performance with a small sample of 108 participants. Tested on four small-scale spatial abilities tasks, it predicted performance more accurately than standard regression and pointwise deep learning baselines. Iterative cortical-region removal identified brain regions supporting performance common to all four tasks as well as task-specific regions.
In briefWhen Can LLM Reasoning Reliably Reuse Intermediate Results?
SummaryThis paper studies when large language models can reliably reuse intermediate results across multi-part problems, and how much computation this requires. It gives conditions for correct completion and bounds the average number of attempts for local retries. Experiments on eight reasoning models show that keeping the first subproblem's answer or reasoning raises average success on the second subproblem.
In briefPLaCA: Learning Patch-Level Future Latent Change for Robot Action Generation
SummaryThis paper proposes PLaCA, a framework for vision-language-action robot models that directly predicts spatially structured patch-level future latent change from current multimodal input. It uses patch-level change-aware supervision and a future-to-action interface to feed predicted changes into action generation. PLaCA reaches a 79.0% success rate on LIBERO-Long, 5.0 points above the Action-only baseline, and raises CALVIN's average sequence length from 2.784 to 3.564 while updating only 2.95% of parameters in the second training stage.
In briefPACE-MoE: Adaptive Domain Sampling and Domain-Specific LoRA for Medical Visual Grounding
SummaryThis paper proposes PACE-MoE, a framework for medical visual grounding that aligns clinical descriptions with spatial evidence across multiple imaging domains. It dynamically assesses each domain's training value from current-policy feedback and adjusts domain coverage through admission, replay, and final balancing, while domain-specific LoRA adapters route updates to the matching imaging domain. Experiments on seven medical imaging domains report a 3.64% relative improvement in mean box IoU over the Vanilla baseline. Adaptive sampling alone, under matched adapter architectures, further improves mean box IoU by 2.07%.
In briefDoes multi-agent collaboration help when agents serve users? A test in dual-control settings
SummaryThis paper tests a previously derived rule about when multi-agent systems beat single agents, using a benchmark where a user participates in the conversation under a written policy. On τ²-bench's airline, retail and telecom domains, it compares single-agent and multi-agent systems across six models over about 7,950 simulated conversations. Under a fixed simulated user, the multi-agent system performed worse in most valid settings and showed no benefit. The authors conclude that single-agent designs should be the default for user-facing agents, and multi-agent designs need compute-matched evidence from the target setting.
In briefBetter Corpus Quality, No Retrieval Gain: Ablating Preprocessing and Refinement in a Speech-to-QA Pipeline
SummaryThis paper ablates a production pipeline that converts conference speech into question-answer pairs, switching a transcript rewrite stage and a pair refinement stage on and off to produce four corpora from 40 Korean medical and finance sessions. Corpus quality scores rise as stages are enabled, but downstream retrieval accuracy does not follow the same ranking; the transcript rewrite alone lowers accuracy by 0.061. The authors conclude that higher rubric scores did not predict how useful a corpus was to the models tested.
In briefBalancing What LLMs Say and What They Know About Opinions
SummaryThe paper observes that opinion distributions verbalized in LLM output align better with human opinions than token log-probability distributions, yet the two differ in their most likely answer for about 37% of World Values Survey questions. The authors argue that verbalized distributions reflect model knowledge while log-probabilities govern actual behavior, so evaluating one does not describe the other. They propose a learnable combination that follows the behavior distribution's most likely answer while preserving the knowledge distribution's shape, controlled by one hyperparameter. Experiments on three datasets and six models show improved behavioral consistency with a slight rise in distance to human distributions.
In briefCUA-Recovery: A Benchmark for Error Recovery in Long-Task Computer-Use Agents
SummaryThis paper introduces CUA-Recovery, a benchmark that tests whether computer-use agents can detect and recover from errors during long, multi-application desktop workflows. It is built with the ReRail framework and includes a test suite of 2,250 human-verified erroneous states and a training suite of 8.6K trajectories. Across six agents, the best Error Awareness Rate reached 44.0% and the best Pass@3 reached 25.5%. Fine-tuning Qwen3.5-35B-A3B on the training suite improved both error awareness and task completion.
In briefBCT: Bayes-Consistent Tempering for Flow Language Models, a method that builds temperature into generative flow language models
SummaryThe paper argues that the usual post-hoc temperature scaling in flow language models disrupts both the endpoint distribution and the intermediate noisy-state distribution, and mismatches the transport dynamics of the untempered flow. It proposes Bayes-Consistent Tempering (BCT), which makes temperature explicit in the generative flow's distributions and dynamics, and also derives training-free variants for existing checkpoints. Reported results show BCT improves GSM8K generation across broad temperature ranges, extends gains to Sudoku, and is more robust than the training-free variants at very low temperatures.
In briefEnsemble-Aware Structural Encoding (EASE): Better Peptide Prediction Using Multiple Shapes
SummaryThe paper introduces EASE, a framework that encodes peptides using a set of conformational structures rather than a single structure. It selects central conformers representing major structural states and fuses them with a permutation-invariant operator. On six peptide property-prediction tasks under leakage-controlled splits, EASE consistently outperformed matched single-structure baselines, with a mean absolute gain of about 3.0% across two structural encoders. Performance kept improving as the number of conformers increased up to K=20.
In briefSCODEL: Shared canonical definitions linking food terms to ontology concepts
SummaryThis paper proposes SCODEL, a model-agnostic method that gives contextual food mentions and ontology concepts referring to the same entity a shared canonical definition, used as a semantic normalisation layer for existing entity linking models. The authors also introduce ReciLinkD, the first benchmark for contextual food ontology entity linking in recipes. Experiments report improvements for three linking models and generalisation to a biomedical ontology, and the method is used to build an ontology-grounded recipe knowledge graph.
In briefWhen Feedback Helps: Auditing Learned Voltage-Clamp Experimental Design
SummaryThe paper audits learned controllers on two synthetic voltage-clamp tasks. Five recurrent policies repeated one fixed protocol across all 1,200 off-grid rollouts, and masking current changed no decisions. In a four-hypothesis task, policies used the first current reading but tied a correctly specified Bayesian controller, and the authors state these cases show neither biological discovery nor general RL superiority.
In briefA Two-Way Street: Survey of How Code-Mixing and Native Languages Transfer in Text and Speech
SummaryThis systematic survey examines transfer between code-mixed (CM) text and speech systems and their native languages in both directions. The authors organise prior work into a taxonomy of transfer direction, route and modality. They find most transfer runs from native languages into the CM setting, while the reverse direction is younger and mainly helps by aligning the two languages. The paper also identifies open challenges, including how to mix native and CM data in training and how to evaluate CM text written in different scripts.
In briefPAR: Progress-Aligned Recurrence for Longer State Tracking
SummaryThis paper introduces Progress-Aligned Recurrence (PAR), which aligns each iteration's prediction of a looped language model with the state after the corresponding number of operations and rescales recurrent states to their first-iteration norms. On swap-S5, a Qwen2.5-0.5B model trained on at most 32 moves reaches high full-state accuracy at 512 moves, about 16 times the training length. Ablations show that removing norm anchoring or using only final-answer supervision reduces this extrapolation ability.
In briefTriGeoBench: A Multimodal Geometry Benchmark with Figures and TikZ Code
SummaryThis paper introduces TriGeoBench, a geometry-focused multimodal benchmark built from Chinese middle- and high-school math exams, containing 7,000 question-answer pairs and 12,235 figures with TikZ code. It is designed to test whether multimodal large language models reason over geometric structure, using image-to-TikZ and multimodal reasoning evaluations. The results show persistent difficulties with fine-grained structure and image-text alignment.
In briefFrom Outcomes to Steps: Reliable Success-Aware Critic Learning for Long-Horizon Language Agents
SummaryThis paper studies reinforcement learning for long-horizon language agents, where only trajectory-level supervision is available and standard critics struggle to judge whether intermediate actions raise the chance of eventual success. It proposes SCALE, an actor-critic framework that adds a Success Value prediction, the probability of eventually completing the task from each interaction prefix. Changes in this value are converted into multi-step action credit via Success-GAE and weighted by the predictor's estimated reliability. According to the authors, SCALE improves task performance over compared RL baselines in most settings across five benchmarks and three backbones, with success-rate gains of up to 4.1 points.
In briefElectronics-free glove demo collection using one head-mounted camera
SummaryThis paper presents a method for collecting robot demonstration data in which a single head-worn egocentric camera reads printed fiducials on an electronics-free glove and paper tiles around the workspace, recovering the gripper's 6-DoF pose and jaw width. At deployment, the same camera is mounted on a low-cost arm beside the robot. Labels are accurate to a median of 6.5 mm and 1.6 degrees against robot encoders, and a policy trained on them pays 3% in validation loss. Adding 194 glove demonstrations to 48 teleoperated episodes raises rollout success from 10/25 to 17/25.
In briefA radiation-hardened neural network that repairs itself after bit-flip errors
SummaryThe paper proposes a runtime method that repairs quantized neural network weights after bit flips on commercial off-the-shelf hardware, without storing extra replicas. It extends fault-aware training with an alignment penalty and reconstructs corrupted values from undamaged ones. Across six regression benchmarks, the method extends the operational scrub interval more than classical SECDED error correction does.
In briefBuilding robust, efficient LiDAR perception for self-driving cars
SummaryThis paper studies LiDAR-based 3D object detection for autonomous driving, aiming to keep detection accurate when point clouds are corrupted while meeting latency and GPU memory limits. The method simulates extreme-weather corruptions, trains detectors by ordering samples by perturbation sensitivity, and narrows the network with density-based sampling settings. On KITTI with PointRCNN, Moderate AP under corruption improves by 3.32 points, while mean latency drops 19.8% and peak GPU memory drops 53.9%.
WhyJointly targets corruption robustness and latency/memory constraints via sensitivity-ordered training and density-driven pruning; evaluated against original detectors on two datasets.
In briefLearning interactions between different molecular parts in chemical systems to predict solar cell properties
SummaryThe paper proposes a data descriptor for complex chemical systems and a graph neural network that learns interactions between organic and inorganic compounds. Existing methods mostly assumed interactions only between structures of the same type. The method is applied to predict physical properties of hybrid solar cell materials. It reportedly achieves state-of-the-art accuracy in predicting band gaps for 1,682 such materials.
In briefABench: Adversarial Benchmark for LLM Alignment with Taxonomized Chinese Values
SummaryThis paper introduces ABench, a handcrafted benchmark for testing whether large language models hold stable stances on taxonomized Chinese values. It is built on a hierarchy of 158 value subcategories and contains 2,895 open-ended questions that use role-playing and logical traps. Experiments on 16 mainstream LLMs found that most models are significantly vulnerable to these adversarial value attacks, scoring below 60.(wait)
In briefAuditing Evaluation Evidence for Self-Improving AI Claims
SummaryThe paper codes 114 arXiv studies from a frozen October 2026 search, 90 of which describe their systems as recursive self-improvement. It finds that product-level evaluation is common, while evidence on repeatability of the improvement loop and on whether improved procedures really outperform earlier ones is much rarer. The authors propose a claim-indexed reporting standard linking each claim to the evaluation design needed to test it.
In briefWaveSpec: Wavelength-Conditioned Spectral Coherence Adaptation for Multi-Spectral Iris Attack Detection
SummaryThis paper proposes WaveSpec, a framework that detects presentation attacks on multi-spectral near-infrared iris images. Each band is encoded with DINOv2, then a set-conditioned adapter updates band features using cross-band consensus, per-band deviation, and wavelength identity before classification as bona fide or attack. Evaluated on the MSIrPAD benchmark under a cross-artefact protocol, it outperforms SpectraIrisPAD on average error metrics. Ablations indicate the adapter and spectral residual drive the gains, and it remains effective with partial band sets.
In briefAsk to Learn: A study of improving reinforcement learning beyond offline data by querying a collaborator for extra information.
SummaryOffline reinforcement learning is limited by the coverage of its data, while online exploration can be costly, unsafe, or infeasible. This paper proposes query-based collaborative reinforcement learning, in which a learner starts from a fixed offline dataset and adaptively queries state-time pairs, with a collaborator sampling actions from a fixed, possibly suboptimal policy and a simulator returning rewards and next states. For finite-horizon tabular problems, the authors derive an instance-dependent information-theoretic lower bound on the additional queries needed and design an adaptive query-allocation algorithm whose complexity nearly matches this benchmark under suitable conditions. Experiments on synthetic and semi-synthetic environments demonstrate its empirical efficiency.
In briefDART: Extending a Shape Model Across Organs and Imaging Types
SummarySegmentation models often produce anatomically implausible shapes, such as broken vessels, on hard-to-segment structures. The authors propose DART, which learns a shared shape representation from segmentation probabilities and uses anatomical text descriptions to correct them without medical images as input. Tested on 13 public benchmarks, the approach scales one shape representation from 50 to 377 anatomical classes without new voxel-level annotations for correction training.
WhyA text-guided diffusion model corrects segmentation probabilities across 377 classes and CT-to-MRI transfer, validated on 13 benchmarks, showing strong novelty and evidence.
Koushik Swarna, Pedro R. A. S. Bassi, Lam Ngo et al.
In briefHiQGS: Hierarchical spatiotemporal querying for feed-forward 3D Gaussian scene reconstruction
SummaryThis paper addresses feed-forward 3D Gaussian scene reconstruction from sparse monocular frames, where observations vary in both viewpoint and time. It proposes HiQGS, which uses a Hierarchical Spatiotemporal Encoder to organize multi-resolution visual features by image location, observation time, and camera pose. A Time-Adaptive Decoder then uses target-timestamp queries to retrieve relevant evidence and predict the Gaussian scene state. Experiments on static and dynamic benchmarks report strong rendering quality and geometric accuracy.
In briefFrom Encoding to Generation: Toward Empathetic Language Models
SummaryThis paper examines how Llama-3.2-3B-Instruct internally represents, attends to, and carries emotional information into its responses, using 10 emotion categories from EmpatheticDialogues. Layer-wise probing shows emotion categories are most decodable in middle layers, and emotion-related user tokens receive more attention during generation. However, alignment between the target emotion and the generated response is lower than for the input representations.
Samhita Bollepally, Takashi Yamauchi, Marissa West
In briefSelf-Decomposing Diffusion for Code: Planning Program Splits Before Parallel Implementation
SummaryDiffusion language models can generate code in parallel and quickly, but complex programs have interdependent parts that parallel decoding can break. This paper proposes SDDC, a framework that first generates a contract specifying how a program is decomposed and how its units depend on each other, then uses it to guide coordinated parallel implementation. Across five benchmarks and two 7B diffusion backbones, the authors report better accuracy-latency trade-offs, with pass@1 raised by up to 31.1 points and decoding 7.5 to 13.7 times faster than the base models.
In briefMeta Expert: Sharing Experts Across Layers to Make MoE Models More Efficient
SummaryThis paper addresses redundancy in Mixture-of-Experts (MoE) models, where experts in different layers store similar content. It proposes Meta Experts shared across layers and compares them with the standard per-layer Local Experts. Models using Meta Experts show lower parameter redundancy, lower cross-entropy loss, and higher downstream accuracy. The authors also argue theoretically that Meta Experts offer more flexible residual pathways, improving information retrieval efficiency.
In briefWORTH Benchmark: Testing How Models Find Clues in Mystery Stories
SummaryThis paper introduces the WORTH Benchmark, in which models read mystery narratives without clue labels, pick out key clue sentences, identify case participants, and explain their reasoning. A pilot across two episodes ran 70 formal runs, 67 of which were scorable. In Episode 02, 31 of 37 scorable runs selected the same incorrect participants, while suspicion-pointing and exoneration-pointing clue sentences were chosen 25 and 0 times respectively. The authors stress this is a two-episode pilot that does not establish model rankings or general mechanisms, and that human records are not yet cleaned and no human-model comparison is reported.
In briefCausal spatiotemporal state modeling for monocular 4D scene reconstruction
SummaryThis paper proposes a feed-forward framework that learns causal 4D visual geometry from monocular videos. Its Causal Spatiotemporal Mamba propagates geometry-aware states across frames in a directional way using two complementary scans. Unified Motion Supervision builds a forward motion field from camera and depth predictions to constrain transitions between frames. The authors report state-of-the-art results on video depth, camera pose, and optical flow estimation, with reduced temporal drift.
In briefEditing scene text using only real images via self-paired training
SummaryThis paper studies Scene Text Editing, which changes the text in an image while keeping its original style. Instead of relying on synthetic paired data, the authors disrupt each real image with cropping, shuffling, and masking to create self-paired training samples, and build a baseline called TextRIRO on a conditional diffusion model. The paper reports that combining character-column shuffling with cropping works best, and uses the method to create TextRIRO-3M, a large dataset for long-text recognition whose training improves recognition accuracy.
In briefRandom-Token Replacement Under Repeated Data: Costs, Benefits, and Ten-Corpus Evidence
SummaryWhen unique pretraining data is scarce and models see the same tokens repeatedly, random-token replacement, an input-corruption augmentation, is reported to lower validation loss. The authors use a controlled source with known entropy to show that replacement shrinks predictions toward the corpus average, and they decompose validation loss into approximation, estimation, and memorization terms. Across ten real corpora, replacement lowers final loss on nine and harms only the most redundant corpus, program-generated JSON.
In briefOrthogonal Residual Learning Keeps Tire Stiffness Adaptation Working in Sideslip Estimation
SummaryThe paper addresses a conflict in vehicle sideslip estimation, where online cornering stiffness adaptation and learned model correction compete for the same filter innovation. The authors build a corrector orthogonal to the stiffness sensitivities, enforced through a null-space parameterization. In simulation and public real-vehicle data, the orthogonal corrector keeps its stiffness response and lowers sideslip error relative to the unconstrained corrector.
In briefBeyond Model Weights: Fairness Depends on Sampling in Diffusion Language Models
SummaryThis paper studies social bias in diffusion language models (DLMs), an emerging alternative to autoregressive language models. Testing LLaDA, LLaDA-1.5, and Dream on Open-BBQ, the authors find that both the measured bias magnitude and the ranking of models change substantially with sampling settings such as block length and remasking strategy. Debiasing methods developed for autoregressive models transfer and reduce bias, but sampling dependence persists afterward, so DLM fairness cannot be assessed independently of the sampling configuration.
Koshiro Saito, Daisuke Oba, Danushka Bollegala et al.
In briefTeCoP: Faster Diffusion Language Model Decoding via Tentative Context Pipelining
SummaryDiffusion large language models predict several token positions per forward pass, but block-wise decoders generate serially and repeatedly recompute stable predictions. The paper introduces TeCoP, a training-free decoder that uses staggered block-parallel pipelining and draft-and-verify, reusing stable but uncommitted predictions as provisional cross-block context. Commitment is governed by confidence and stability checks. Across four LLaDA and Dream models on GSM8K, Minerva MATH, and MBPP, TeCoP reduces forward evaluations by up to 87% and reaches up to 7.50x throughput without KV caching and 11.77x with caching, with cached accuracy within one point of the original decoder.
Jui-Chien Tsou, Hsuan-Chi Liu, Wei-Chin Wang et al.
In briefDuplexLedger: A Benchmark for Transactional Memory in Full-Duplex Speech Agents
SummaryThis paper introduces DuplexLedger, a bilingual benchmark of 1,094 persistent sessions across four full-duplex speech scenarios, with event-level labels for writing, updating, ignoring, and retaining memory. Evaluating five full-duplex models, the authors find that none combines strong interaction-boundary control with coherent final-state memory. They also propose ACT-Mem, an 8.92M-parameter sidecar, which raises latest-value accuracy on held-out sessions but slows interruption stopping.
In briefFactual access states in diffusion language models for knowledge editing and retrieval
SummaryThis paper studies how diffusion language models (DLMs) store and retrieve facts through an internal representation the authors call the factual access state. They find that in DLMs this state is assembled early, in the first attention layer and early feed-forward layers, and varies with the relation's wording and query direction. Building on this, the authors propose Multi-View Edit, a knowledge editing method, and show that the state can serve as a training-free retrieval vector for two-entity queries.
In briefWhy Static Probes Fail: Deception Detection Depends on Context
SummaryThis paper studies how to detect deceptive behavior in large language models, which can appear benign while pursuing misaligned goals. The authors show that single-direction linear probes fail to generalize across diverse deception scenarios, and propose representing deception as a multi-dimensional subspace. They add a domain-adaptive gating mechanism that composes these directions with few-shot adaptation at test time, reaching an average AUROC of 0.97 and 94% recall at 5% FPR using only 10 examples.
In briefFrom POMDP Theory to Deep RL with Particle Filters: A Practical, Theoretically Grounded Method for Partially Observable Decision-Making
SummaryThe paper proposes an end-to-end algorithm for reinforcement learning in partially observable environments. It estimates hidden states with sequential Monte Carlo (SMC) particle filtering, updates a latent dynamics model online, and performs approximate online planning for a fixed number of steps before bootstrapping with a state-value critic (a QMDP approximation). The authors prove an end-to-end regret bound that decomposes error into model estimation, particle filter, Monte Carlo rollout, and planning-depth-dependent structural bias terms. The method is instantiated in an MPC-based deep RL framework and evaluated on several environments.
In briefWhen LLMs Manipulate: Measuring Behavioral, Social, and Temporal Effects
SummaryThis paper proposes a framework for measuring behavioral vulnerability, meaning how susceptible users are to manipulative AI responses. Using the DarkPatterns-LLM benchmark, the authors generate manipulative stimuli targeting authority, urgency, and social proof biases, and test them with 120 participants. Manipulative framing significantly changed trust calibration and compliance (p < 0.01), and the effects also spread across social contexts.
Sadia Asif, Israel Antonio Rosales Laguan, Shumaila Asif et al.
In briefRaspAssistant: An AI Agent Framework for Automating Adsorption Simulations in Porous Materials
SummaryThis paper presents RaspAssistant, a hierarchical multi-agent framework that coordinates adsorption studies in porous materials, from force-field preparation through simulation, validation, and separation analysis. In twelve workflow-reproduction runs across four adsorption and separation benchmarks, overall scores ranged from 8.50 to 10.00. In an eight-case force-field extraction benchmark, the mean weighted score was 0.967. The system also diagnosed and refined an acetylene force field and screened 197 pure-silica zeolites using a CO$_2$ force field validated against MFI experiments.
Marko Petkovi\'c, Susanne Gooijer, Jos\'e Manuel Vicent-Luna et al.
In briefTAME: Text-anchored multimodal sentiment analysis with uncertainty-guided expert routing
SummaryThis paper proposes TAME, a model for multimodal sentiment analysis of user-generated videos. Text serves as an always-active shared expert, while audio-text and video-text cross-modal branches act as dynamically routed auxiliary experts. Lightweight Prophet Heads estimate per-sample uncertainty for each modality, which guides reliability-aware expert aggregation. The authors report state-of-the-art results on most metrics across CMU-MOSI, CMU-MOSEI, and CH-SIMS, with up to 78.7% fewer parameters and 41.2% fewer FLOPs than the mixture-of-experts baseline.
In briefBayesian neural networks with Dirichlet process priors on the data distribution for randomized value functions in deep RL
SummaryThe paper introduces a class of Bayesian neural networks that quantify epistemic uncertainty by placing Dirichlet Process priors on the data-generating distribution rather than on network weights. Building on this, the authors propose DP-DQN, a reinforcement learning algorithm that maintains posterior distributions over value functions and performs Thompson sampling. The method combines Bayesian Bootstrap reweighting with generative prior hypotheses injected through the DP base measure, aiming for deep exploration using a single neural network instead of an ensemble.
In briefTwo Is Enough: Binary Pretraining Priors Also Support Multi-Class In-Context Tabular Prediction
SummaryThis paper studies pretraining for tabular foundation models, showing that capping the class count of synthetic classification tasks at two still yields a model that performs zero-shot K-way classification. On 38 TabArena classification datasets, the binary two-stage recipe reaches 1590.1 Elo, above the official default at 1568.0. On 60 real-world credit-scoring institutions, the same checkpoint improves test AUC on 44 of them. The authors argue that the number of classes is a tunable hyperparameter rather than an architectural constraint.
WhyShows binary-only pretraining enables zero-shot multi-class prediction, validated on 38 benchmark datasets and 60 institutions against the default.
In briefTRCM: Topology-Aware Route-Consistency Modeling for Robust Neural Vehicle Routing
SummaryThis paper proposes TRCM, which explicitly models whether customer pairs belong to the same route in autoregressive neural solvers for the Capacitated Vehicle Routing Problem. The predicted same-route probabilities guide customer selection as a route-consistency signal. On CVRP50 and CVRP100, the optimality gap decreases from 3.67% to 3.28% and from 2.11% to 1.99%, respectively.
In briefAudiPsy: Evidence-based multimodal reasoning for assessing depression from spoken counseling conversations
SummaryThe paper proposes AudiPsy, an omni-LLM that assesses depression from spoken counseling dialogue by combining transcripts, textual and acoustic emotions, and their incongruence. It uses an Observation-Assessment-Summary reasoning paradigm that incrementally organizes evidence and maintains a dialogue summary across turns. The authors also build AudiPsy-Counsel, a synthetic counseling corpus, and train with a three-stage curriculum. Experiments on DAIC-WOZ report better assessment performance and reasoning quality than specialized depression detectors and general speech/omni-LLMs, with supplementary results on EATD-Corpus.
In briefLearned Readout Tokens for Measuring How Confident LLMs Are
SummaryThis paper proposes a new way to estimate how reliable a large language model's output is. Instead of reading hidden states from existing vocabulary tokens, it learns a new readout token and trains a probe on its hidden states to predict whether the answer is correct. The authors report that this method, ScratchProbe, outperforms the existing-token version across several datasets and LLMs, and that a multi-token extension gives further gains. Applied to a state-of-the-art probe, the learned readout improves AUROC on every backbone, both in and out of domain.
In briefAnomalyJev: Checking Image Anomaly Explanations by Verifying Individual Claims
SummaryMultimodal LLM-based image anomaly detectors can produce text explaining their predictions, but judging these explanations reliably and cheaply is hard. AnomalyJev extracts verifiable claims from each explanation and checks them against ground-truth annotations to score faithfulness and completeness, without generating tokens. The authors built a human-annotated set of 640 explanations with over 1,600 expert ratings. AnomalyJev agrees with human annotators more than most open MLLM judges and approaches flagship judges while using far less runtime and resources.
In briefSearching Solar System Surfaces for Signs of Artificial Design
SummaryThis paper extends technosignature searches beyond radio signals to planetary surfaces, looking for non-random, intentional patterns on terrain similar to long-lasting human-made landscapes. It argues such traces could persist for hundreds of millions of years on geologically inactive bodies like the Moon and Mars. The authors highlight high-resolution imaging datasets and ML-based anomaly detection as making this search timely, and outline supervised and unsupervised modeling approaches. They suggest combining these with LiDAR, ground-penetrating radar, and reflectance spectroscopy.
In briefCityPaint: Generating City-Scale 3D Textures Using 2D Diffusion Models
SummaryThis paper proposes CityPaint, a framework that uses 2D diffusion models to generate globally consistent textures for city-scale 3D environments without requiring city-scale 3D training data. It combines geometric priors with a top-K camera baking strategy for iterative texture projection. A decoupled refinement mechanism separately optimizes terrain and building textures to reduce artifacts from 2D-to-3D projection. Experiments report better visual quality and cross-view consistency than strong baselines.
In briefHypoNet Nankai PS: fast earthquake hypocenter locating for the Nankai Trough using neural networks with P- and S-wave 3D models
SummaryThis paper extends HypoNet Nankai, a neural-network tool for locating earthquakes in the Nankai Trough, so it can use both P- and S-wave travel times. The authors build an S-wave velocity model from the P-wave model and revise the network training to handle steeper travel-time gradients. Benchmarks against the fast marching method show lower root-mean-square deviations for both wave types. Applied to 223 earthquakes recorded by DONET stations off the Kii Peninsula, the hypocenters largely agree with the JMA catalog beneath the network, while offshore estimates are shallower and lie in the oceanic crust and uppermost mantle.
Ryoichiro Agata, Ayako Nakanishi, Dan Bassett et al.
In briefTesting Small Recurrent-Depth Language Models on Many Clinical Text Tasks
SummaryThis paper tests whether small recurrent-depth language models can handle diverse clinical text tasks, including information extraction, classification, and question answering, using one jointly fine-tuned model. The authors fine-tuned several models on 9 datasets and evaluated them on 4 additional held-out datasets, comparing them with standard transformers. Ouro-1.4B-Thinking outperformed Qwen3-1.7B on 11 of 13 datasets, and Ouro-2.6B-Thinking was competitive with Qwen3-8B on most tasks.
In briefAnchorRM: Parallel Batch-Internal State Reconstruction for Temporal Graph Networks
SummaryThis paper addresses a problem in memory-based temporal graph neural networks, where nodes appearing multiple times in one training batch are evaluated with outdated memory. It proposes AnchorRM, which keeps the pre-batch node memory as a fixed anchor and predicts a residual from earlier same-batch interactions to reconstruct an event-specific memory for each interaction. Because these reconstructions depend only on the anchor and causal history, they can run in parallel. Experiments on four temporal graph datasets and three backbones report an average 6.16x training speedup (up to 10.05x) and 28.67x batch scaling (up to 40x) with competitive predictive performance.
In briefCan Native Safety Travel? Transferring Safety-Pretrained Knowledge Between Models via On-Policy Distillation
SummaryThe paper introduces the native safety reuse problem: whether safety acquired during pretraining can be transferred from a small safety-pretrained teacher to standard-pretrained student models. Using on-policy distillation with cross-tokenizer alignment, the 1.7B student's attack success rate rises only 5.00% after fine-tuning, versus 11.58% for safety post-training. Its average utility of 48.20% also exceeds the teacher's 27.42%.
In briefLower bounds for smooth nonconvex zeroth-order optimization and the cost of determinism
SummaryThis paper studies how many exact function-value queries are needed to find approximate stationary points of smooth nonconvex functions under a low curvature trace (effective rank) assumption. It shows randomized methods need and achieve Ω(ET) and O(ET) calls, respectively, which is optimal. Deterministic methods require Ω̃(d·min{T, √(ET)}) calls, and adaptive subspace search matches this bound, so their cost grows linearly with dimension d.
WhyProves matching upper and lower bounds showing deterministic zeroth-order methods pay about d times more calls than randomized ones under low curvature trace.
In briefTime-aware, interpretable self-attention for detecting task switches from wearable event data
SummaryThis paper studies how to detect when a person switches from one task to another using discrete behavioural events recorded by wearable devices. The authors propose a time-aware self-attention framework that incorporates each event's onset timestamp and duration, encoding timing with a parameter-efficient kernel-based representation. On a semi-naturalistic wearable dataset, the model outperforms existing event-based approaches and matches or exceeds statistical and deep-learning models using a more compact input, with improved interpretability via integrated gradients.
In briefSynCalib: Using controllable synthetic data to fix LiDAR-camera calibration shortcuts
SummaryThis paper addresses online LiDAR-camera calibration for robots, where deep learning methods often overfit to fixed nominal extrinsic parameters because their training data lacks extrinsic diversity. The authors propose a data generation framework that uses omnidirectional image datasets with depth ground truth to synthesize multi-modal calibration data under strict extrinsic constraints. Experiments show the generated data improves the generalization and robustness of typical learning-based calibration networks. Models trained only on this synthetic data also transfer zero-shot to various real-world camera-LiDAR configurations.
WhyProposes extrinsic-controllable synthesis from omnidirectional depth datasets addressing dataset shortcuts; validated across several calibration networks and real-world zero-shot transfer, giving fair
In briefBEACON: A framework for detecting anomalies in cloud systems by combining multimodal data and asynchronous propagation
SummaryLarge cloud systems produce metrics, logs, and dependency graphs that reflect their health, and anomaly detection models analyze these to find abnormal states. The paper notes that existing methods overlook the semantic gap between data types and asynchronous spatio-temporal dependencies across system instances. It proposes BEACON, which includes a Dynamic Interpretable Multimodal Aggregator and an Asynchronous Spatio-Temporal Propagator. Experiments on three real-world cloud system datasets show BEACON outperforms state-of-the-art baselines in anomaly detection.
In briefEviDial: Benchmark Checking Whether Correct Depression Predictions Have Dialogue Evidence
SummaryThe paper introduces EviDial, a benchmark of 382 depression dialogues and 2,744 participant-symptom cells that separates agreement with benchmark labels from whether the dialogue actually supports the assessment. Evidence references are calibrated from multiple judgments using statistical measurement methods. The authors find that 82.8% of cells lack sufficient textual evidence, and 69.2% to 96.9% of predictions counted correct by labels are unsupported by sufficient dialogue evidence.
WhyIntroduces a new evaluation distinguishing label agreement from evidence support, calibrated with multiple psychometric methods, revealing many correct predictions lack evidence.
In briefLess Is More: Concise Retrieval and Critic Re-ranking for Composed Image Retrieval
SummaryThis paper addresses zero-shot composed image retrieval, which finds a target image from a reference image and modification text. The authors propose Less-is-More CIR, a training-free two-stage framework. The first stage generates concise target captions that keep only the most discriminative modification semantics to reduce retrieval noise. The second stage uses a critic re-ranking module to evaluate candidates for modification consistency and reference preservation, and the authors report extensive experiments showing its effectiveness and robustness.
In briefWhat drove the alarm score? Causal attribution with exact computation and scarce-incident estimation
SummaryThis paper addresses alarm diagnosis in industrial systems and scientific instruments, where incident data is scarce. It formulates the change in an alarm score as a Shapley game over hybrid structural causal model states, treating structural functions and exogenous laws as separate players. For ordered linear SCMs and quadratic scores, a causal moment recursion combined with quadrature computes the full Shapley vector exactly in polynomial time. To estimate this from few incidents, the authors adapt a reference model using cross-fitted estimates, with attribution-error bounds and top-k recovery guarantees, and experiments show scaling to hundreds of variables.
In briefAutomated design of multi-agent organizations that learn from execution
SummaryThis paper proposes Automated Multi-Agent Organization Design (AMOD), which treats the organization of agent teams, including their roles, coordination, and decision-making, as the object of automated design. It uses an execution-driven loop in which a designer assigns responsibilities and rules, then revises them based on task outcomes and interaction traces. The authors also propose testing discovered responsibilities by removing, replacing, or reassigning them under matched resource budgets.
In briefJudging the Justification: Supervising Explanations for Generative Reward Models
SummaryGenerative reward models produce written rationales for their preference judgments, but a correct verdict does not guarantee a sound rationale. The authors propose DeCert, which assesses the validity and verdict consistency of generated rationales without reference explanations. It turns these assessments into preference supervision that accounts for both verdict correctness and justification quality. Experiments across two model families and several benchmarks report consistent gains over strong generative judge baselines and reduced positional bias.
In briefMemDev-Bench: Does Memory Help Code Agents? A Benchmark Using Real Repository Histories
SummaryThis paper introduces MemDev-Bench, a benchmark for evaluating how memory systems affect code agents working on real repository histories. Each task is paired with relevance-annotated earlier tasks from the same repository, built through an automated pipeline that yields 4,869 historical tasks for 555 base tasks. The study finds that memory mainly improves fine-grained localization and editing, but it can also cause failures on solvable tasks, especially when history is unrelated, and that providing raw historical trajectories reduces agent steps more than memory systems do. The authors suggest future memory systems should preserve process information and reassess memory applicability against the current task context.
In briefOmniVLoc: Topo-Metric Visual Localization for Omnidirectional Cameras
SummaryThis paper presents OmniVLoc, a visual localization system for omnidirectional cameras based on a topo-metric framework. It combines global place recognition and monocular depth estimation, then solves a scale-aware pose graph and applies spherical bundle adjustment to recover metric poses for each query frame. A hierarchical verification mechanism filters erroneous outputs to improve robustness. Experiments on public and custom datasets report superior accuracy and reliability, and the source code and dataset will be released upon publication.
In briefGMSI-Net: EEG-centered multimodal emotion recognition with shared and modality-specific representations
SummaryThis paper presents GMSI-Net, a framework that combines EEG with other physiological signals to recognize emotions. It uses a multiscale encoder for EEG patterns, a branch that keeps modality-specific cues, and a branch that learns shared representations with a contrastive loss. Evaluated on DEAP and MAHNOB-HCI with within-subject ten-fold cross-validation, it reports high valence and arousal accuracy. The authors conclude that combining EEG with complementary physiological information improves emotion recognition.
In briefDo Health AI Assistants Agree Too Readily? A Benchmark for Urgent Self-Diagnosis
SummaryThis paper introduces SelfDx, a paired benchmark testing whether AI models wrongly agree with users' self-diagnoses in health questions where delayed urgent care could be dangerous. It contrasts 100 pairs of caution cases, where a serious alternative remains possible, with agreement cases, where the user's evidence rules it out. Across 27 frontier models, the authors find they often inappropriately agree when caution is needed, with the best model reaching only a 77.0% appropriate caution rate on time-critical conditions.
Marilyn Zhang, Tianfeng Chen, Patricia M Pechter et al.
In briefHealBench: A benchmark for judging whether AI-generated images feel psychologically comforting
SummaryThis paper introduces HealBench, a fine-grained benchmark of 30,000 image-text pairs for assessing whether text-to-image generation and image editing outputs feel psychologically comforting. Grounded in cognitive evaluation theory, it defines psychological comfort as an affective response to restorative and emotionally supportive visual cues. Testing 14 representative image generation models on the HealBench-Eval set, the authors find that strong semantic alignment and aesthetics do not guarantee emotionally restorative images.
In briefUnsupervised adaptation of vision-language models via progressive subset selection and cascaded consistency
SummaryPretrained vision-language models lose accuracy when the target-domain data differs from pretraining, and this paper addresses adapting them without labeled target data. The method combines progressive, class-balanced pseudo-label subset selection based on prediction margins with cascaded consistency learning from temporally smoothed teacher predictions, filtered by confidence. Only normalization parameters of the visual encoder and text-initialized class prototypes are updated. Experiments with CLIP and SigLIP backbones show consistent gains over zero-shot baselines and performance competitive with recent unsupervised adaptation methods.
In briefStructured-token attention for telling similar liquids apart with millimeter-wave radar
SummaryThe paper studies how to distinguish similar liquids from compact millimeter-wave radar measurements. The authors propose StructAttn, which splits a radar feature tensor into antenna, feature, branch, and global views, embeds them as eleven tokens, and feeds them to a three-layer transformer. On the public FG-LiquID dataset (30 liquids, 3,000 samples), it reaches 96.70% mean best-epoch accuracy over five seeds, 1.50 points above a scalar-token attention model. Five-fold results give StructAttn 95.30% out-of-fold accuracy.
In briefStreet design simulation balancing mobility and safety for planners
SummaryThis paper presents a microscopic simulation framework to quantitatively evaluate Complete Street designs with heterogeneous road users, including autonomous, connected, and human-driven vehicles, PUDO services, and micromobility. Safety is measured with a collision-physics-based conflict severity measure, and street layouts are optimized with a genetic algorithm. Applied to Belmont Avenue in Chicago, dedicated bicycle lanes reduced conflict severity but lowered vehicle throughput, while higher AV penetration reduced congestion.
Sharika J. Hegde, Amirmohammad Khakpour, Pedram Beigi et al.
In briefKeeping Object Identity Across Visibility Gaps in 4D Scene Graphs (CVG-SGM)
SummaryThe paper addresses re-identifying objects that vanish inside an opaque carrier and reappear as new detections in 4D scene graphs. It proposes CVG-SGM, a memory model that stores the target-carrier binding and uses it to recover the historical identity without appearance or motion cues. In a controlled setting it reaches 1.000 re-link accuracy versus 0.570 for no-memory and parameter-matched controls, and re-links 36/36 held-out detections on real RGB-D/SAM2 features versus 28/36 for the best fixed-binding control. The authors note that shared-carrier identity ambiguity remains a limitation and that end-to-end raw-video re-identification is not provided.
In briefComparing how well robots suit tasks in mixed multi-robot task assignment
SummaryThis paper compares ten deterministic methods and two LLM-based variants for estimating how suitable a robot is for a task, using rule-based, similarity, requirement, fuzzy, and semantic approaches. The methods score, rank, and select robots on identical robot and task profiles. The authors find that the methods often disagree on the same scenarios, and that suitability should be treated as an explicit modeling choice.
In brief1D Soliton Profiles: Enhanced Neural Solvers and Constrained Deep Ritz Methods
SummaryThis paper studies neural-network solvers for the one-dimensional stationary profile equation Q'' = bQ − γQ^p, which governs ground states in nonlinear Schrödinger, generalized KdV, and nonlinear Klein-Gordon equations. The authors improve the standard PINN baseline with Gauss-Newton/Levenberg-Marquardt and L-BFGS optimization and Fourier-feature PINNs, and they introduce two constrained Deep Ritz methods based on the Nehari manifold and the Weinstein quotient. The Deep Ritz variants perform best overall, with the Weinstein version giving the best error-runtime balance. In 1D, however, these neural solvers remain less accurate and efficient than classical Petviashvili iteration, and the authors present their work as a proof of concept for higher-dimensional and more complex problems.
Chandler Haight, Alex D Rodriguez, Svetlana Roudenko
In briefSelf-evolving agents learn from their own experience via retrospective self-distillation.
SummaryThe paper proposes Retrospective Self-Distillation (RSD), where one model acts as both student and teacher, using cross-task lessons, reflections, and hindsight to re-score its actions and distill them back into the policy. It also adds frontier-guided evolution, where a stronger model proposes constrained edits to how privileged context is built. Experiments on ALFWorld, ScienceWorld, and SearchQA with Qwen3-8B and Qwen3-14B report up to 59.7% relative improvement over the base model.
In briefMiCLA: Aligning Mixed Audio With Its Sound Parts for Multi-Sound Audio Learning
SummaryThis paper proposes MiCLA, a two-stage self-supervised pre-training method that learns audio representations for clips containing several overlapping sounds. It mixes clips synthetically in log-mel form and aligns each mixture with its constituent clips, using residualization to suppress one component and align the remainder with the other. The authors report 43.2% mAP on AS-20K and 50.3% mAP on AS-2M under fine-tuning, and the highest mAP among evaluated models on both AudioSet settings under frozen Convex Gated Probing.
