In briefSpatially-aware self-organizing graph autoencoder using localized Hebbian learning for stable visual features
SummaryThis paper proposes S-LooM, a self-organizing graph autoencoder that connects feature maps using spatial cross-correlation and learnable spatial kernels. It replaces scalar message passing and nearest-neighbor upsampling in decoders, and uses a soft-blending curriculum to align the latent topology smoothly during training. According to the experiments, S-LooM yields a more stable topology than current dynamic architectures, with greater resistance to feature entanglement and noise, and forms communities tied to illumination, shape, and texture.
Robin Ko\c{c}, Hazal Mo\u{g}ultay \"Ozcan, Fato\c{s} T. Yarman Vural
In briefBirth-death structural learning for 3D Gaussian Splatting using variational scores instead of heuristics
SummaryThe paper addresses adaptive density control in 3D Gaussian Splatting, which currently relies on proxy signals such as position gradients and opacity. It proposes a first-variation birth-death control method that derives a score from the multi-view reconstruction loss to select where new Gaussians should be added and which can be pruned. The authors report that integrating this score into 3DGS-MCMC achieves state-of-the-art reconstruction quality on standard benchmarks, with the largest gains under tight Gaussian budgets.
In briefXing USP: A shared semantic contract for information extraction and intent-slot parsing
SummaryThe paper proposes USP, a common interface that represents categories, fields, values, and instance links across named-entity recognition, relation extraction, event extraction, and intent-slot parsing. It defines five diagnostic matching views and normalizes 61 datasets into a unified corpus. The fine-tuned Xing4.0-29B-Intent model ranks first in every task family and on all five views, with its margin growing under stricter instance-level matching. The authors argue that category agreement alone understates structural errors.
In briefLong-short contrastive pruning for more compressible reasoning chains in pruned models
SummaryThis paper proposes Contrastive Set-Difference (CSD), which contrasts neuron activations on long and short correct solutions to the same problem to prune redundant-reasoning neurons first, and extends preference pairs to problems a pruned model always or never solves. On four math benchmarks, the combined approach uses fewer thinking tokens than pruning with reasoning-aware calibration (RAC) followed by the same compression, with accuracy within 0.5 points.
In briefCoLink: Compressing and Linking Dialogue History for Better Multi-Turn Chat Models
SummaryThis paper proposes CoLink, a framework that helps large language models keep track of the overall context in long, domain-specific multi-turn dialogues. It clusters dialogue history into topic graphs linked to external knowledge, and retrieves knowledge using both semantic relevance and graph structure. An information-theory-based objective compresses the history and knowledge into compact memory and knowledge slots. The authors report strong results on public datasets and their own MMD dataset.
In briefCrossPair: A method for transferring visual edits from a single before-after example
SummaryCrossPair adapts a pretrained multimodal image editor so it can apply diverse transformations shown in one before-after exemplar pair. A vision-language model extracts the operation's meaning while a VAE preserves fine appearance and structure. A training-only DINOv3 objective aligns feature-change directions, and the authors built the Transform-360K dataset to train it. The paper reports better target agreement and edit-direction similarity than released systems on the evaluated transformations.
In briefSHIFT-Jev: Certified Abstention and Gain-Gated Continual Adaptation for LLM Decision Models
SummaryThe paper studies how LLM-based decision models fail under distribution shift and proposes SHIFT-Jev, a procedure that scores decision risk, certifies abstention thresholds, monitors shifts, and accepts model updates only when they yield significant held-out gains without harming earlier regimes. Across ten shifted test environments, fine-tuning with KL anchors and temperature scaling reduced calibration error under shift from 0.289 to 0.087. In continual adaptation, SHIFT-Jev outperforms baselines on Qwen3-0.6B but trails them slightly on Qwen3-4B.
Muhammad Atif Saeed, Azka Atiq, Adarsh Kumar et al.
In briefTopX: Generating text-driven motion for unseen skeletons from the skeleton file alone
SummaryThe paper presents TopX, a 73M-parameter flow-matching transformer that generates joint rotations and root trajectories for any skeleton from its tree structure, rest pose, and joint names, without motion clips, templates, retargeting, or fine-tuning for that character. The authors also build SkelShift, a benchmark of 143,334 captioned clips with held-out skeletons organised by topology, joint count, species and rig source. Compared with UniMate trained on the same clips, TopX reports lower Fréchet distances on both kinetic and descriptor features, and fewer frames sinking below the ground.
In briefDeveloper Delegation of Tasks to LLMs and the Role of Trust
SummaryThis mixed methods study examines how software developers delegate development tasks to large language models, based on 22 interviews and a survey of 160 practitioners. Developers reported four delegation modes: full delegation, partial delegation, self-delegation, and task avoidance. Statistical analysis found that trust is a significant factor associated with task delegation, alongside task difficulty, task risk, and developer motivation.
In briefInsertion-based diffusion text generation with self-correction: δMDM
SummaryThe paper addresses insertion-based masked diffusion models, whose irreversible parallel decoding can cause errors that later steps cannot fix. It proposes δMDM, which adds substitution and deletion operations so the model is trained to correct its own mistakes and interleaves correction steps during inference. Reported results include halving generative perplexity versus FlexMDM on OpenWebText at the highest inference budget, and matching FlexMDM accuracy on TinyGSM with half as many NFEs.
Durga Prasad Maram, Dhruvesh Patel, Tahira Naseem et al.
In briefGeneralization in learning: Eyring formula and predator-prey model, a study of how machine learning generalizes via a biological-style population model and thermodynamics.
SummaryThis paper studies generalization in machine learning using a predator-prey model and reviews biological and chemical interpretations of stochastic gradient Langevin dynamics and generative adversarial networks. In an example of recovering an unknown function with a fixed-degree polynomial from noisy data, the predator-prey model gives a better approximation than gradient descent. The authors also use thermodynamic arguments, particularly the Eyring formula, to explain the delayed generalization known as grokking.
In briefPrototype-based object re-identification with boundary repulsion MixUp
SummaryThis paper presents PBR-ReID, a prototype-based framework that learns a prototype for each identity and uses their similarity matrix to find the most confusable identity pairs. It introduces Boundary Repulsion MixUp (BRM), which builds a boundary proxy between confusable pairs and pushes their features away from it. Training uses a prototype warmup stage followed by EMA-stabilized joint fine-tuning. The authors report strong results on person and vehicle re-identification benchmarks and ablations supporting each component.
In briefReal-time conveyor belt defect detection on a low-power edge device
SummaryThis paper presents a low-power edge vision system that runs on a Raspberry Pi 5 with a Google Coral Edge TPU to detect conveyor belt defects in real time. It uses an INT8-quantized YOLOv11 nano model trained on industrial and open-source images. The system reaches an mAP@0.5 of 0.764 across four defect classes. It was deployed at a cement plant, running at 30-40 FPS with 70-90 ms latency.
In briefEditDistill: Guiding Video Editing with Image Editing
SummaryThe paper proposes EditDistill, which edits the first frame of a video with an off-the-shelf image editor and distills the change into a compact edit descriptor. A lightweight fusion module injects this descriptor into a video diffusion model to steer denoising toward the desired edit. Training uses edit-descriptor alignment and DINO structural consistency losses to preserve motion and layout. The authors report strong benchmark performance and high-quality results across diverse scenarios.
In briefSame Tools, Different Interfaces: How MCP and CLI Change AI Agent Behavior
SummaryThis paper studies whether language agents behave differently when the same capabilities are exposed through Model Context Protocol (MCP) or command-line interfaces (CLI). The authors introduce MICA, a method that presents both interfaces from a shared action contract and compares them across six models and three interactive environments. They report substantial differences in task performance and interaction cost, and propose DART, which repairs interface-specific realization and improves benchmark scores in all six evaluated settings.
In briefAutomatically dating Babylonian astronomical diaries using language models and astronomy tools
SummaryThis paper presents a system that automates the astronomical part of dating broken Babylonian Astronomical Diaries tablets. A rule-based parser and a language model turn recorded observations into checked clauses, and astronomical tools test them against all years from 626 BCE to AD 75, ranking candidate years. On 120 dated tablets, the top-ranked tied years include the published year in 96% of cases. The output lists the supporting passages and computed positions for each candidate.
Lucas Carrit Delgado Pinheiro, Yuan-Sen Ting, Jon Clindaniel et al.
In briefChronoReplay: One Answer Format Isn't Enough for Streaming Spatial Reasoning
SummaryThis paper proposes ChronoReplay, a training method for spatial vision-language models that replays heterogeneous spatial supervision as chronologically ordered episodes, keeps each source's native targets, and balances supervision across task families. A single checkpoint improves on OVO-S-Bench, OST-Bench, and UCS-Bench together, with a 13.41-point gain on OST-Bench under the same observation access. The authors also report that post-training can concentrate gains on one answer format, and that adaptive frame selection at inference adds further gains at lower visual cost.
In briefA logic-based framework for comparing multi-agent path planning variants
SummaryMulti-Agent Path Planning requires collision-free paths for several agents, with applications such as automated vehicle navigation. The paper takes a specification-centered view, showing that many MAPP variants can be modeled in HyperLTL and its omega-regular extension HyperQPTL. It does not claim that current hyperproperty model checkers replace specialized MAPP solvers. Instead, it offers a unified language for comparing variants, expressing cross-agent constraints and optimality, and linking MAPP to verification, explanation, and causality techniques for hyperproperties.
In briefFair Weighted Scheduling for Shared Earth Observation Satellites
SummaryAs satellite constellations grow costlier, multiple stakeholders fund them and expect access proportional to their investment. Most prior fairness work assumes equal user weights, which this paper relaxes by weighting users by their contributions. The authors propose weighted utilitarianism, egalitarianism, leximin, and WOWA solution concepts, formulated as mixed-integer linear programs. These are compared against weighted greedy and weighted round robin heuristics on literature scenarios under strict runtime limits.
Maxime Lucet, Nawal Benabbou, Aur\'elie Beynier et al.
In briefFrom Transformer to GPT-5: Review of Large Language Model Progress, Limits, and Future Directions
SummaryThis paper reviews the development of large language models, from early models such as BERT to GPT, Gemini, Claude, and Grok, arranged along a timeline. It covers training datasets, training compute, hardware, architecture, and parameters, along with limitations, computational challenges, and open problems. The authors report that models balancing architectural complexity and data quality show stronger reasoning stability and better alignment without proportionally higher computational load.
In briefThree Answers to One Question: Measuring How LLM Perspectives Vary Across Languages on Contested Topics
SummaryThe paper introduces PerspectivesQA, a manually verified multilingual dataset of globally contested topics across 18 languages, and ROOTS, a framework for profiling linguistic, perspective, and framing diversity. Eight open-source and proprietary models were evaluated. The authors find that models hedge in English but take positions in lower-resource languages, and that larger models keep a more consistent cross-lingual stance than smaller ones. All models are also highly sensitive to minor surface-level input changes.
In briefDoubleFree: Reusing Exploration History for Rollout-Free Multi-Task LLM Training
SummaryThis paper addresses the high system cost of multi-teacher on-policy distillation (MOPD) in multi-task LLM post-training, which keeps several teachers online and generates new student responses. It proposes DoubleFree, which reuses historical trajectories from domain reinforcement learning for rollout-free consolidation and precomputes teacher supervision so students train without online teachers. Experiments on math, code, and instruction following with Qwen3-4B, Qwen3-8B, and SmolLM3-3B report consolidation speedups of 2.99 to 5.75 times and comparable multi-domain performance.
In briefRobust Single-Parameter Point Cloud Registration Using Truncated Lp Functions
SummaryThis paper proposes TLP, a point cloud registration method that needs only one tunable parameter. It uses truncated Lp functions with different p values to score correspondences, select hypotheses, and refine the final model, handling the high outlier rates common in real 3D feature matching. GPU parallelization allows registration from thousands of correspondences in under a second on average. On public benchmarks, TLP is more accurate than existing single-parameter methods and outperforms two widely used multi-parameter methods on three of four datasets, with competitive runtime.
In briefRacerGS: A residual-autoregressive context method for compressing 3D Gaussian Splatting scenes
Summary3D Gaussian Splatting (3DGS) renders novel views quickly at high quality, but its large parameter size makes storage and transmission costly. This paper proposes RacerGS, a compression framework that uses a hash projection model, a FiLM-modulated autoregressive GRU context model, and entropy coding of residuals relative to context-predicted means. The authors report more than 150x compression over vanilla 3DGS and 28x over Scaffold-GS with comparable rendering quality. Experiments on five datasets are said to show better storage reduction and rate-distortion performance than prior 3DGS compression methods.
In briefMultiEvoSkill: Multi-agent evolution of structured, reusable skill documents for AI agents
SummaryThe paper introduces MultiEvoSkill, a framework that optimizes skill documents used by frozen large language model agents. It represents skills as typed graphs to capture parallel branches and dependencies, and uses heterogeneous agents with a structure-aware debate and evolution mechanism to revise them. Multi-dimensional gating is used to ensure each evolution step's information gain and topological correctness. The authors report experiments and ablations against state-of-the-art methods supporting its effectiveness and generalization.
In briefRecovering business rules as cascade lookups to predict enterprise fields
SummaryEnterprise sales-order systems depend on manually maintained rules for fields like payment terms, and stale rules cause errors downstream. This paper proposes an LLM agent that offline proposes business keys, validates them with SQL profiling, and compiles them into deterministic cascade lookups. At inference, prediction is a transparent key lookup with no model serving. Cascades work best when fields follow stable business rules, while learned graph models are better when predictions depend on temporal or neighborhood patterns.
In briefVariational memory framework helping robot vision-language-action models handle long tasks
SummaryVision-language-action (VLA) models perform well in robotic manipulation but struggle with long-horizon tasks due to weak long-term memory. The paper proposes CogMemory, a variational cognitive memory framework that encodes multimodal trajectories into compact latent memories using a selective state-space encoder, multi-slot temporal abstraction, and variational compression. Relevant memories are retrieved via dynamic cross-attention and gated fusion during action generation. With only 3.45M additional parameters, the authors report consistent gains across diverse VLA backbones on long-horizon manipulation benchmarks.
In briefDoes Notation Matter? MusicXML Tokenization for Symbolic Music Models
SummaryThis paper introduces the first general-purpose tokenization framework for MusicXML, a standard score-interchange format. The authors train 26 autoregressive and masked language models and compare them against MIDI and ABC-based representations on music understanding and generation tasks. They report that MusicXML-tokenized models outperform MIDI and text-based counterparts on both kinds of task.
Andrea Poltronieri, Matteo Spanio, Simone Chieppa et al.
In briefAlignment via Generalization: A Learning-Theory Framework for AI Alignment
SummaryThis paper proposes Aligned-PAC learning, which treats AI alignment as a generalization question: does success on training requirements carry over to requirements the system has not seen? Intended behavior is expressed as a collection of specific requirements rather than a single objective. The authors show how independent evaluation can give rigorous evidence of alignment and how much evaluation is needed, and they present an optimization method that combines learned models under explicit assumptions. Choosing requirements that faithfully capture intentions, and understanding when learning generalizes to unseen requirements, remain open challenges.
In briefΔτ-Bench: A Benchmark for AI Agents Facing Changing Tool Ecosystems
SummaryThis paper introduces Δτ-Bench, an executable benchmark that tests AI agents as the set of available tools changes over time. It contains 1,588 validated multi-turn tasks across 42 domains and six successive ecosystem states. The strongest evaluated model or harness reaches only 42.3% task success, and gains from existing skill-learning methods do not consistently carry over as the ecosystem evolves. The authors propose Steward, an ecosystem manager that maintains tools and procedural knowledge incrementally, and report that it sustains higher task success under these changes.)
In briefParetoGS: Efficient 3D scene rendering at any Gaussian count
SummaryThis paper proposes ParetoGS, a method for 3D Gaussian Splatting that produces high-quality reconstructions across a continuous range of Gaussian-count budgets. It uses a small set of adapted budget anchors and a compact residual atlas to encode parameter changes, then interpolates between anchors to decode intermediate budgets without separate optimization. According to the authors, across 13 scenes it achieves better average PSNR, SSIM, and LPIPS with fewer Gaussians than compared prefix methods, and reduces mean additional state from 390.0 to 85.7 MB with a 0.01 dB average PSNR change.
In briefBoundedGS: a memory-efficient way to train 3D Gaussian Splatting at high resolution
SummaryThis paper presents BoundedGS, a method that reduces GPU memory use when training high-resolution 3D Gaussian Splatting models. It separates rasterization and projection backward passes and limits the active local rasterization working set based on measured load, with candidate-indexed projection and compact gradient storage. In tests on native 5K outdoor scenes, the mean rendering working set dropped from 9.97 to 1.76 GiB, with similar reconstruction quality and training time reduced from 36.7 to 33.9 minutes.
In briefConcept-bottleneck models that let users interpret and steer molecular graph generation
SummaryDeep graph generators used in drug discovery are largely hard to interpret, which limits control over the desirability of generated molecules. The paper adds concept bottlenecks to these generators so users can interpret, debug, and steer the generation process. The approach consistently lowers FCD, indicating better distributional fidelity, with only modest, non-systematic effects on validity and uniqueness. The authors also examine Concept Controller, an optimization-based steering method, and evaluate the approach on a one-shot molecular graph generation framework.
In briefA conditioned GAN that generates commercial aircraft shapes meeting target aerodynamic performance
SummaryThis paper proposes CEGAN, a Condition-Enhanced Generative Adversarial Network that embeds geometric and aerodynamic information into aircraft configuration generation. It combines a geometry-guided feasibility penalty, a performance penalty from surrogate models, and an output filter that discards infeasible designs. Tested on Cessna Citation X-like configurations described by 28 shape parameters and evaluated with OpenVSP, it reduces average generation error by at least 36% for interpolation and at least 46% for extreme extrapolation compared with a baseline.
In briefCropReward: A Human-Preference Benchmark and Reward Model for Aesthetic Image Cropping
SummaryThis paper finds that current evaluation methods for aesthetic image cropping, such as IoU and off-the-shelf aesthetic assessment models, align only weakly with human preferences. The authors build CropBench, with 106,995 human preference pairs across 1,019 images and 14 cropping methods, and propose CropReward, an MLLM-based reward model that scores crops against the original image. They also release CropReward-300k and CropAlign, a model trained using only CropReward supervision, which reportedly outperforms state-of-the-art cropping methods.
In briefFrom Presence Labels to Bounding Boxes: Can Multimodal AI Models Scale Surgical Tool Localization?
SummaryThis paper proposes a framework that uses multimodal large language models to convert surgical frames and existing instrument-presence labels into bounding boxes, and analyzes its reliability, cost, and design sensitivity. Tested on laparoscopic cholecystectomy videos from Cholec80 with six models, a multi-stage pipeline reaches 0.784 mIoU, above the best single-call baseline (0.735) and close to human-human agreement (0.814). The authors release a spatial extension of Cholec80 with 285,977 bounding boxes across all 184,498 annotated frames from 80 videos.
Mohamed Hamdy, Muraam Abdel-Ghani, Fatmaelzahraa Ahmed et al.
In briefSynthetic linguistic data: model-generated text in linguistic research
SummaryThis glossary entry examines synthetic linguistic data, which is increasingly used in linguistics. Extending Bresnan's garden-and-bush distinction, it uses an artificial plant nursery metaphor to separate corpus-like collections of model-generated language (synthetic bush data) from outputs elicited under controlled conditions to test specific hypotheses (synthetic garden data). The authors argue synthetic data is a distinct object of study, and comparing it with natural language can reveal shared patterns and divergences, but interpretation must account for the conditions of generation.
In briefh1: Testing the Bitter Lesson in Automated Theorem Proving
SummaryThis paper presents h1, a simple agentic framework that lets a language model work with the Lean 4 proof assistant through an efficient interface, using verifier feedback and self-reflection without imposed decomposition or search structure. The authors report state-of-the-art results on several benchmarks and solutions to eight open problems in FormalConjectures. At matched models, they report that h1 matches or outperforms specialized workflows such as Goedel-Architect and AlphaProof Nexus. They conclude that an efficient Lean interface largely determines cost and solve rate, so simplicity alone is not enough.
In briefStaX-Net: A Stackelberg-Inspired Asymmetric Network for Medical Image Translation Across Imaging Modalities
SummaryThe paper proposes StaX-Net to translate medical images between imaging modalities while avoiding anatomically inconsistent synthesis that could affect diagnosis. It combines a partially coupled encoder-discriminator design with a FAST layer that distills target-domain priors during training and applies only a source-side transform at inference. Training uses a Stackelberg-inspired asymmetric optimization with shared-parameter freezing and asymmetric learning rates for stability. The authors state that the code is publicly available on GitHub.
In briefHigher reasoning effort makes GPT-5.5's references more verifiable and changes its reasoning summaries
SummaryThe study examines GPT-5.5 generating scholarly references for 1,000 papers under low and high reasoning-effort settings, matching the references to external bibliographic records. High effort raised the mean paper-level resolver-verified reference rate from 78.3% to 84.8% and produced longer reasoning summaries. All four predefined summary features scored higher under high effort.
In briefMIS-VQA: An Annotation-Driven Surgical Vision-Language Dataset for Complex Surgical Understanding
SummaryThis paper introduces MIS-VQA, a surgical vision-language dataset built from EndoVis18 and CholecT50 annotations converted into structured text scenes that guide large language models to generate question-answer pairs. The dataset supports four response types, enabling evaluation from basic visual grounding to multi-step reasoning. The authors also present LLaVA-SurgXpert, a family of surgical vision-language models, which show improvements over existing surgical VQA baselines and reveal gaps between perceptual grounding and higher-level reasoning.
Mohamed Hamdy, Iheb Zouari, Fatmaelzahraa Ahmed et al.
In briefCORMA: A Mamba-based framework for generalizable multi-constraint offline safe reinforcement learning
SummaryThis paper studies offline safe reinforcement learning, which learns policies from static datasets while satisfying safety constraints at deployment. The authors propose CORMA, which splits policy inference into constraint-aware return-to-go estimation and adaptive action generation, using constraint-specific experts fused by a router-guided product-of-experts mechanism. Experiments on OSRL benchmarks report a 37% reduction in inference latency compared with Decision Transformer-based methods and stronger zero-shot generalization to unseen safety-budget combinations.
In briefMatOS: An AI Agent Operating System for Organizing Materials Research Tasks
SummaryThis paper presents MatOS, an operating system for AI agents doing materials research. It organizes data, knowledge, tools, skills, roles, and workflows as reusable assets, and runs long tasks as resumable, traceable execution graphs that revise assets based on records and user feedback. The authors report an 11.15% improvement over the previous best agent on ResearchClawBench and describe eight research cases, including narrowing over 30 million candidate structures to 97 solid-electrolyte candidates.
In briefCompact diabetic retinopathy grading via knowledge distillation from different teacher architectures
SummaryThis paper studies offline response-based knowledge distillation on APTOS 2019, using ResNet-50, MobileNetV3-Large, and Swin Transformer-Tiny as teachers to train a MobileNetV3-Small student for five-class diabetic retinopathy grading. The best single-seed quadratic weighted kappa was 0.9093, and the three-seed mean was 0.9024, above the 0.8804 no-distillation baseline. The student has 1.52 M parameters, an 18.07x reduction relative to Swin-T, and the authors state that independent external validation is still needed.
In briefSRA: Selective Residual Adaptation to improve LLM reasoning in small distilled models
SummaryKnowledge distillation transfers reasoning from large language models to small student models, but distilled students still struggle with multi-step reasoning. The authors propose SRA, a plug-and-play framework that uses a Discrepancy Gate to find positions where teacher and student distributions diverge and a Candidate Residual Adapter to calibrate top-K logits. On mathematical benchmarks with Qwen3 models from 0.6B to 4B, SRA raises average Pass@16 by up to 4.37 percentage points while preserving general capabilities.
In briefComparing 2D, 2.5D, and 3D Networks for Lung Nodule Malignancy-Risk Classification
SummaryThis paper compares 2D, 2.5D, and 3D residual networks for classifying the malignancy risk of pulmonary nodules using the LIDC-IDRI dataset of 2,654 nodules. It also tests momentum-contrast pretraining, mild augmentation, and excluding indeterminate nodules. The augmented 2.5D model achieved the highest test Macro-F1 of 0.6537. The labels reflect radiologist-assessed malignancy risk rather than pathology-confirmed cancer.
In briefTiming Matters Twice: When to Query and Skip in Flow-Matching Distillation
SummaryThis paper introduces PhaseOPD, a framework for on-policy distillation of flow-matching image models that decides when to query specialized teachers and when to skip the rest of the rollout. It identifies generative phases online from revisions of predicted clean latents, using student velocity predictions without extra model evaluations. On SD3.5-M, PhaseOPD executes only 5.56 of 16 rollout positions per local batch, cutting training time by 41.3% relative to DanceOPD while improving Aesthetic, GenEval, and OCR scores.
In briefGeoPrompt: geometry-guided prompt learning for segmenting underwater objects with open vocabularies
SummaryUnderwater images are degraded by color attenuation, scattering, and low contrast, which makes matching visual region features to text category embeddings unreliable, especially for unseen classes. The paper proposes GeoPrompt, which uses depth-derived geometric structure, considered more stable than appearance, to inject structural priors into visual features and generate geometry-aware textual classifiers. These classifiers are combined with fixed and image-conditioned textual classifiers. Experiments on the MARIS-OpenIS benchmark report state-of-the-art performance and consistent improvements in underwater open-vocabulary instance segmentation.
In briefAudioFM: A Self-Supervised Method for General-Purpose Audio Representation Learning
SummaryThis paper proposes AudioFM, a self-supervised framework for learning audio representations. It combines recording-level contrastive learning with patch-wise masked latent discrimination, and uses weaker supervision from temporal neighbors to encourage local temporal consistency. The authors report state-of-the-art results across diverse audio domains in linear probing, fine-tuning, and zero-shot transfer.
In briefEntropic Control for Homeostatic Optimization: Preventing Catastrophic Forgetting with Dynamic Feedback
SummaryThis paper addresses catastrophic forgetting in supervised fine-tuning of LLMs and proposes ECHO, a closed-loop control framework. It uses output entropy as an efficient online proxy to monitor forgetting risk, adjusting the gradient weight of replay data with a proportional controller. Experiments on models from 0.6B to 9B parameters show ECHO lies on and in some cases extends the plasticity-stability frontier, remaining robust across different data mixing ratios.
In briefWrite What Matters: Separating Token Selection from Slot Allocation in Recurrent Memory
SummaryThis paper studies how recurrent-memory sequence models decide what to write into limited memory slots. It proposes Factorized Write Routing (FWR), which first decides whether to keep or skip each token and then picks how many slots to use. Across ten seeds on two synthetic long-context tasks, always writing performed better on dense facts, while learned selective writing performed better on sparse facts.
In briefFreezing settled regions and caching smooth features to speed diffusion image editing
SummaryThis paper studies how to split a speed-up target between step caching, which reuses features across denoising steps, and region freezing, which skips regions an edit leaves untouched, for diffusion-transformer image editors. The proposed FreezePlan method keeps five full steps and assigns the remaining budget to TaylorSeer. On a pre-registered test split it runs 5.05 times faster end to end at 29.8 dB PSNR relative to the original output.
In briefReasoning or Just Prompting? Probing How Latent Chemical Reasoning Works
SummaryThis paper examines whether latent reasoning, which performs intermediate computation in continuous space instead of explicit tokens, forms a true hidden reasoning chain in chemistry tasks. The authors find that swapping or moderately perturbing latent states barely affects performance, while removing them entirely can cause a substantial drop. They argue the latent states mainly act like soft prompts that steer later computation rather than encoding step-by-step reasoning.
In briefMetaARIMA: Transfer Learning for ARIMA Model Settings
SummaryThis paper introduces MetaARIMA, a pretrained policy based on meta-learning that transfers knowledge across forecasting tasks at the ARIMA configuration level. It learns from historical experiments which configurations perform best and recommends candidates for new series. Trained on M4 and tested on M1, M3, and Tourism, it outperforms AutoARIMA and other classical methods and remains competitive with recent zero-shot foundation models.
In briefTask success alone misses unauthorized side effects in tool-using AI agents
SummaryThis paper argues that evaluating tool-using language model agents only by task completion overlooks unauthorized side effects, such as writing to the wrong object or contacting unintended recipients. The authors introduce SABench, a benchmark of 201 email, calendar, and ticket tasks with deterministic trace-based oracles for goal satisfaction and authorization compliance. Testing 11 frontier LLMs, they find 15.2% of runs cause collateral damage, with ambiguity tasks showing a 38.8% violation rate.
WhyIntroduces a collateral-damage metric with a deterministic oracle, tested across 2,211 runs of 11 LLMs, revealing safety risks that task success alone hides.
In briefReinforcement learning for adjusting crowd barriers to ease pedestrian flow
SummaryThis paper presents a simulation-driven reinforcement learning framework that learns crowd barrier placements without labelled training data. It builds a simulation of pedestrian distribution from a segmented crowd scene and trains a Soft Actor-Critic (SAC) agent to adjust barrier positions continuously. On unseen pedestrian counts, the learned policy reduced pedestrian stops compared with a manually designed barrier layout while keeping flow regulation effective.
In briefPredicting battery remaining life from a single cycle using compressed wavelet features and neural networks
SummaryThis paper proposes a method to predict a lithium-ion battery's remaining useful life from only its early single cycle data. Voltage, current, and capacity signals from each cycle are converted into time-frequency representations using a continuous wavelet transform, with scale compression applied to reduce redundancy. Evaluated on eight benchmarks derived from seven public battery datasets, the framework outperforms prior approaches in single-cycle prediction and stays robust as input cycles increase from 1 to 100.
In briefLast-iterate convergence of Muon with finite Newton-Schulz steps
SummaryThis paper studies whether practical Muon, which uses a fixed number of quintic Newton-Schulz iterations, converges in gradient at its last iterate on smooth, lower-bounded nonconvex objectives. Without input stabilization, the authors show Muon can cycle even on a smooth convex quadratic, while with positive stabilization they prove last-iterate gradient convergence for deterministic gradients. For stochastic gradients, they establish almost-sure and mean-square convergence by combining decreasing learning rates with increasing momentum, and construct a counterexample showing no pointwise polynomial rate is guaranteed.
WhyProves last-iterate convergence for finite-step Newton-Schulz Muon, with counterexamples and density-one rates, a solid theoretical contribution.
In briefAdaptive Neural Activation Using Parametric Tanh Edge Functions
SummaryThis paper proposes a compact neural framework that replaces the B-splines in Kolmogorov-Arnold Networks (KANs) with a bounded parametric tanh function having four learnable parameters, optimized edge by edge via backpropagation. The authors report accuracies of 95.63% on MNIST and 44.74% on CIFAR-10, gains over affine baselines. They state the method matches or exceeds the expressiveness of a 7-parameter exponential baseline while requiring up to 85% less training runtime.
In briefExecution Logs Shift VLM Judgments of Robot Progress
SummaryVision-language models are now used to grade robot videos, and text beside the frames can change their verdicts. Adding a single log line claiming progress to clips where the robot was actually losing progress flipped correct verdicts 10 to 35 percentage points more often than a neutral line across Gemini, Qwen and GPT. A one-line warning that the log may be wrong removes most of this effect.
In briefHumans in Space: Can MLLMs Build Human-Centric Spatial Understanding Through Active Exploration?
SummaryThe paper introduces SATE-Bench, an interactive benchmark with 316 real-world scenes and 2,525 multi-turn QA pairs for testing visual self-awareness and active viewpoint exploration. Models begin from a heavily occluded view and must query adjacent viewpoints until they gather enough evidence, with an efficiency reward that penalizes extra steps. Evaluations show that current multimodal LLMs often stop exploring too early while reporting high confidence, and they struggle with fine-grained 3D geometric tasks.
In briefMuSTEL: A Large-Scale Synthetic Multilingual Dataset of Text-Line Images for OCR Training
SummaryThis paper presents MuSTEL, a synthetic OCR dataset with over 14.5 million annotated text-line images in Persian, Arabic, and English. The images are generated from natural-language corpora with varied fonts, mixed writing directions, highlighting, underlining, and visual distortions such as blur and skew. The dataset is intended for training and evaluating multilingual line-level OCR systems.
In briefRethinking Retrieval for Long Contexts: Training-Free Semantic Trees for Adaptive Context Building
SummaryThis paper introduces TreeFuse, a training-free framework that builds a reusable semantic hierarchy from a full document in one generation, linking node descriptions to source text. For a query, it selects relevant nodes through structural-semantic scoring and lexical attribution, adding ancestor context to preserve relations and local text. Across four frontier models on three LongBench tasks, it improves average QA-F1 by 1.72 points over full-context prompting while shortening the answer context.
In briefMeasuring which fine-tuning weight changes in CLIP actually drive performance, via reversion and retention tests.
SummaryThe paper studies which weight changes in fine-tuned CLIP models actually support task performance. It pairs reverting selected parameters to pretrained values with keeping only selected fine-tuning updates, evaluated on several image datasets. Attention value-output (VO) updates alone recover 43-87% of the accuracy gain and outperform equally sized random subsets.
In briefDecoupled prune-and-reconstruct: training GNNs on a compact node core for memory-limited graphs
SummaryThe paper proposes DPR, which prunes a large graph to a compact node set for training a GNN encoder, then reconstructs pruned nodes at inference with a sparse structure-aware decoder. The authors report 1,450 MiB peak training memory on a 6 GB GPU and 96.9% recovery of the dense-reconstruction reference. They also report comparisons with the DSpar baseline and a 2.45M-node production graph.</summary>
In briefSelf-Evolution Reasoner: Continuous Self-Improvement of LLM Reasoning Through Policy-Intrinsic Exploration
SummaryThis paper proposes Self-Evolution Policy Optimization (SEPO), a framework for continuously improving large language models' reasoning. It uses labels to guide the policy toward correct reasoning, then treats the policy conditioned on that reasoning as a self-teacher, distilling newly found solutions back into the policy. The authors report that SEPO achieves the best performance under identical training time or steps across scientific, tool-use, and complex logical reasoning tasks, and maintains stable gains where baselines underperform or fail.
In briefStickFigure: Pose-only IoT cameras that keep edge vision private
SummaryThe paper presents StickFigure, an IoT camera system that keeps raw video local and exports only compact skeleton pose data for human activity recognition, reducing exposure of faces and household details. It introduces an activation-aware pose backbone and a lightweight hierarchical MLP refiner to run accurate pose estimation on resource-constrained cameras. On NTU RGB+D and Penn Action, it reports the highest pose accuracy and activity recognition utility versus MCU pose models, with 96.5% fewer parameters, 99.3% fewer MACs than RTMPose-S baselines, and 0.107 s p95 latency on an STM32H753 MCU.
In briefFine-tuned open LLMs for classifying Eclipse bug severity versus simpler baselines
SummaryThe paper compares small open language models, Llama 3 8B and Qwen 2.5 7B, with TF-IDF classifiers and a CodeBERT encoder on four-level bug severity classification using Eclipse reports. Zero-shot models score below all TF-IDF models. After fine-tuning, Llama 3 shows no detectable difference from CodeBERT, while Qwen 2.5 scores lower. The authors also report duplicate rows in the data and release split identifiers, predictions and analysis code.
In briefHumanoidNDM: Using a learned neural dynamics model to close the sim-to-real gap in humanoid motion tracking
SummaryHumanoid robots that track motions can perform poorly when deployed in the real world due to extra physical disturbances or inaccurate simulation models. The paper proposes HumanoidNDM, an online fine-tuning framework that uses a neural dynamics model trained on real-world data. It collects data with low-frequency reference perturbations, trains a proprioceptive history-based dynamics model using differentiable forward kinematics, and fine-tunes the policy via backpropagated model gradients. Physical experiments on the Unitree G1 robot are reported, and the authors claim it is the first online, motion-capture-free fine-tuning framework for humanoid tracking policies.
In briefUsing evolutionary algorithms and machine learning to improve type 2 diabetes diagnosis
SummaryThis paper proposes a model that combines a Random Forest classifier with the Bat Optimization Algorithm to perform feature selection and hyperparameter tuning together. The authors applied SMOTE-ENN to the Pima Indians Diabetes Database to correct class imbalance and remove noisy samples. The optimized model reached 89% accuracy, 88% precision, 90% recall and an F1-score of 89%, outperforming the baseline Random Forest and other existing approaches. The authors suggest it could be a practical, low-cost tool for early detection of type 2 diabetes in clinical practice.
Parisa Rezaei, Mohsen Saffar, Hamid Reza Naji et al.
In briefSynthetic Data Augmentation for Functional Data: Wasserstein Risk Bounds and Diffusion-Based Conditional Generation
SummaryCollecting functional data such as curves is costly, often yielding only tens to hundreds of samples. The paper studies augmentation with a class-conditional functional diffusion generator and gives guarantees in the Wasserstein-1 distance, including the best number of synthetic curves and when synthetic data beats real data. Applied to temperature, hyperspectral and gait curves, augmentation improved classification and functional regression, with the largest gain from directing synthetic curves to the minority class under class imbalance.
In briefRoboDuo: A Code-Based Attempt, Check, and Atomic Recovery Framework for Vision-Language-Action Models
SummaryVision-Language-Action models do not check whether an instruction was completed, so one failed step in a long task can leave it unfinished. RoboDuo is a closed-loop framework where a coding agent, via code, first has a full-task VLA attempt the task and check completion, and an atomic VLA finishes remaining steps on failure. The method wraps existing VLAs without updating their weights. The authors report consistent gains for X-VLA and π0.5 on LIBERO and LIBERO-PRO, largest on long-horizon tasks, and recovery of failed rollouts on a real robot.
In briefCan LLMs Trust Appropriately? Calibrated Reliance on Imperfect Partners
SummaryThis paper examines whether large language models turn evidence from partners, such as humans or other AI agents, into appropriate reliance policies. The authors built a benchmark of 324 scenarios across 23 collaboration categories, each with three sequential decisions and explicit costs and feedback rules. The strongest frontier model made near-optimal choices, while other models showed recurring problems such as buying verification with zero expected value.
In briefT2D: Pruning and adapting a target model into a faster draft model for speculative decoding
SummaryThis paper proposes T2D, a framework that builds draft models for speculative decoding directly from a large language model. It combines latency-aware structured pruning with knowledge distillation, jointly optimizing measured inference latency and cross-entropy loss on the target model's responses. Experiments across benchmarks and model families show that T2D prunes a large portion of the target model to obtain draft models that outperform competitive baselines, and these drafts can also be used plug-and-play with other methods.
In briefTemporal causal reasoning via causal topology: two architectures for LLMs
SummaryThe paper addresses the weak causal reasoning of LLMs on temporal tasks and proposes shifting from discrete narratives to causal topologies. It presents TCR-Agent, a training-free neuro-symbolic framework using external causal graphs and a symbolic intervention engine, and TCR-Former, an end-to-end model that embeds these mechanisms into Transformer attention. On the TempoCausal benchmark, TCR-Former reportedly scores 16.5% higher than prior SOTA methods with about 8.6B parameters.</summary>
In briefLegoOcc++: Consistent Open-Vocabulary 3D Occupancy Prediction
SummaryThis paper studies open-vocabulary 3D occupancy prediction, which estimates occupied space and its semantics in environments whose categories go beyond fixed label sets. It uses only binary occupancy labels for geometric supervision and builds on Language-Embedded Gaussians, adding an opacity-aware Poisson-based formulation, progressive temperature decay, instance-level semantic consistency, and covisibility-aware support-view selection. The authors report strong results on both indoor and outdoor benchmarks.
WhySubstantive refinements to Language-Embedded Gaussians with opacity-aware occupancy and instance consistency; broad indoor and outdoor experiments support useful progress for open-vocabulary occupancy
In briefLeta: a few-step video distillation method that learns the score difference and future credit for autoregressive video generation
SummaryThis paper presents Leta, a few-step distillation framework for autoregressive video generation that combines two modules, ScoreDelta and FutureCredit. ScoreDelta directly estimates the teacher-student correction from denoising residuals, while FutureCredit learns a cross-chunk update signal from vector-Jacobian products of next-chunk teacher residuals. In the authors' controlled experiments, Leta improves over DMD by 0.9 points on VBench and 0.66 on VideoAlign.
In briefAgentBaton: Efficient Model Collaboration in Agentic Tasks
SummaryThis paper studies how a capable, expensive model and a cheaper, efficient model can share a single long-horizon agentic task to cut costs. It examines design choices such as handoff direction and context transfer, finding that handing off from the capable model to the efficient one at suitable stages largely preserves or improves accuracy. The authors propose AgentBaton, a lightweight detector that learns when to hand off, which reduces API costs by 65.9% on average across 11 model pairs and two benchmarks, with accuracy gains of up to 8.8 percentage points.
In briefTracking Hidden Objects by Position and Velocity with Doppler-aided mmWave Radar
SummaryThis paper presents an indoor non-line-of-sight tracking method using mmWave radar that exploits the full Range-Azimuth-Doppler information. The authors release a dataset of 576 sequences totaling 720,856 frames with synchronized position and velocity ground truth. The method estimates position and velocity sequentially and refines them into a motion-consistent trajectory. It reports errors of 8.95 cm in position and 1.75 cm/s in velocity on held-out trajectories.
In briefOvershoot in filters for event restarts after long silences
SummaryThis paper studies how explicit Gaussian point-process filters amplify the first update exponentially when events resume after a long silence. It gives the exact distribution of that update and a computable index that predicts when overshoot occurs, along with second-order silence expansions of the state variance. The predictions are checked across all 73 crossing cells and on exchange trade data, and variance caps are evaluated as a remedy.
In briefLALA Benchmark: Testing LLMs on Literary and Linguistic Analysis
SummaryThis paper introduces LALA, a benchmark of 236 open-ended questions written by linguistics and literary experts, with expert ideal answers. Seventeen language models were tested and scored by human experts and two LLM evaluators. Base models showed little of the assessed linguistic competence, while post-trained versions performed much better, and parameter count had no significant link to partial-credit scores. Performance was uneven across domains, with logical relations and some discourse tasks strongest and ambiguity, style, reference, and misleading assumptions still difficult.
Michael Foster, Barbara McGillivray, Gard B. Jenset et al.
In briefCommunication-efficient speculative decoding for distributed LLM inference
SummaryThis paper presents DISCO, a framework for distributed speculative decoding in device-server collaborative large language model inference. It places a replica of the draft small language model on the server so the server can locally obtain the draft-side distribution needed for rejection sampling. Instead of transferring full vocabulary-scale distributions, only drafted token indices and scalar probabilities are exchanged. Experiments on multiple datasets report higher decoding speedup than existing distributed speculative decoding methods, especially under strict bandwidth constraints.
Joohyun Shin, Byeonghyeon Park, Keumseo Ryum et al.
In briefIdentically trained language models fail on different jailbreak prompts
SummaryThe paper asks whether another pretraining run on the same data, with the same safety training, would be broken by the same jailbreaks. The authors trained 150 models across 25 pretraining data recipes and compared their non-refusal rates on shared jailbreak prompts. They find that which jailbreaks succeed depends much more on the pretraining seed than on the data recipe.
WhyShows jailbreak vulnerability depends more on pretraining seed than data recipe, validated across 150 models and 3,211 prompts.
In briefMemory Tax: Measuring the Net Value of Agent Memory
SummaryThis paper evaluates dedicated memory systems for long-horizon agents across multiple benchmarks and model backbones. It finds that many underperform naive RAG, which retrieves directly from raw history, indicating negative net value. The authors separate the gains and losses relative to RAG into "memory benefit" and "memory tax", focusing on the overlooked tax and how it arises.
In briefINTRA: A hardware-aware sparse attention design using relay communication across layers
SummaryThe paper introduces INTRA, a static sparse attention method for Transformers that uses complementary relay groups so information can travel between distant tokens across layers. A custom GPU kernel loads and densely packs non-contiguous tokens to make the sparse computation efficient. Reported results include a 1.53x speedup for 2K FLUX.1-dev image generation on an A100, improved long-context prefill on LLaMA-3.1-8B, and up to 8.4x and 10.4x attention-only speedups over MoBA and NSA.
In briefExtended LASMP: A Language-Guided Sampling Planner for Drone Navigation
SummaryThis paper introduces a sampling distribution that uses navigation commands to guide exploration, making sampling-based motion planning for quadcopters more efficient in 3D spaces. It extends command-aware subset sampling to aerial vehicles and uses a prompt-based large language model to turn natural language instructions into yaw- and altitude-aware geometric constraints. Simulations in complex environments show collision-free, consistent trajectories with significantly less sampling than uniform RRT-based planners.
Saswati Bhattacharjee, Anirban Sinha, Mukulika Ghosh et al.
In briefUsing Dale's Principle as a Built-in Bias in Recurrent Neural Networks
SummaryThe paper introduces EISep RNN, a recurrent network that enforces Dale's principle, where each neuron is either purely excitatory or purely inhibitory, through a learnable excitatory-inhibitory separation. Across multi-task, continual, and reinforcement learning, it reports more stable training and competitive or better performance than vanilla RNNs and earlier Dale-constrained models. The authors also report sparser, more modular connectivity and stable, informative latent dynamics.
In briefBEHAVIOR-Spatial: a benchmark checking whether robots follow spatial instructions, not just finish tasks
SummaryThis paper introduces BEHAVIOR-Spatial, a benchmark built on BEHAVIOR-1K that tests spatial question answering, point-based target grounding, and robot action to see how VLM spatial understanding relates to VLA behavior. It includes 3,256 spatial questions, 1,240 grounding queries, and action evaluations across 14 manipulation and 10 navigation tasks. The baseline reaches 41.6% on held-out configurations but drops to 16.1% when instructions change, while oracle visual target cues raise these rates to 60.3% and 38.0%.
In briefOptimal regret limits for kernel-based reinforcement learning with Sobolev smoothness
SummaryThis paper studies how well episodic reinforcement learning can perform when rewards and transitions are smooth functions in a Sobolev kernel space, with continuous states and actions and adversarial initial states. The authors prove that the minimax expected regret over T episodes scales as T raised to an exponent determined by action and state dimensions. They introduce Sobolev Optimal Value Iteration (SOVI), which achieves this rate, and their matching upper and lower bounds fix the optimal dependence on the number of episodes. The work partially resolves an open problem posed by Vakili (2024).
In briefA sign law explaining entropy drift from sampler mismatch in RLVR
SummaryThis paper studies why entropy in reinforcement learning with verifiable rewards rises or falls when choices like masking or failing over-length responses are made. The authors prove that a constant advantage shift changes softmax policy entropy at first order, with the direction set by the sampler, such as top-k or top-p truncation. Experiments on 2B and 9B models show entropy tracks the net advantage under truncated sampling, and baseline corrections or renormalization improve held-out accuracy by up to 8.8 points.
In briefContinual agent learning that checks and revises its own skills and lessons over time.
SummaryThe paper introduces KEEL, a framework that uses verifier evidence to turn successful agent behavior into reusable skills and failures into corrective lessons. As new evidence arrives, individual skill claims can be supported, contradicted, replaced, or retired. On the SkillFlow benchmark, KEEL improves task success by 4.2 to 21.7 percentage points over a no-knowledge baseline, and ablations show skills and lessons are complementary.
Wenzhe Fan, Benjamin Steenhoek, Gaurav Mittal et al.
In briefWhat sketches reveal about how vision-language models draw and locate parts
SummaryThis paper proposes a training-free sketch probe to test whether large vision-language models (LVLMs) retain human drawing conventions, such as the order in which object parts are drawn. The authors evaluate part order on nine LVLMs and attention-based grounding on five, using two part-annotated sketch datasets. They find that model part orders track human conventions and that part-level grounding survives on sketches, with the signal concentrated in a narrow mid-network layer band.
Chaitat Utintu, Ahmed Bourouis, Arkaprabha Basu et al.
In briefPretrained network for in-context off-policy evaluation of decision policies
SummaryThe paper proposes InCOPE, a prior-data fitted network for off-policy evaluation, which estimates a decision policy's expected return from logged data when online testing is costly or unsafe. The model is pretrained on synthetic tasks built from Markov decision processes and behavior-target policy pairs with exact target-policy values. For a new tabular problem, it takes the behavior-policy dataset, target policy, and task settings as context and predicts the return without task-specific parameter updates. Experiments on four tabular domains, including a semi-synthetic clinical setting, show its effectiveness for policy-value estimation.
In briefFovea-JEPA: Fixing V-JEPA's Geometric Blind Spots for Camera Motion Recognition
SummaryThis paper addresses camera motion recognition, which requires distinguishing camera movement such as translation, rotation, or zoom from object motion. The authors find that V-JEPA underrepresents geometric cues in some regions, which they call geometric blind spots. Fovea-JEPA uses V-JEPA as the base model and VGGT as a geometry expert, with their prediction errors guiding a blind-spot predictor that adaptively fuses visual and geometric features. The method reports gains of 9.67, 8.67, and 24.38 percentage points over prior state-of-the-art baselines on CameraMotionDataset, CameraBench, and ShotBench.
In briefStructured Time-Frequency Representation for Channel Prediction
SummaryThis paper addresses channel state information (CSI) forecasting in high-mobility wireless systems, proposing FAR-CSI, which places historical uplink observations and future downlink targets in a joint temporal and operating-band positional field. A unified structured routing rule keeps the same FDD geometry across encoding, decoding, and target selection. Under matched conditions, FDD-aware routing reduces NMSE by 21.44% relative to generic routing, and the method keeps its predictive advantage across varied mobility, horizons, and noisy observations.
In briefTCSP3.0: Template-based crystal structure prediction with composition-preserving element substitution
SummaryThis paper presents TCSP3.0, a template-based method that predicts crystal structures for new compositions by substituting elements in known structures. The authors found that 20% of structures selected by the original TCSP2.0 on the T180 benchmark had incorrect compositions, and they fixed this with coefficient-preserving mappings and simultaneous substitution. The updated method adds structural deduplication and machine-learning-potential relaxation, reaching Top-1 and Top-20 match rates of 72.2% and 90.6% on T180, versus 28.3% and 45.6% for TCSP2.0.
In briefExplaining Robot Navigation Decisions with Stage-Aware Visual Attribution
SummaryThis paper addresses how to explain the navigation decisions of vision-language models that guide robots by language instructions. The authors note that existing explanation methods fail to distinguish between different instruction stages, and they propose a stage-aware attribution framework that aggregates relevant historical observations. An instruction-guided spatial filter keeps explanations focused on relevant regions. Experiments across multiple navigation policies and benchmarks show better attribution quality, and the method is also shown as a tool for failure analysis and tested on a Unitree G1 robot.
In briefWhen Does GEPA Merge Help? A Mechanism-Level Diagnostic
SummaryGEPA is a reflective prompt optimizer whose merge operator behaves inconsistently across models, improving gpt-4.1-mini on average while degrading Qwen3-8B. Across five datasets and two models, the authors ran a nine-configuration sweep over merge timing and operator choice, finding that merge can improve both models when configured appropriately. They conclude the instability is better explained by configuration sensitivity than by a flaw in the merge operator, and show prompt-level static features do not predict merge-accept decisions. A behavior-based probe that abstains on high-risk tasks avoids large merge failures as a proof of concept, but does not consistently outperform NoMerge.
In briefSymmetry of the Differential Neural Tangent Kernel under Orthogonal Transformations: A study of how a kernel describing neural network training behaves under rotations.
SummaryThis paper studies the symmetry of the infinite-width differential neural tangent kernel (DNTK), a tool for analyzing training dynamics of physics-informed neural networks, under an orthogonal group. The analysis covers two-layer and deep networks, using the explicit DNTK expression and a recursive formula respectively. The authors show that integral operators on the unit sphere induced by operators in the same group orbit share spectra, and their reproducing kernel Hilbert spaces differ by an orthogonal transformation. They also show that the DNTK exhibits zonal symmetry for symmetric operators, allowing eigenvalue decay rates to be derived for shallow RePU networks with zero bias initialization.]
