In briefGood Views Are Not Enough: Counterexamples for Gaussian-Regularized JEPAs
SummaryThis paper shows that good views alone do not guarantee scene invariance or uniqueness up to rotation at the global optima of Gaussian-regularized JEPAs. The authors give globally certified counterexamples showing that the uniqueness premise of the Gaussian extension in Good Views fails for specified soft objectives. They also show that better scene alignment can coincide with worse prediction performance.
In briefLHNav: Long-Horizon Humanoid Navigation with Multirate Coordination and Physics-Grounded Motion Data
SummaryThis paper studies humanoid robots navigating to multiple destinations in sequence under long-horizon language instructions. The authors propose LHNav, which splits navigation into three time-scale roles: a Brain for task reasoning, a Cerebellum for local motion generation, and a Brainstem for whole-body execution, coordinated asynchronously through progress and arrival feedback. The Cerebellum is trained with a physics-grounded data engine that produces natural, route-faithful motion along collision-aware routes. In simulation, the system outperforms OmniNav and DualVLN on single-goal success and motion smoothness and reaches a 75.56% success rate on long-horizon tasks, and it was also deployed on a Unitree G1 robot.
In briefOne-shot federated graph learning with better shared proxies and transfer-aware collaboration
SummaryThis paper proposes FedHFT, a framework for federated graph learning that allows only a single communication round. It identifies two problems: aggressive one-shot compression distorts clients' local graph topology and semantics, and the server cannot reliably judge cross-client collaboration potential. FedHFT addresses these with fidelity-driven proxy refinement and transferability-steered personalized collaboration, and the authors report experiments on homophilic and heterophilic graphs.
In briefRAPA: A Ray-Aware Physics Augmentation Method for Robust Camera-LiDAR Calibration
SummaryThis paper frames checkerboard localization in LiDAR data as a 2D detection problem using cylindrical reflectivity images. To address the scarcity of labeled training data, it proposes RAPA, which generates diverse synthetic samples from a single checkerboard observation to improve detector robustness. Across four indoor and outdoor locations, the synthetic augmentation outperforms the strongest image-space augmentation baseline by 16.1 to 43.1 percentage points in mAP@0.5.
In briefToward Sustainable Agentic AI: Carbon Tradeoffs in Self-Refinement and Tool-Augmented Coding
SummaryThis paper reports two empirical studies estimating the energy use and carbon footprint of agentic AI systems at increasing levels of complexity. Study 1 compares GPT-4.1-mini with 1 to 5 steps of iterative self-refinement against a GPT-4.1 single-pass baseline on two MATH benchmark splits. Study 2 evaluates GPT-4.1-mini, Claude Opus 4.6 and GPT-5.3-codex on SWE-bench Lite, where agents use code execution and file editing tools inside Docker-isolated environments.
Boris Ruf, Adrien de Mailly Nesle, Marcin Detyniecki
In briefRounding Errors Make Hybrid Reasoning Models Think Longer: bf16 Recurrent State
SummaryThis paper studies how storing the cached recurrent state of hybrid reasoning models in bf16 with round-to-nearest rounding makes them produce longer outputs, hit the 32K token cap more often, and lose accuracy. Experiments on six open hybrid models and 590 competition-math problems show that stochastic rounding largely removes the length increase. The authors recommend that evaluations of hybrid models report state dtype, cap rate and output length alongside accuracy.
WhyThe authors study bf16 round-to-nearest state caching in six hybrid reasoning models on 590 math problems, reporting longer outputs and accuracy drops, with stochastic rounding removing much of the ef
In briefCARE-PO: A Method for Controlling Tool-Call Efficiency in LLM Agents
SummaryThis paper presents CARE-PO, a training method that reduces how many external tool calls an LLM agent makes while keeping task success. It prices each call against a reference budget tied to task difficulty and fixed before training, and injects successful training trajectories to prevent the policy from collapsing into making no tool calls. An effort tag in the prompt shifts the trade-off between efficiency and capability from a single training run. Results are reported on the GAIA and four held-out multi-hop benchmarks across model sizes.
In briefSystem optimizations to lower costs for AI agent harnesses
SummaryThis paper studies how the software harnesses that connect AI agents to computing environments affect the cost of running them. The authors analyze how costs accumulate across agent execution traces and identify ways to reduce token use and LLM calls without changing the model or harming task accuracy. They present Agency, a platform that applies these harness-side optimizations to existing tools such as Claude Code and OpenAI Codex. Across several real-world benchmarks, the proof-of-concept optimizations cut costs by up to 35% without reducing accuracy.
In briefSCALE: A small-model-first framework that abstracts concepts from experience with LLM guidance
SummaryThis paper proposes SCALE, a framework that lets autonomous agents build reusable knowledge from experience while relying less on large language model reasoning. A lightweight model handles routine predictions, and prediction errors or accumulated cross-episode evidence trigger LLM reflection that generates reusable concept hypotheses, which are then built into hierarchical neural representations. Across ten random seeds, historical recall raised held-out accuracy from 87.11% to 98.28%, hierarchical concept knowledge raised difficult-case accuracy from 55.33% to 66.47%, and active experience records dropped from 6,000 to 1,528.
In briefJoi-Bench-Task: Benchmarking Task Coordination in Full-duplex Voice Agents
SummaryThis paper introduces a benchmark that tests whether full-duplex voice agents coordinate task execution with conversation, tracking what is executing, completed, communicated, and still to report. It proposes an interactive user simulator that injects configurable interventions during tool execution and spoken result delivery. Five commercial voice agents were evaluated, with the highest result completion rate under request changes at only 55%, and four agents showing average information omission rates of at least 23% after user interruptions.
In briefForecast Value Add: Reward-Based Action Selection for Demand Forecasts
SummaryThis paper proposes refining the output of a strong machine-learning ensemble for demand forecasting by treating refinement as a reward-driven decision problem. An agent chooses among values the ensemble already produced, such as the base forecast, bounds, quantiles, and a bias-corrected estimate, with choices evaluated against context-matched historical periods and constrained by a seasonal guard. Across five retail, consumer-goods, and direct-to-consumer deployments over nine months, the refined forecasts reduced mean WMAPE by 6.3% relative to the stacked ensemble and achieved the lowest error in 35 of 45 customer-months.
In briefOGDA: Ontology-Grounded Data Agents for Reliable Analysis Across Mixed-Schema Data Lakes
SummaryThis paper presents OGDA, an ontology-grounded data agent that helps LLM agents perform end-to-end analysis over data lakes containing many tables with different schemas. OGDA builds a compact ontology through agentic hypothesis testing, where a proposer agent induces schemas, concepts, and relations and a verifier checks them against the data. Evaluated on KramaBench and DiscoveryBench with three LLM backbones, OGDA outperforms the strongest baselines by up to 10.82 points.
In briefSkillDuet: Co-evolving models and reusable skills using response feedback
SummaryThis paper presents SkillDuet, a framework that coordinates parameter updates and skill evolution in LLM reinforcement learning using model response feedback. Skills extracted from on-policy trajectories are ranked by a response-based score called Strength, and the top-scoring skills are periodically consolidated into a persistent global skill set. With Qwen3-4B, SkillDuet reaches 98.57% success on ALFWorld and 92.97% on WebShop, exceeding the matched SEED-API baseline by 2.86 and 5.47 percentage points.
In briefRisks and opportunities of AI in public climate services
SummaryThis paper examines how generative and machine-learning AI systems are being used to develop and deliver climate services. The authors combine a consensus-based expert elicitation method with the PESTLE framework to analyse political, economic, social, technological, legal and environmental risks and opportunities across the co-production value chain. They draw policy conclusions on data, models and infrastructure, and outline challenges and benefits of a pan-European, decentralised public climate service built on sovereign AI infrastructure and open governance.
Francesca Larosa, Sandro Calmanti, Matteo De Felice et al.
In briefReal-time method predicting volumetric light transport without path tracing
SummaryThis paper presents a real-time method that reconstructs full pixel radiance in volumetric media under direct and ambient lighting. Instead of relying on costly path-traced radiance estimates, it infers the final light transport directly from inexpensive primary-ray and environment samples. An attention-augmented image-to-image network performs the reconstruction. The authors report real-time performance and high accuracy in challenging volumetric cloud scenes.
In briefSPARK: A Simple Method for Adapting Pretrained Vision Models to Robot Control
SummaryThis paper proposes SPARK, a post-training method that adapts general-purpose vision foundation models for robot control. It combines temporal abstraction, which yields compact features that capture temporal changes and action-relevant scene structure, with knowledge preservation, which aligns patch-level representations with the original model. Experiments on multiple robotics benchmarks show higher success rates and better generalization than the vanilla foundation models, and these gains carry over to real-world robot manipulation.
In briefAdjusting Within-Document Boundaries to Improve Contextual Chunk Retrieval
SummaryThis paper addresses the difficulty of separating a correct chunk from semantically similar sibling chunks from the same document in Late Chunking, where whole documents are encoded before pooling at chunk boundaries. It proposes a two-part framework: one component adapts the discriminative scale using within-document margins, and the other lets attention see chunk partitions during document encoding. Experiments on ConTEB and NanoBEIR show improved contextual chunk retrieval across the tested encoder settings while keeping competitive passage-level retrieval quality.
In briefFrom Fluent Text to Ontology-Compatible Definitions: A Quality-Control Agentic Workflow
SummaryThis paper addresses the problem that LLM-generated biomedical ontology definitions can read fluently yet reference ungrounded entities or violate ontology constraints. It proposes two metrics, Entity Error Rate (EER) and Logical Inconsistency Rate (LIR), to measure entity grounding and logical consistency. The authors also introduce an agentic quality-control workflow that generates definitions, extracts axioms, checks them, and repairs them iteratively. Experiments show the workflow lowers EER while keeping strong text-generation scores, and LIR reveals sparse logical errors that text similarity metrics miss.
In briefCRAFT-AI: Aligning Capabilities and Requirements for Flexible Manufacturing with Agentic AI
SummaryThis paper presents CRAFT-AI, an agentic AI approach that combines natural language requirement interpretation, Asset Administration Shell (AAS)-based capability descriptions, and reusable engineering knowledge from existing manufacturing configurations. A tool-using AI agent matches requirements to available capabilities and reference implementations to generate manufacturing scenes and corresponding AAS instances. The approach was tested with a Factory I/O prototype using twelve reference scenes and five manufacturing tasks of varying complexity. The authors also highlight the need for safeguards for risk perception and communication in foundation model-driven solutions.
In briefTSFFormer: A Frequency-Domain Spike Transformer for Low-Power Activity Recognition
SummaryThis paper proposes TSFFormer, a purely spiking Fourier Transformer for recognizing human activity from wearable inertial sensors with low power consumption. It introduces a frequency-domain multi-head Fourier attention mechanism and a dual-compartment spiking LIF neuron with dynamic thresholds. The authors report that it outperforms existing spiking networks and DeepConvLSTM on eight public IMU benchmarks, with about 3x faster training and an energy footprint of 0.1 pJ.
In briefSearch-free reinforcement learning for Slay the Spire 2 that restarts training from failure points
SummaryThis paper trains agents to play the deck-building roguelike Slay the Spire 2 by choosing each action with one forward pass of a neural network, without lookahead search. Training and evaluation run the game's own compiled code, and a failure-driven curriculum allocates training restarts according to where complete runs are lost. On 12,800 held-out games at Ascension 10, the released policies win 8.05% averaged over five characters.
WhyThe authors use failure-driven restart curriculum with lookahead-free single-pass policies, reporting 12,800 A10 games and controlled comparisons against uniform restart allocation on Regent.
In briefCompass 1: Persistent Open-World Robot Navigation from Single Camera Videos
SummaryThis paper presents Compass 1, an embodied navigation foundation model that unifies language understanding and video generation and can backtrack through nearly two minutes of visual history. It introduces a History Representation that combines sparse keyframes with an explicit map, and a structured VQA supervision protocol for history utilization. The authors also release a dataset of over 500 hours of real-world interactions and over 2,000 hours of annotated web videos. Reported open-world success is 79.4%, compared with 17.8% for the strongest baseline.
In briefCorrect First, Fast Second: Local 7B Code Model for GPU Kernel Optimization
SummaryThis paper studies how interaction strategies around a fixed local 7B code model (Qwen2.5-Coder-7B-Instruct) affect its ability to replace PyTorch operators with custom CUDA kernels. On 30 KernelBench tasks, it compares zero-shot prompting, profiling-guided prompting, iterative refinement on execution feedback, and a multi-agent loop. Correctness, not speed, is the main constraint, the best strategy depends on task difficulty, and a naive agent scaffold performs worse than zero-shot. Best-of-n sampling under an execution oracle markedly improves correctness.
In briefSadaqahStories: A Checked English-Arabic Charity Story Collection for Children, and a Review of AI Errors in Religious Stories
SummaryThis paper presents SadaqahStories, a parallel English-Arabic corpus of 41 illustrated four-page stories on Islamic charitable giving (Sadaqah, Zakat and Waqf) for readers aged 7 to 18. The stories were first drafted with AI help, then checked against hadith, Qur'an references and historical sources, and rewritten for children with scholar approval. An audit found substantive errors in 39 of the 41 drafts, most often unsupported specific details, invented quotations, and conflated events or people. Rewriting lowered the English text's estimated Flesch-Kincaid grade from 14.0 to 6.1.
In briefHow Model Architecture and Performance Affect Explainability: Grad-CAM and CRAFT in CNN Image Classification
SummaryThis paper studies how the complexity of convolutional neural network architectures and their training performance affect the quality of explanations from explainable AI methods. The authors trained VGG16, ResNet50, and ConvNeXt-Tiny on ten ImageNet-1K subsets and evaluated Grad-CAM and CRAFT across faithfulness, correctness, sensitivity, and simplicity. As training progressed, ResNet50 and ConvNeXt-Tiny showed improved faithfulness and lower sensitivity, while VGG16 produced simpler but less faithful and less correct explanations.
Kilian G\"oller, Haadia Amjad, Steffen Seitz et al.
In briefColor wave: systematic test of screen-light challenges for detecting injected faces in remote login
SummaryThe paper addresses injection attacks against remote face authentication, classifying injected content into static images, prerecorded videos, and real-time deepfakes. The authors implement an active screen-illumination method called color wave and build a dedicated evaluation protocol covering detection effectiveness, generalization, robustness to signal degradation, and resistance to selected evasion strategies. Results show a strong signal under controlled conditions but limited performance under overexposure, low screen brightness, and subject motion, along with vulnerabilities to some evasion strategies.
Daniyar Kurmankhojayev, Elmurat Nigmatov, Dmitrii Gordin et al.
In briefMPEG-Diffusion: Faster Video Generation via Mixed Resolution
SummaryVideo diffusion transformers produce realistic videos but are computationally expensive. This paper proposes MPEG-Diffusion, which generates the visually important region at high resolution and the rest at low resolution, using dual-resolution attention, a mixed-resolution decoder, and self-selection of the high-resolution region. Processing only 30% of full-resolution tokens, it speeds up denoising by 4.8x on Wan2.1 14B. In a user study, raters preferred it over prior mixed-resolution methods in over 90% of pairs.
In briefMIRIX-2.0: A Multi-Agent Memory System for LLM Agents
SummaryThis paper presents MIRIX-2.0, a multi-agent long-term memory system for LLM agents, aiming to be practical, accurate, and efficient. Six typed memory agents keep complete records, and a provenance hypergraph indexes episodic and semantic records. The paper reports higher accuracy than compared systems on several QA benchmarks and procedural memory tasks. The system needs no fine-tuning, works with closed and open models, and its code is public.
In briefXposedMIA: One-shot membership inference attacks using explainable AI
SummaryThis paper presents XposedMIA, a black-box membership inference attack that exploits the explanations returned by machine learning services. It shows that LIME's local surrogate geometry can serve as a proxy for a sample's distance to the decision boundary, and extends this idea to tabular and image counterfactual explanations. After a one-time calibration with as few as 30 labeled samples, the attack needs only one query per target and reaches success rates up to 0.95. The results indicate that releasing explanations creates an additional channel for membership leakage.
WhyThe authors repurpose LIME surrogate geometry and counterfactual explanations as boundary-distance proxies, reporting up to 0.95 single-query membership inference success on tabular and image classifi
Omar Hamdache, Serdar Kara, K{\i}van\c{c} Kuzey Dikici et al.
In briefWriting Medical Image Classifier Weights Directly with an LLM, Without Training Data
SummaryThe study tests whether large language models can write the weights of a medical image classifier from medical knowledge alone, without training data. Across six MedMNIST tasks, networks written by two models performed above chance, with mean test AUCs of 0.719 and 0.711, and beat zero-shot MedSigLIP on five tasks while using over 10,000 times fewer parameters. However, the same architecture trained with Adam on 256 labelled images scored higher on every MedMNIST task, with mean AUC 0.828. The authors conclude that LLMs can write functional classifiers but do not yet replace training data.
WhyThe authors report LLMs writing medical image classifier weights directly, without training data, beating zero-shot MedSigLIP but trailing Adam-trained networks on MedMNIST tasks.
Maximilian Lindholz, Richard Ruppel, Annette Milnik et al.
In briefTransferring KV Caches Between Hybrid-Attention Language Models
SummaryThe paper studies translating a language model's key-value cache into another model's format, so the target continues a conversation without re-reading the full history. The authors extend a prior grouped-query attention method to compressed MLA latents, sparse-attention indexer keys and linear-attention states, describing how to capture and inject each. Experiments on hybrid model pairs such as Kimi-VL-A3B with Kimi-Linear-48B and Qwen3.5-27B with Qwen3.8-27B report quality within the paired confidence interval of native models, and time to first token 18.7 to 46.8 times shorter than a re-read at 64K tokens.
In briefBARS: Beat-Adaptive Rap Synthesis Using Flow Matching
SummaryThis paper presents BARS, a non-autoregressive model that generates rap vocals from a beat track, lyrics, and a short reference voice. It controls rap tempo independently of the beat, allowing flexible output length without explicit duration prediction. The model uses a pretrained ByT5 encoder and an Asymmetric Scaled RoPE to improve lyric-audio alignment, and supports lyric-level editing through span masking. Experiments report high lyric fidelity, rhythmically coherent flow, and robust tempo control.
In briefCOMPASS: A learned policy for retrieving optimization templates in LLM-based solver modeling
SummaryCOMPASS is a framework for retrieving optimization templates when large language models translate problem descriptions into solver code. It organizes the template library as a directed acyclic graph and learns a traversal policy with a deep Q-network from template-level judgments and execution feedback. The method embeds problems and templates in a shared semantic space so new templates can be added without redefining the action space. Experiments across several LLM backbones and optimization-modeling benchmarks show gains over heuristic structured retrieval, largest on complex and mixed-category problems.
Nguyen Le Minh Hoang, Van Duc Cuong, Bui Trong Duc et al.
In briefA black-box reduction linking regret to multi-level conformal coverage
SummaryThis paper studies online conformal prediction, which builds prediction sets over a data stream so that their average coverage converges to a target level, even for adversarial data. It shows that isotonic projections can turn any online learning algorithm into a multi-level prediction method, with the algorithm's regret controlling coverage and consistency via Fenchel duality. Two instantiations, p-norm online mirror descent and coordinate-wise universal portfolio methods, improve the dependence on the number of levels to √log d. Experiments show favorable performance against baselines.
In briefExact Natarajan Dimension of Homogeneous Multiclass Linear Predictors
SummaryThis paper studies the Natarajan dimension, a measure of capacity for multiclass hypothesis classes that is important for analyzing generalization. By establishing lower and upper bounds, the authors determine the exact Natarajan dimension of homogeneous multiclass linear predictors with pairwise distinct class weights. The result is n(k-1)-1, where n is the feature-space dimension and k the number of classes, for n, k at least 3.
In briefLow-rank, low-precision training of Gaussian processes for faster large-scale learning
SummaryThis paper presents a scalable method for training Gaussian processes that combines AnchorNet landmark selection with an FITC correction. The objective and its analytic gradients are computed using low-rank factors and small dense matrices, avoiding the full kernel matrix. The full procedure runs in single precision and was tested on nine large-scale regression datasets, reaching reference predictive accuracy with substantially less training time, though the required rank varies by dataset. The authors also find that single precision can limit accuracy on ill-conditioned problems, and that further reductions to fp16 or TF32 give only modest speedups and can be unreliable when the system is poorly conditioned.
In briefBECON: Belief-Conditioned Constrained Multi-Objective RL Under Shifting Preferences and Budgets
SummaryThis paper addresses constrained multi-objective reinforcement learning, where agent preferences and safety budgets may drift over time and external task signals may be unreliable. The proposed BECON framework infers latent task context from recent histories to estimate preferences and safety budgets, and fuses these estimates with the exposed task signal using uncertainty-dependent gates. According to the paper, experiments on diverse multi-objective control tasks show BECON improves robustness and constraint satisfaction under task drift over state-of-the-art MORL baselines.
In briefGeo-locating ground image sequences by matching them to satellite images using attention-based fusion
SummaryThis paper addresses cross-view image sequence geo-localization, which determines where a sequence of ground images was taken by matching it to geo-tagged satellite imagery. The proposed ASML method combines a Sequential Content Fusion module, a Spatial Feature Aggregation module, and a Geo-InfoNCE loss that uses both feature and geographic distances. The authors also introduce the CrossRoad-66K dataset. On SeqGeo and CrossRoad-66K, ASML improves R@1 over state-of-the-art methods by 3.37% and 2.10%, respectively.
In briefTesting AI Agents on Industrial Chip Floorplanning: A Evaluation Framework and Study
SummaryThis paper presents a framework for judging AI agents by the physical quality of the chip floorplans they produce, checking geometric validity, congestion, timing, and resource use. It applies the framework to one industrial design with 99 macros occupying about 71% of a fixed L-shaped region. The specialized system Nova-SI scored a mean of 57.97 on a 0–100 composite quality score, compared with 31.17 for the strongest general-purpose configuration, GPT-6 Astra.
In briefDeep learning reconstruction of very sparse thyroid cancer scans using post-treatment I-131 images
SummaryThis single-center prospective study examines whether high-count post-therapeutic I-131 whole-body scans can serve as training references for reconstructing sparse diagnostic planar images in differentiated thyroid cancer. It trained five Restormer-based models on 111 patients and evaluated them with PSNR, SSIM, and blinded dual-reader scoring. Models using transfer learning from post-therapeutic data achieved the best quality, and selected configurations matched or exceeded the standard scan at twice the acquisition speed.
In briefFast four-step generation of images containing text via distillation
SummaryThis paper presents a four-step distillation framework for generating images that contain text. It distills the GLM-Image DiT using distribution matching, adds a path-length penalty to keep local responses of the conditional denoising field consistent, and uses progressive training with synthetic strings and teacher-generated images. The student reports higher GenEval and CVTG-2K word accuracy than four-step TDM, and single-image inference drops from 5.818 s to 1.742 s.
In briefSMO: Compiling Natural Language Skills into Reusable State-Move-Outcome Brick Graphs
SummaryThis paper proposes State-Move-Outcome (SMO), a framework that decomposes natural-language skills used by LLM agents into reusable procedural bricks and organizes them as a symbolic graph. On SkillsBench v1.1, 232 raw skills from 87 tasks are represented with 95 shared bricks, reducing skill representation from 366,472 to 149,154 tokens. Across 232 evaluation instances, the average task score rises from 0.6379 to 0.7478, and average runtime token consumption drops by 38.38%.
In briefPre-training framework combining event frame and voxel representations with RGB guidance
SummaryThe paper proposes CM3AE 2.0, a unified pre-training framework for RGB frames and event voxels/frames. It introduces Event-aware Spatiotemporal Masked Reconstruction, which uses event responses to direct masking toward dynamically salient regions, and Hierarchical Cross-modal Generation, which performs event-conditioned RGB generation with a pre-trained diffusion model and video-level text description generation. The authors report consistent performance gains across four downstream tasks.
In briefDPGS-SLAM: A Monocular Gaussian Splatting SLAM Method for Dynamic Scenes
SummaryThis paper introduces DPGS-SLAM, a monocular RGB SLAM framework for dense mapping in scenes with moving objects. It detects dynamic regions with optical flow and SAM2, and uses a diffusion inpainting model to generate plausible backgrounds for occluded areas, which are tracked with provenance labels so they stay out of pose estimation. Reported results include average ATE RMSE of 1.2 cm on TUM RGB-D and 2.7 cm on BONN RGB-D, and a 4.60 dB average PSNR gain over WildGS-SLAM on StaDy4D.
Yifan Luo, Takafumi Matsumaru, Vibekananda Dutta et al.
In briefSAFE-Hair: Exportable Single-Photo 3D Hair Reconstruction Anchored to the Scalp
SummarySAFE-Hair reconstructs explicit hair strands from a single portrait by decoding hair from fixed follicle positions on a canonical scalp UV domain, so outputs can be used directly in grooming, editing, or simulation. It uses a conditional Rectified Flow model to generate latent scalp fields, with attachment, collision, and bending priors to improve geometric validity. On the new HairBench-3D benchmark, it improves over the strongest baseline, reducing Chamfer distance from 3.74 mm to 3.28 mm and penetration ratio from 4.8% to 3.9%, while raising F-score from 77.9 to 81.7.
In briefScale-invariant empirical Bayes Laplace approximation for ReLU networks
SummaryReLU networks have continuous scale symmetries, so many different weight settings represent the same function. The paper shows that the empirical Bayes Laplace approximation is not well-defined for these networks, because the marginal likelihood depends on the Hessian, which changes under ReLU scaling. The authors introduce γ-canonicalization, which first converts the network into a canonical form that balances incoming and outgoing weights, then fits the approximation there. This makes the marginal likelihood invariant under ReLU scaling and yields consistent prior precision and predictive uncertainty.
In briefMambaPD: Bidirectional State-Space Model for Detecting Parkinson's Disease from Resting EEG
SummaryThis paper introduces MambaPD, a bidirectional selective state space model that detects Parkinson's disease from resting-state EEG recordings. It is compared with four baseline models (TCN, BiLSTM with attention, EEGNet, and a patch Transformer) on two public datasets, UNM and SanDiego, using subject-independent 5-fold cross-validation. Differences among the models are small relative to differences between folds. The author reports the full pipeline, per-fold results, and limitations of the evaluation protocol.
In briefSingle-image super-resolution using semantic prompts and cross-modal training
SummaryThis paper proposes a super-resolution method that combines pixel intensity with semantic information to recover image details better than models focused only on pixels. It extracts and encodes semantic data such as textual descriptions and spatial positions. A Cross-Modal Feature Fusion module generates semantic feature maps, and a Prompt Traction Head feeds these insights into the backbone network. The authors report that experiments across various datasets confirm the approach's superiority.
In briefFocusBranch: A Combinatorial Branch-and-Bound Method for ℓ0 Robustness Verification
SummaryThis paper addresses verifying neural network robustness against ℓ0 perturbations, which requires reasoning over all sparse sets of perturbed pixels. It proposes FocusBranch, a combinatorial branch-and-bound framework that uses node abstractions to decompose and branch the ℓ0 ball. The method selects its starting level and branching strategy based on sampled estimates of safe rates and costs. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 classifiers report that it outperforms state-of-the-art verifiers.
In briefTIRAD: Text-guided image residual learning for detecting unseen anomalies
SummaryThis paper introduces TIRAD, a framework that treats the difference between text descriptions of normal and abnormal states as a text residual feature. This feature guides the image residual feature through a module that converts patch-level text residuals into a gate, keeping defect-related visual differences and suppressing others. Experiments on industrial benchmarks and industrial-to-medical transfer show it outperforms existing image-only residual methods.
In briefA general transition framework for finding reduction domains in non-binary constituency parsing
SummaryThe paper presents a common transition calculus that covers top-down, in-order, and bottom-up non-binary constituency parsing. At each reduction, the rightmost delimiter marks the left boundary, the nonterminal marker gives the parent label, and the completed items within the delimited segment determine the children and their count. The authors prove soundness and completeness with a stack invariant and derive the exact length of the static oracle's action sequence. Controlled experiments on PTB and CTB show that delimiter-guided parsers reach labeled accuracy similar to their canonical counterparts across all traversal orders.
In briefFeeling Easier, Working Harder: Cognitive Load in AI-Assisted Code Comprehension
SummaryThis study had 20 novice computer science students read Java code snippets with or without generative AI assistance, then take a quiz without AI. A detection-response task measured their cognitive workload in real time. AI assistance significantly increased objective workload, raising reaction times by 11.2%, while self-perceived workload stayed the same. Comprehension scores were high in both conditions, with no clear evidence that AI improved understanding.
Maryangela White, Maximilian Rose, Matthew Buckman et al.
In briefGrasp-type labels switch a robot hand between grips for climbing holds
SummaryThe paper presents BetaGrasp, a 3D diffusion policy that takes point-cloud input conditioned on a four-class grasp label (crimp, sloper, pinch, jug). It runs on a Franka arm with a 16-DoF LEAP hand and is evaluated on a climbing-hold benchmark that measures slip force in newtons. In 80 paired hardware trials, average slip force rose from 9.6 N to 14.3 N, and fully slipped grasps fell from 44% to 10%.
Rumi Loghmani, Abraham George, Amir Barati Farimani
In briefSpeculative Self-Distillation: a more efficient way for models to learn new knowledge
SummaryThis paper proposes Speculative Self-Distillation (SSD), a method that helps language models internalize knowledge missing from pretraining, updated facts, and new domain skills. The student model generates training rollouts by default, and the teacher intervenes only where its next-token distribution substantially diverges from the student's. On three closed-book QA benchmarks, SSD nearly matches on-policy distillation while using 38-69% fewer supervised tokens. It also retains the most general capability in a forgetting stress test.
Shayan Talaei, Agam Bhatia, Arshia Soltani Moakhar et al.
In briefBenchmark testing AI models' scientific reasoning for materials research
SummaryThis paper introduces MatScope, a benchmark for testing how well language models and multi-agent systems reason about materials science. It assesses condition-specific reasoning on 200 tasks, agent-assisted knowledge transfer to unresolved questions, and forecast revision as observations accumulate. Across eleven model deployments, models handle given-input calculations and many physical explanations well, but fine-grained estimates, comparisons between similar states, and quantitatively supported choices remain substantially weaker.
In briefGraph-guided preference data for better multi-step chemical mechanism prediction
SummaryThis paper presents a graph-guided framework for predicting multi-step chemical reaction mechanisms using a chemical mechanism knowledge graph to guide large language models. It introduces GRMP, a graph-aligned predictor that uses graph neural networks to generate synthetic candidate pathways, and a relation retrieval module that pulls graph evidence to refine pathways. GRMP reports pathway accuracies of 48.1% on ChemMechQA and 57.9% on oMe-Gold, and 36.4% top-1 accuracy on 110 out-of-distribution reactions.
In briefRobust budget scheduling for merging multiple domain experts into one student model
SummaryThis paper studies multi-teacher on-policy distillation, which combines the skills of domain-expert teachers in math reasoning, code generation, and instruction following into one compact student model. Because the training budget is shared across domains, the authors apply distributionally robust optimization to prioritize the worst-performing domain. They propose RDS-MOPD, which shifts allocation only when domain differences clearly exceed noise, bounds the influence of single anomalies, and defines the worst case as regression from each domain's own best level. The method outperforms recent competing approaches and ranks first on most math reasoning, code generation, and instruction-following benchmarks.
In briefHIST3R: Using History to Correct State Updates in Streaming 3D Reconstruction
SummaryThis paper addresses streaming 3D reconstruction, where a recurrent scene memory is updated frame by frame and errors can accumulate over long videos. It splits each recurrent update into a direction and a magnitude, and notes that prior training-free methods only adjust magnitude. The proposed HIST3R, built on CUT3R, corrects update direction using locally consistent historical states and scales update magnitude by the visual difference from history. It reports improved long-horizon results in camera pose estimation, 3D reconstruction, and video depth estimation.
In briefScalable Bayesian optimization for finding ideal patterns in layered materials
SummaryThis paper introduces FruBO, a flexible Bayesian optimization framework for finding optimal configurations in computationally demanding materials design. The case study examines multilayer Moiré patterns formed by six rotated hexagonal lattices, seeking the rotation that maximizes pore-size uniformity measured by the Jain fairness index. Using a CatBoost surrogate model, FruBO found the global optimum after 1,470 acquired samples in about 1.15 hours, outperforming random sampling and Gaussian Process-based optimization.
Panagiotis Krokidas, Christos Lazaridis, Vassilis Sioros et al.
In briefHyPLM Technical Report: An Efficient Protein Language Model
SummaryThis report presents HyPLM-Flash, an open protein language model with 650 million parameters, trained by the research agent Hyra 2.0 on 726 billion tokens. Its inference engine uses fused kernels and INT8 matrix multiplication, embedding about 130,000 residues per second on one NVIDIA H20 GPU. In a matched evaluation on 32 tasks, it reaches a ProteinGym zero-shot Spearman correlation of 0.409, statistically tied with ESM-2 650M and others, and gives the best long-range contact precision of all models tested. The model weights and inference engine are released on GitHub.
WhyThe authors report matching ESM-2 650M and similar 600M-class models on ProteinGym and contact precision using less training data and faster inference across 28 public checkpoints and 32 tasks.
In briefEstimating Hypergradients for Bilevel Reinforcement Learning When the Follower's Reward Is Unknown
SummaryThis paper studies bilevel reinforcement learning where the leader cannot see the follower's reward function and can only observe its trajectories. The authors propose Reward-Estimated Hypergradients (RE-HG), which recovers the reward through inverse reinforcement learning and then computes the hypergradient. They prove that for linear rewards, the bias from reward shaping cancels when estimates are aggregated across sufficiently diverse dynamics. Experiments show gradients close to a white-box oracle with comparable performance, and the method avoids some local optima.
In briefHyInv: A Protein Inverse Folding Model Conditioned on Backbone Structure (Technical Report)
SummaryHyInv is a 1.96-million-parameter graph network that takes a protein backbone structure and designs amino-acid sequences that fold into it. It is pretrained with random-order autoregressive modeling and then refined with reinforcement learning using a structure-based reward. A grouped draft-and-rewrite sampler improves refolding TM-score at matched diversity. The report states that HyInv achieves higher average scTM than ten external baselines across five datasets, and that the model supports de novo, pocket-constrained, enzyme, and loop redesign tasks. Code and model are publicly released.
In briefInfrathon: A Benchmark for Testing AI Agents on Long-Running AI Infrastructure Optimization
SummaryThis paper introduces Infrathon, a benchmark of long-horizon tasks for optimizing real-world AI infrastructure across several levels. Six models were evaluated over 150 runs with 24-hour budgets, totaling over 10,000 GPU-hours. Model rankings change with the evaluation horizon, and most runs keep improving in the second half of their budgets.
In briefEmoHarness: Improving multimodal emotion recognition by evolving the surrounding harness
SummaryEmoHarness wraps a frozen small multimodal model in a harness of prompts, settings, and verification code. The harness labels visual observations, transcripts, and prosody descriptions as citable evidence for the model to weigh when choosing emotions. A stronger LLM offline reads the model's errors and traces and edits the harness, with validation gates deciding which edits are kept. On 5,507 MME-Emotion items, it raises task-macro recognition by 7.6 to 13.4 points over a video-only call, mainly through transcripts and prosody descriptions.
In briefMUTE: A Multi-Level Audio Protection Method Against Talking-Head Misuse
SummaryThis paper proposes MUTE, an audio protection framework that prevents speech-driven talking-head generation models from being misused in deepfakes. It adds perturbations that both disrupt cross-modal attention alignment and degrade temporal audio representations, constraining them to high-energy STFT regions for robustness. Experiments show consistent lip-sync degradation on both white-box and black-box models while preserving perceived audio quality.
In briefWhen Audio and Video Mislead: Multimodal AI for Emotion Recognition in Conversation
SummaryThe paper audits off-the-shelf omni models on conversational emotion recognition by holding the transcript fixed and adding, removing or replacing audio and video. Adding the target audio and video improved accuracy across the tested model-dataset pairs, while unrelated audio and video consistently lowered accuracy. The results separate genuine benefits from the target modalities from sensitivity to misleading input.
In briefA survey of trajectory density ratios when two processes have weak overlap
SummaryThis survey reviews how many scientific fields compare two stochastic processes through the sequences they generate, including reinforcement learning, molecular simulation, climate attribution, and language-model post-training. It explains that the trajectory density ratio may not exist when the processes' supports differ, or its variance may grow uncontrollably with sequence length. The paper proposes a three-level diagnosis of weak support and maps nine methods used when the full ratio is unavailable, and it closes with open problems.
In briefThermo-VLA: A Robot Manipulation Model That Uses Thermal Sensing
SummaryThis paper presents Thermo-VLA, a vision-language-action model for robots that uses thermal images alongside RGB. Its ThermoRes Encoder splits thermal images into cold and hot attribute representations and feeds them into visual tokens through language-conditioned residual modulation. The authors tested six real-world tasks on a dual-arm Unitree G1 robot, where the model reached an average success rate of 0.86 on four thermal-dependent tasks, 18 points above a π0.5 baseline. The same modulation mechanism was also applied to colour attributes with improvements reported.
In briefAnyBreak: Jailbreaking Any Text-to-Image Model for Any Concept
SummaryThis paper presents AnyBreak, a jailbreak framework that builds adversarial images through multi-scale progressive semantic alignment. These images steer the latent-space generation of text-to-image models toward unsafe content, transferring malicious concepts from text to image. A variant, AnyBreak-V, adds a compensation term to bypass post-hoc safety checkers. The authors report high attack success rates in a fully black-box setting using only clean text prompts, and argue that existing defenses leave cross-modal attack surfaces inadequately addressed.
In briefFireGate: Limiting Behaviour Change Before Continual RL Fine-Tuning
SummaryThis paper addresses forgetting when a deployed language model is fine-tuned on tasks one after another, proposing FireGate to bound output changes before training. Each task's update, such as a new LoRA, sits behind a classifier gate that decides per prompt whether the update is used, and where no gate fires the model matches the old one exactly. On reinforcement-learning streams of up to five tasks, gated training changed 3.3% of held-out chat responses versus 91.8% for ungated training.
In briefMixRoute: Rethinking Single-Distribution Training for Generalizable Neural Routing
SummaryThis paper proposes MixRoute, a neural framework for routing problems that adaptively combines normalization statistics at different granularities to handle varying distribution shifts. The authors also build a benchmark of 179 fine-grained datasets across three representative routing problems. Trained only on uniform distributions, MixRoute generalizes to many unseen distributions without adaptation, and the authors report state-of-the-art zero-shot results.
In briefShiftCache: a fast, exact way to predict DNA insertion and deletion effects
SummaryThe paper presents ShiftCache, a method for predicting the effects of DNA insertions and deletions (indels) using sequence-to-function models. It encodes each shifted reference sequence once and shares the result among all variants in the same window with the same shift. On held-out AlphaGenome windows, it runs 3.5 to 6.5 times faster than deduplicated dense execution. For default AlphaGenome and Borzoi runs, its outputs match dense execution within a tolerance of 10^-5.
In briefAutoQS: Automated Quantization Strategy Design with Large Language Models
SummaryQuantizing large language models reduces deployment costs, but effective strategies must balance quality and storage. AutoQS uses LLMs to repeatedly propose and evaluate group-size allocations across layers and operators under deployment constraints, keeping weight precision fixed. It combines an inexpensive quality proxy with periodic actual quantization to guide search within a limited evaluation budget. On Qwen2.5-7B-Instruct at four-bit weight precision, AutoQS improves seven-task average accuracy over uniform-group GPTQ and AWQ by 0.70 and 1.12 percentage points.
In briefCalibrating Regret for Robust Contextual Optimization Beyond Prediction Errors
SummaryThe paper introduces Conformal Regret Sets, a post-hoc, model-agnostic framework for robust predict-then-optimize decision-making. It conformalizes SPO+, a pointwise upper bound on exact regret, and uses the calibrated level set to refine decisions for minimax regret. Experiments on contextual fractional knapsack show that continuous refinement reduces the 0.95-quantile of regret by 9.1% and CVaR at 0.95 by 13.1% relative to nominal PTO.
In briefA pairwise distortion distribution for compression and quantization
SummaryThe paper introduces the pairwise distortion distribution (PDD) and its asymptotic limit, the pairwise dimension (PD), which extend the correlation integral to general distortion measures. It shows that PD bounds the Rényi information dimension (RID). It also shows that PDD gives a lower bound on the performance of random codebook quantization, which can guide compression, quantization, and dimensionality reduction.
Carl Nuzman, Jeongmin Chae, Kursat Rasim Mestav et al.
In briefChainCheck: Auditing Whether Evidence-Sufficiency Scores Really Measure Multi-Hop Reasoning
SummaryThe paper proposes ChainCheck, an audit method that tests whether benchmarks for evidence sufficiency truly measure multi-hop reasoning. It finds that a language-model scorer that never reads the reasoning chain, and even a one-string match, can achieve very high scores, while the 11 evaluated systems score between 0.308 and 0.484. The authors therefore interpret positive CS4 results as operational bridge-consistency sensitivity rather than semantic multi-hop reasoning. They also report that patterns found on one set of datasets did not replicate on an independent HotpotQA set.
In brief3R-Adapter: A Method for Efficiently Adapting 3D Reconstruction Models via Retrieval, Rewiring, and Refinement
SummaryThis paper addresses the weakness of feed-forward 3D reconstruction models in long-tail scenarios such as low-light reconstruction, joint human-scene reconstruction, and test-time adaptation. The authors identify the collapse of cross-view relational attention structure as a main cause and frame adaptation as recovering correspondences under domain shift. They propose 3R-Adapter, a parameter-efficient fine-tuning method with Deformable Retrieval, Relational Rewiring, and Consistency Refinement components. Experiments show improved robustness over existing parameter-efficient fine-tuning methods while keeping adaptation lightweight.
In briefShaping Useful Noise: Energy Distributions Predict Visual Pretraining Quality
SummaryThis paper proposes using a dataset's energy profile, the distribution of log-probability levels under a normalised real-image density model, to audit and curate synthetic data for visual pretraining. It introduces Energy-Level Distribution Distance (ELD), which compares real and synthetic energy distributions. The authors report that ELD reveals mismatches missed by FID and predicts pretraining performance better than FID, LPIPS variation, and feature-space recall in CNN and ViT self-supervised settings. They also use ELD to resample and mix synthetic sources, matching the real-image energy histogram while preserving diversity and improving downstream transfer.
Ching Lam Choi, Antonio Torralba, Phillip Isola et al.
In briefHospital operations-aware multimodal system for assessing ICU discharge readmission risk
SummaryThis paper presents a clinical decision support system that predicts unplanned 30-day ICU readmission by combining patient data with hospital operational context. It integrates progress-note text, vital signs, departmental workload indicators, admission flow features, and medical codes. Evaluated on 20,000 ICU stays from MIMIC-III, an end-to-end multimodal fine-tuning model achieved the highest AUROC (0.8288) and PR-AUC (0.4373), and the authors report that operational indicators supply independent risk signals.
In briefCHO-Bench: How Context Hints Affect Ad Rankings in LLM Advertising
SummaryThis paper studies how advertisers can write natural-language context hints to improve how their ads are ranked in LLM-based advertising, where ads are chosen from user conversations. The authors define this as Context Hint Optimization and create seven hint strategies covering what, who, and when, plus a unified strategy, releasing a benchmark called CHO-Bench. Across over 1.19 million paired evaluations on five datasets, context hints improved rankings, but strategies performed similarly on average and gains from larger hint budgets diminished.
In briefBharatSynthQA: A Question-Answering Dataset on Indian Temples and Religious Heritage
SummaryThis paper presents BharatSynthQA, a passage-grounded question answering dataset about Indian temples and related religious sites. Drawing on Wikipedia articles, the authors generated Boolean, factoid, and explanatory question-answer pairs with a large language model, then filtered them for correctness, retaining 98,677 pairs, and released 300 human-annotated pairs as a reference set. Benchmarking five current LLMs, factoid token F1 was 0.18 to 0.30 without the source passage and 0.73 to 0.77 with it.
Riya Naik (1), Shivam Sharma (2), Swati Agarwal (3) ((1) BITS Pilani et al.
In briefUltraVoxelGS: Fast 3D Ultrasound Reconstruction with Feed-Forward Gaussian Splatting
SummaryThis paper presents UltraVoxelGS, a method for 3D ultrasound reconstruction that predicts a complete Gaussian scene directly from posed ultrasound images without per-scene optimization. It lifts all slices into a fixed-size voxel support to predict Gaussian parameters, and uses an ultrasound-aware rendering strategy that separates reflection from cumulative attenuation. On real and simulated datasets, it achieves the best PSNR/SSIM among per-scene optimization baselines while reducing inference time from 10-20 minutes to about 10 seconds.
WhyThe authors report a first feed-forward 3D ultrasound Gaussian method, cutting inference from 10-20 minutes to about 10 seconds with best PSNR/SSIM on real and simulated data.
In briefSelf-Auditing Visual Reliance for Video Reasoning
SummaryThis paper proposes VIOS, a post-training method for video reasoning that identifies tokens where a model underuses visual evidence. It measures each prediction's dependence on the video by comparing next-token distributions with and without the video, and contrasts a student with an answer-privileged teacher to find where the teacher relies on visuals the student ignores. VIOS turns this signal into dense, targeted supervision using an EMA teacher, a contrastive objective, and a reliance-gap gate, requiring only a single on-policy rollout. The authors report consistent gains across three backbones on 12 benchmarks, with fewer training steps, FLOPs, and generated tokens than prior methods.
In briefHow much demonstration data is each operator worth? Time limits and operator mixes in behavior cloning
SummaryThe paper measures how many demonstrations from one group of robot teleoperators equal those from another, using a data multiple defined by the demonstrations needed to reach the same success rate. Behavior-cloning policies were trained on robomimic's Square and Can tasks using demonstrations from two teleoperator groups. On Square with a 25-second limit, an all-Worse set needs 1.55 times as many demonstrations as an all-Better set, mostly due to later completion; at 100 seconds the multiple drops to 1.11.
In briefDecision-aware estimation of origin-destination travel demand using structurally constrained traffic data
SummaryThis paper proposes a decision-aware framework for estimating origin-destination travel demand. It uses structurally constrained traffic data to define a feasible demand region and selects estimates that minimize regret in downstream decisions. A geometric approximation of the feasible region and a convex surrogate of regret make the optimization tractable. Experiments show the framework reduces decision regret compared with conventional estimation benchmarks.
In briefHigher-order attention: when it helps and when it doesn't
SummaryThis paper examines why standard Transformers fail at graph search and the card game SET, while higher-order attention has an uneven record. The authors treat each layer as computing the partition function over a family of witnesses and reading it at a temperature, yielding a depth law (L layers reach 2^L - 1 steps) and a size law (temperature times margin must exceed the log of the candidate count). Models built on these laws achieve results on graph tasks comparable to or better than much larger models.
In briefConstrained variational policy optimization extended to hybrid action spaces
SummaryThe paper addresses constrained reinforcement learning where agents must coordinate discrete modes and mode-conditioned continuous controls under long-horizon cost constraints. It extends Constrained Variational Policy Optimization to hybrid action spaces as H-CVPO, using exponential reward-cost reweighting to form a joint target distribution and fitting separate discrete and conditional-continuous policy components under trust regions. Experiments in simulated hybrid-control environments report more favorable reward-cost trade-offs than hybrid-action baselines.
In briefMonocular fisheye visual odometry with scale-consistent depth learning guided by spherical bundle adjustment
SummaryThis paper presents S²D-FVO, a monocular fisheye visual odometry framework that combines spherical bundle adjustment (SBA) with unsupervised depth learning to improve scale consistency and depth accuracy. SBA uses tangent-space angular reprojection residuals and a median depth-based scale regularization term. Experiments on multiple public datasets and two newly collected large-scale datasets show better depth and trajectory accuracy than representative existing methods, across different camera models.
In briefCan LLM Agents Organize Their Own Memory? Introduces SoMem, a framework for storing and using persistent agent memory as linked ordinary documents.
SummaryMost LLM agents keep fixed weights after deployment, so new information must live in external memory, which existing methods often manage with predefined records and rules. This paper proposes SoMem, which stores memory as linked documents that the agent itself summarizes, groups, and revises, with each summary tied to original interactions as evidence. On LoCoMo, BEAM-100K, and LongMemEval-S, SoMem outperforms the strongest baseline by 2.66, 3.10, and 2.00 percentage points, and the memory can be reused across models.
In briefLanguage processing in the autistic brain: a brain-imaging meta-analysis across modalities and age
SummaryThis meta-analysis combined 45 fMRI and PET studies involving 753 autistic and 777 non-autistic participants to compare brain activity during language processing. Both groups activated a similar left-lateralized language network. Autistic participants showed reduced activity across left perisylvian cortex and increased activity in the right temporal pole, indicating weaker leftward lateralization.
In briefText Alone Predicts Which Classes a Segmentation Readout Keeps
SummaryOpen-vocabulary segmentation models can silently lose classes beyond their training categories: they keep emitting confident masks, yet the class never wins a pixel. The paper finds that which classes collapse is determined by the class-name list alone, so collapse risk can be ranked from text similarity to the training vocabulary before any image is seen. The authors also report that collapsed classes survive in the frozen features, and lowering the readout's amplitude or stopping training early restores them by up to 16 points without retraining.
WhyThe authors predict which classes collapse from class-name text similarity alone, reporting AUROC 0.69-0.96 across nine model families and twelve backbones, plus retraining experiments that flip gains
In briefTesting whether a pretrained table model can act as a world model without training
SummaryThis paper examines whether a pretrained tabular foundation model can serve as a training-free dynamics model by framing next-state prediction as in-context regression over transition pairs. On synthetic dynamical systems and robot locomotion benchmarks, it is more accurate than LSTM and MLP models trained from scratch on the same limited data in almost all settings, but models trained on the full dataset remain substantially better on most systems. Multi-step rollout accuracy depends strongly on the amount of context provided, and inference is about three orders of magnitude slower than an LSTM. The authors conclude that training-free priors are useful when data are scarce and context is sufficient, but are not yet a substitute for trained world models in planning.
In briefLBSNet: lightweight EEG brain-state model with learnable spectral prototypes for fatigue detection
SummaryThis paper proposes LBSNet, a lightweight network that models brain states with spectral prototypes to link local cross-band spectral structure to temporal aggregation. Each window's spectral pattern is matched to learnable prototypes, and the resulting assignments set its temporal weight. Subject-independent experiments on SEED-VIG and SADT report better performance than seven baselines at low computational cost.
In briefCRAFT-GBM: Choosing the Best Features for Faster Gradient Boosting
SummaryThe paper introduces CRAFT-GBM, an extension to LightGBM that uses split-gain evidence gathered during training to adaptively decide which features, and by default how many, are searched in each boosting round. It provides statistical guarantees: with probability at least 1-δ, the features with the largest expected gain are retained after initialization. Across 14 dataset families and four noise-augmented settings, the fully endogenous variant reduced per-fit training time by 39.74% and lowered test loss by 1.04% in the geometric mean.
In briefCHORD: A Multi-resolution Compressor for Multimodal Time Series
SummaryCHORD is a coarse-to-fine compressor for multimodal time series from wearables, mobile devices, and IoT sensors, so one stored payload can serve downstream models at different temporal resolutions. It uses a residual cascade that builds corrections through learned shapelet-inspired atoms, with a shared pathway for common synthesis across modalities and private pathways for sensor-specific morphology. Finite scalar quantization fixes the bit width per modality and scale, so the payload size is known in advance. On seven wearable-sensing and forecasting benchmarks at an 8x target, CHORD's classification accuracy stays within 0.62 points of uncompressed data on average.
Bibhus Luitel, Payal Mohapatra, Subrata Biswas et al.
In briefLO-World: Unified BEV Latent Model for Joint 4D Occupancy and LiDAR Forecasting
SummaryThis paper proposes LO-World, a unified 4D world model that forecasts both LiDAR point clouds and 3D semantic occupancy within a shared bird's-eye-view (BEV) latent space. It combines an efficient hybrid range-voxel autoencoder with a flow-matching generator that caches historical context. Experiments report a 10.9% gain in occupancy forecasting IoU and a 30.1% reduction in LiDAR Chamfer Distance, while running above 25 FPS on a single RTX 4090.
In briefLate but Valuable: How Deadlines Change the Measured Value of LLM Reasoning
SummaryThe paper argues that when an LLM's answer in an automated laboratory arrives after its deadline, the system runs a safe fallback instead, so measuring answer quality and latency separately can overstate reasoning's value. On synthetic sample-decision tasks, the authors compare a fast direct-answer mode with a thinking mode and find that late thinking answers were worth about five times more than on-time ones. In a single-GPU Qwen3-8B experiment, raising concurrent requests from 8 to 56 cut the share of thinking answers meeting a 45-second deadline from 79.2% to 1.4%. The authors recommend evaluating each answer's value and arrival time together, and scoring the action that actually runs.
Sina Barazandeh, Benjamin E. Lundell, Jose Lugo-Martinez
In briefLayer Roles Matter: Adaptive Layer Profiling to Reduce Modality Imbalance in Multimodal Emotion Reasoning
SummaryThis paper addresses modality imbalance in multimodal emotion reasoning with multimodal large language models, identifying over-reliance on a single modality and insufficient cross-modal use as two failure modes. The authors find that different layers of these models play separable roles, stabilizing unimodal representations or integrating cross-modal information. They propose LAMR, which uses Modality Entropy and Cross-modal Intensity to profile layers and applies tailored interventions to each. Experiments on two benchmarks show improved emotion reasoning across several backbones without additional training.
In briefRBA: A Lightweight Adapter for Interactive Retinal Vessel Segmentation
SummaryThis paper proposes the Read-out Basis Adapter (RBA), a parameter-efficient method for interactive vessel segmentation in retinal OCTA images using the SEEM model. RBA adds about 0.47 million trainable parameters by generating image-specific basis maps at the mask read-out stage. On OCTA-500 3mm large vessels under simulated clicks, it outperforms LoRA, Conv-LoRA and AdaptFormer, reaching an average IoU of 85.16% and Dice of 91.93%.
In briefA study proving global convergence of the stochastic Adam optimizer for physics-informed neural networks
SummaryThis paper provides theoretical convergence guarantees for the Adam optimizer when training wide two-layer neural networks, using both full-batch and minibatch gradients. It extends the analysis to physics-informed neural networks (PINNs) solving the Poisson equation, establishing global convergence for both deterministic and stochastic Adam. The authors also characterize how minibatch size affects convergence.
WhyThe authors give global linear convergence for Adam on wide two-layer networks and extend the analysis to PINNs for the Poisson equation.
