In briefAntalia-2 Mini: A 7.6M-parameter Turkish speech synthesis model trained only on synthetic audio
SummaryThis paper presents Antalia-2 Mini, a single-voice Turkish text-to-speech model with 7,620,931 parameters. All its training audio came from the teacher model VoxCPM2, and automatic checks kept about 1,790 hours. On the 495-sentence Freya-TR-Eval benchmark, version 1.0.1 reached a word error rate of 1.09% and a character error rate of 0.22%. A paired bootstrap cannot distinguish it from EMA Lightning, an 8.6M Turkish model.
In briefFastWild-GS: Fast 3D scene reconstruction from varied internet photos using compact per-image appearance adaptation
SummaryThis paper presents FastWild-GS, a method for reconstructing 3D scenes from internet photos that differ in exposure, lighting, and transient objects. Each photo is described by a 128-dimensional appearance token and a six-parameter per-channel scale and offset, and a small MLP predicts Gaussian colours from these instead of spherical harmonics. On PhotoTourism it trains in 8.8 minutes per scene with the lowest LPIPS among compared methods, and on NeRF-OSR it achieves the best average PSNR, SSIM and LPIPS in 9.7 minutes per scene.
In briefRelaxing cross-modal pairing for multimodal remote sensing pretraining
SummaryThis paper examines whether multimodal remote sensing pretraining needs exact instance-level pairing across modalities. It proposes Semantic Pairing (SemPair), which pairs class-consistent samples from different instances instead, and uses a latent-factor model to argue that shared semantic structure is preserved. It also introduces Mixed Pairing (MixPair) to study pairing granularity along a spectrum. Experiments show SemPair matches exact pairing in transfer performance, while MixPair reveals a non-monotonic relationship with the instance-pairing ratio.
In briefCombining graph and pairwise learning to infer gene regulatory networks from single-cell RNA-seq data
SummaryThis paper addresses inferring gene regulatory networks from single-cell RNA-seq data, noting that graph-based methods tend to overfit while pairwise methods ignore global topology. The authors propose scGPL, an end-to-end framework that jointly models global regulatory topology and local interactions using importance-guided dynamic subgraph sampling. Across seven benchmark datasets, it reports state-of-the-art performance, with up to 13.7% improvement over the strongest baseline.
In briefAnchorMTS: Multi-View Forecasting for Long-Term Time Series
SummaryThis paper proposes AnchorMTS, which builds visual and textual views from each observed time-series window and uses them to refine a complete numerical forecast. A trainable path produces a candidate forecast, and a parameter-free projection bounds its deviation from the numerical anchor for each variable. Across six benchmarks and four horizons, it achieves the best macro MSE/MAE (0.336/0.367) among seven supervised baselines, and removing the auxiliary path raises macro MSE to 0.340.
In briefLLM validation of minimal-pair language tests makes them easier for the models tested
SummaryThe paper examines whether large language models can replace costly human judgments for validating minimal-pair benchmarks such as BLiMP (English) and QFrBLiMP (Quebec French). It finds that LLM-based filtering removes items that language models fail but humans do not, so the remaining items are easier for the models being tested. The authors release code, validator outputs and item-level scores.
WhyThe authors test LLM validation of minimal-pair benchmarks against human panels, reporting that filtering makes items easier for models across BLiMP, QFrBLiMP and BLiMP-NL.
In briefThe Target Is a Predicate: Predicate-State Grounding for Semantic-Level Video Reasoning Segmentation
SummaryThis paper introduces Predicate-State Video Reasoning Segmentation (PS-VRS), where a query defines a dynamic predicate over objects and time rather than a fixed object. The authors build PS-Bench, with 2,077 videos and 16,250 queries annotated with active/vacant states, temporal intervals, and pixel masks. They propose Predicate-State Grounding (PSG), a three-stage pipeline that predicts state tokens, aligns them to frames, and emits SAM2 masks only for active states. PSG outperforms baselines by 9.32 J&F on PS-Bench and significantly reduces false-positive mask emissions.
In briefCP-GS: Jointly Balancing Gaussian Count and Precision for Compact 2D Image Representation
SummaryThis paper presents CP-GS, a framework for representing images with 2D Gaussians under a limited budget. It jointly optimizes how many Gaussians are used and how precisely their attributes are quantized, under a shared rate constraint. The Gaussian budget is reallocated from redundant regions to areas with higher representation demand, and quantization is distortion-aware. Experiments on DIV2K and Kodak report better rate-distortion performance than representative 2D Gaussian and implicit neural representation methods across a range of budgets.
In briefMultimodal reinforcement learning method combining action-prefix prediction and cross-modal alignment
SummaryThis paper addresses the difficulty of learning representations from multimodal observations in reinforcement learning with limited interaction data. It proposes Improved Predictive Representation (IPR), which predicts future representations directly from an observed anchor and an action sequence, combined with contrastive cross-modal alignment. On five DeepMind Control Suite tasks, IPR-H achieved 10.0% higher average return than the baseline with visual and depth images and 3.4% higher with visual images and position vectors. In Webots, both IPR configurations outperformed the baseline without LiDAR but not with LiDAR available.
In briefSoftness Is Also Needed: Information Metrics of a Passive Walker with a Compliant Ankle
SummaryThis paper examines the gap in adaptability and dexterity between robots and animals, arguing that robots should be modeled and controlled using principles from biology such as morphological computation. Using passive dynamic walking models, it starts from the McGeer compass gait and adds knee-ankle compliance, foot-rocker effects, and a Ruina-style compass gait with distributed leg inertia. The authors estimate viable phase-space volumes W(γ) using state-dependent Monte-Carlo and reproducing kernel Hilbert space methods.
Fabio Bonsignorio, Elisabetta Barletta, Sorin Dragomir et al.
In briefOmni-JEV: A model for deciding full-duplex speech interaction actions from text or dual-stream speech
SummaryThis paper introduces Omni-JEV, a model that decides how a full-duplex voice assistant should act during a conversation. It takes text or separately encoded user and assistant speech and outputs probabilities over five interaction actions. The model is trained with supervised fine-tuning followed by reinforcement learning, and the Gaussian variant reports 84.08% test accuracy on audio and 96.55% on text.
In briefDisasterAgent: An evidence-driven agent that generates verifiable disaster reports from satellite imagery.
SummaryDisasterAgent is an AI agent that writes disaster reports from pairs of optical remote sensing images taken at different times. It first builds a spatiotemporal graph that stores the changes caused by the disaster, then attaches citations linking each statement to its supporting visual evidence. Experiments on the DisasterM3 and RSCC benchmarks show it outperforms general vision-language models and other remote sensing agents, with score gains over the original model.
In briefRF3DGS: Removing Reflections in 3D Scenes Using Self-Derived Cues
SummaryThis paper introduces RF3DGS, a method for removing reflections from 3D scenes reconstructed from multi-view images. It uses a photometric prior that the minimum pixel intensity across views of the same 3D point tends to reflect reflection-free content, and warps information from auxiliary views to build clean pseudo-labels for self-supervised optimization. An Adaptive Twin Rendering loss helps preserve high-frequency details from the original images. The authors report state-of-the-art results on synthetic and real scenes and release the first 3D reflection removal benchmark built on the Unity engine.
In briefFrom Depth to Safe Depth: Uncertainty-Aware Monocular Perception for Mobile Robot Navigation
SummaryThis paper examines whether a single RGB camera can replace depth sensors for mobile robot navigation. The authors find that off-the-shelf monocular depth models systematically overestimate the distance to nearby obstacles, causing frequent collisions in simulated indoor navigation. They propose a lightweight module on top of a frozen depth model that predicts a depth correction and uncertainty to produce a conservative "safe depth." In their experiments, collisions fall from 63% to 3% and navigation success rises from 37% to 70%.
In briefSelecting Feature Interactions by Checking Whether They Replicate Across Independent Models
SummaryThis paper introduces LENS, a model-agnostic framework for selecting feature interactions. It trains two models with the same learning algorithm on disjoint data and accesses them only through predictions, removing main effects to measure each feature pair's interaction strength. It tests whether the second model reproduces the pattern at a meaningful strength, using a single multiple-testing correction and only two model fits. Experiments on synthetic designs, French motor insurance data, and two credit datasets show it retains few pairs while maintaining or slightly improving predictive performance.
WhyThe authors select feature interactions via independent replication, prove a strength floor restores selection consistency, and report results on synthetic and real data.
In briefSynchrony-based communication for modular neural networks
SummaryThe paper introduces SyncNet, a modular neural network that separates module content from a dynamical phase state. Learned Kuramoto dynamics evolve the phases, and phase alignment determines communication through a shared channel. Evaluated on Sort-of-CLEVR and SQOOP, SyncNet achieves the best held-out accuracy at the three lowest pair varieties, leading by 5.1 percentage points at the hardest setting. A content-addressed variant trained from scratch matches it, suggesting the gains stem from modular decomposition rather than synchrony itself.
Nikolas Willems, Qiyang Sun, Bj\"{o}rn W. Schuller
In briefMixed Signals: Weight-Space Programming for EEG and ECG
SummaryThis paper explores using language models to read and edit the weights of neural networks, offering an alternative to gradient-descent training. The authors test the approach on EEG and ECG biosignal models, finding that frontier models can fix planted faults without being told what they are and can write network weights from scratch using only the architecture, task, and a few training subjects. In one experiment, a model matched standard training with six times fewer subjects and outperformed classic hand-crafted EEG pipelines built from the same five subjects. The authors note limitations, including difficulty improving already well-trained classifiers and unreliable identification of a network's task from its weights alone.
Dara Varam, Taraneh Aminosharieh Najafi, David Atienza
In briefMatching OCR text to ground truth word by word despite layout and reading-order errors
SummaryThis paper addresses aligning OCR output with ground-truth text word by word, a step needed for OCR post-processing. The authors revisit the Zoning Problem from their 2019 paper and introduce a protocol that tests correctness when reading order is permuted. They compare several alignment methods across synthetic noise, historical newspapers, geometry-derived gold data and a small word-box set. Local tracking proves fragile after large displacements, and unique-anchor methods fail when such anchors are scarce.
In briefWhen Covariance Is Not Enough: Kernel-Based Regularization for Single-Cell Counterfactuals
SummaryThis paper studies counterfactual editing of single cells, predicting how a cell would look if one factor were different. The authors show that the linear cross-covariance penalty used in prior concept bottleneck models fails to fully separate known and unknown factors when non-linear interactions exist. Adding a normalized Hilbert-Schmidt Independence Criterion (HSIC) recovers true intervention effects in simulations. Applied to a complex single-cell RNA-seq dataset, the non-linear model improves generation of counterfactual expression profiles across tissue environments.
Korbinian Traeuble, Edward de Brouwer, Archit Verma et al.
In briefDo Language Models Feel Sentiment? A Study of Emotion Vectors
SummaryThis paper examines emotion vectors, directions in language model representations linked to emotional concepts, in Qwen2.5-0.5B-Instruct and Qwen2.5-3B-Instruct. Eight directions from GoEmotions are tested through sentiment probing, activation interventions, and supervised correction. Emotion-only probes reach 88.66% and 90.70% accuracy on binary TweetEval, but the findings do not establish a uniquely emotional mechanism.
Ali Hassan Khan, Saad Bin Ather, Muhammad Saif et al.
In briefADMM-based plug-and-play method for inverse problems using flow matching priors
SummaryThis paper proposes an ADMM-based plug-and-play framework for flow matching priors that uses auxiliary variables and dual updates to enforce consistency between the measurement model and the prior. It also introduces a trajectory correction mechanism that combines deterministic and stochastic projections to restore consistency with the flow. The authors report high-quality reconstructions with reduced computational cost.
In briefLaser beam width and noise effects on tree diameter measurement by laser scanning
SummaryThis study examines errors from laser beam width and measurement noise in mobile laser scanning and how they affect stem diameter estimates. Simulations and experiments show that noise whose variance differs between range and scan angle, together with the lidar beam width, makes diameter bias depend on both distance and tree diameter. For trees with a diameter at breast height above 10 cm, the bias can be predicted and compensated using only sensor specifications, reducing stem diameter error to sub-centimeter level at boreal forest sites. For trees below 10 cm, bark and stem out-of-roundness limit accuracy.
Jesse Muhojoki, Nuutti Ervasti, Eric Hyypp\"a et al.
In briefDeterministic, reproducible AI inference using exact integer arithmetic, enabling verifiable distributed inference
SummaryThe paper presents tritify, an open-source runtime that runs model inference entirely in exact integer arithmetic, using floating point only for offline quantization and verification. The authors report bit-identical outputs across CPUs, instruction sets and CPU/GPU, decision agreement with a float reference, and identical results when computation is split across multiple machines. From this they argue that distributed inference can be verified by re-deriving the exact output bytes, though protection against a dishonest server is left to future work.
WhyThe authors run real models' forward passes in exact integer arithmetic, report bit-identical outputs across hardware and machines, and validate on several models including a 106B MoE.
In briefWeighted-AoI-based UAV trajectory and reporting for tracking dynamic channel knowledge maps
SummaryExisting channel knowledge map construction assumes quasi-static environments and fails when channels decorrelate rapidly. The paper studies continuous online tracking of dynamic channel knowledge maps over a 2D airborne plane using a UAV fleet, formulating a long-term error minimization problem under UAV mobility and reporting constraints. It proposes a scheduling approach based on Weighted Age of Information (W-AoI), which combines temporal staleness with spatial uncertainty predicted by a 2D U-Net, and proves that top-K W-AoI selection maximizes a one-slot error reduction bound. Simulations show steady-state error reductions of 3.8 dB² and 4.8 dB² over AoI-only and FIFO baselines, closely approaching an idealized oracle.
In briefTRUER: Fixing Test-Time Rewards with an Error-Aware Corrector for RL
SummaryThis paper introduces TRUER, a framework for test-time reinforcement learning in language models. Standard TTRL uses majority voting over the model's own answers as rewards, which can reinforce incorrect solutions, and the authors find that false positive rewards dominate reward errors. TRUER trains a reward rectifier with limited labeled data to separate correctness from consensus, emphasizing cases where incorrect majorities and correct minorities disagree. The rectifier is then frozen and used to correct rewards during test-time adaptation. The authors report that TRUER consistently outperforms naive TTRL in reward accuracy and model performance across all evaluated configurations.
In briefToward Autonomous RTL and Synthesis Optimization: Benchmarking and Self-Improving Agents
SummaryThis paper explores using large language models and autonomous agents to optimize hardware design, covering realistic RTL optimization benchmarks, a multi-agent framework called Dr. RTL, and agent-driven synthesis tuning. Dr. RTL automates critical-path analysis, RTL rewriting, and implementation evaluation, and improves through reusable optimization skills. On 20 real-world RTL designs, it reports average WNS/TNS improvements of 21%/17% and a 6% area reduction over commercial synthesis.
In briefReusing External Advice in Reinforcement Learning via Latent State Encoding
SummaryThis paper proposes Encoder BPA, a method for improving Deep Interactive Reinforcement Learning. A trained latent encoder maps state sequences into latent representations, which are used to retrieve and reuse stored external advice. Top-k aggregation and a consistency condition are used to make advice reuse more stable. Experiments report better early-stage learning efficiency and advice utilisation than other Deep IntRL approaches, especially in complex, partially observable environments.
In briefTreating Masks as Hints to Remove Video Overlays Robustly
SummaryThis paper presents HintErase, a method for removing burned-in subtitles, karaoke lyrics, titles, credits, logos and watermarks from video. It treats the detector's mask as a hint rather than a hole, training on full video with a hint channel corrupted like real localizer errors. The model removes overlays the hint misses, keeps background it wrongly covers, and preserves genuine scene text. The report also describes a streaming pipeline for full-length deliveries and shows qualitative results.
In briefAutoManifold: Agent-Designed Data Visualisation Algorithms via Manifold Embedding
SummaryThe paper introduces AutoManifold, a pipeline in which multiple large language model agents compose new manifold embedding algorithms from a constrained set of affinity, cost, and optimisation primitives. Each design stage is conditioned on user-specified structure-preservation preferences and refined through an LLM-guided, metric-based iterative loop. On real-world datasets, the generated algorithms are compared with t-SNE, UMAP, and PaCMAP, and the authors report that they better steer toward user-specified properties while preserving the intrinsic structure of the data.
In briefFedVP: Federated Learning Across Heterogeneous Models by Pooling Locally Trained Verifiers
SummaryThis paper proposes Federated Verifier Pool (FedVP), a federated learning approach in which clients collaborate through shared verification instead of merging model parameters. Each client trains a verifier on local data and responses from its own LLM, and a server combines these verifiers' judgments to select answers. Evaluated on five public medical question-answering datasets with Qwen and LLaMA backbones of 3B to 14B parameters, FedVP improves performance for all clients and beats the strongest conventional federated baselines by 1.46 to 5.79 percentage points in mean accuracy.
In briefAgentic AI and Formal Verification for Hardware Design
SummaryThis invited paper surveys how agentic AI is advancing hardware formal methods across specification, proof construction, and reasoning engines. It examines how formal methods support reliable AI-driven hardware development through RTL generation, debugging, and correctness-preserving optimization. The authors review representative research and their recent contributions, and discuss future directions toward proof-native hardware development.
In briefSeqFold: Compressed, Timestamp-Addressable User States for Action Sequence Backbones
SummaryThis paper studies how to export the multi-layer user states of a frozen long-history sequence backbone as compact representations that can be looked up by timestamp for downstream training backfill and serving. The proposed SeqFold combines learned multi-channel fusion, channel-specific temporal decay, and Lookup-Free Quantization (LFQ) to turn backbone states into compact discrete codes, which task-specific decoders map into downstream embedding spaces. On large-scale proprietary datasets, it cuts per-state representation storage by 3,400 times while keeping most of the uncompressed model's downstream ranking performance.
In briefTopology of Network Weights Under Changing Labels
SummaryThis study examines how the persistent topology of neuron weight point clouds changes under stable and non-stationary supervision. An identical multilayer perceptron was trained on Fashion-MNIST with correct labels, a fixed random label permutation, or a fresh permutation every epoch, yielding 15 models over 100 epochs. Correct-label models reached 89.39% mean true-label test accuracy. Topological features did not provide incremental predictive improvement over conventional diagnostics for early condition classification.
In briefCJRR-Net: Candidate-joint reprojection ranking for choosing camera views in human pose estimation
SummaryThis paper proposes CJRR-Net, a network that selects the most informative future camera view for human pose estimation in a multi-camera system. It combines current-view pose observations, camera configuration, and geometric features obtained by reprojecting the estimated 3D pose and its uncertainty into each candidate view. On eight calibrated cameras in the MPI-INF-3DHP dataset, the selected view achieves a 2D mean per-joint position error of 105.18 pixels, lower than the shared-input MLP and adapted PoseErrNet baselines.
Huiting Xi, Marsha Mariya Kappan, Eduardo Benitez Sandoval et al.
In briefLearning decisions directly for battery-assisted grid-import smoothing at AI data centers
SummaryThis paper addresses how rapid power fluctuations from AI data centers challenge grid integration. It proposes a risk-aware battery energy storage scheduling method that coordinates advance reserve commitments with demand-responsive dispatch, accounting for operating costs, grid-ramp penalties, and their conditional value-at-risk. The authors introduce battery scheduling decision-focused learning (BS-DFL), which trains a constraint-aware neural optimization layer directly from the objective without optimal-solution labels, then uses it to guide demand-predictor learning. In a 10,000-minute simulation with loads derived from measured AI workloads, BS-DFL reduced the mean absolute minute-to-minute grid-import change by 71.70% and total scheduling cost by 18.01% compared with an MSE-trained predictor paired with a mixed-integer linear programming solver, with an average decision time of 18.77 ms.
In briefWalrus: A Unified Speech and Audio Foundation Model Technical Report
SummaryThis report presents Walrus, an 8B-parameter foundation model that handles speech recognition, speech translation, text-to-speech, emotion recognition, audio captioning, and general audio understanding. Its main contribution is a data-centric training recipe built on a standard decoder-only backbone initialized from Gemma 4, with a three-stage data curriculum and task- and dataset-aware weighting. Trained on over 600M samples, it leads on most multilingual, multi-talker, and code-switching benchmarks. On most benchmarks it matches or surpasses Qwen3-Omni, which is four times larger.
In briefAURA: Auditory Spatial Intelligence Using a Unified Representation from a General-Audio Encoder
SummaryThis paper asks whether a general-audio codec trained without spatial supervision can both identify and locate sounds in multi-source scenes. The authors propose AURA, which keeps the codec and language model frozen and uses complete alignment to answer spatial questions. On SpatialSoundQA, AURA exceeds the strongest dedicated system in dual-source direction accuracy with about 40% of its trainable parameters.
In brief"You Are What You Read": A Hindi-Language Dataset for Detecting Bias in Indian Political News
SummaryThis paper introduces a dataset for studying media bias in Hindi, one of India's most widely spoken languages. It contains 448 political news articles from four leading Indian news agencies, annotated by 14 experts for bias and classified into selection, framing, clickbait, and omission bias. Annotators also wrote explanations for biased articles to support explainable approaches. Inter-annotator agreement was strong, with a Cohen's Kappa of 0.81 for bias versus no-bias and an average Kappa of 0.80 for the multi-label task.
Kushal Trivedi, Mayank Mahor, Dhruv Sanghvi et al.
In briefEgo Guidance: Using Action-Free Human Videos to Improve Robot Learning
SummaryThis paper studies whether human egocentric videos without action labels can help train robot policies. Building on a pretrained vision-language model, the authors propose SPAR, a Mixture-of-Transformers architecture in which future visual state prediction complements action prediction. Representations learned from egocentric videos improve performance on out-of-distribution robot tasks, with gains increasing as video data grows.
In briefBootstrapped On-Policy Distillation: Train an Intermediate Model with Reference Guidance, Then Distill to a Fresh Model
SummaryThis paper studies why on-policy self-distillation (OPSD), which uses a reference-conditioned copy of a language model as its teacher, can degrade under continued training. The authors attribute this to a growing mismatch between the evolving reference-free student and the frozen privileged teacher. They propose Bootstrapped On-Policy Distillation (BOPD), a two-stage method that first uses OPSD to transfer reference-based task knowledge into an intermediate model, then distills that model into a fresh copy of the initial model. The paper reports gains in mathematical reasoning and code generation over both the initial models and the OPSD teachers.
In briefOpticalRAG: Token-efficient retrieval-augmented generation using pixel-space compression
SummaryThis paper presents OpticalRAG, which renders documents as page images, encodes them with a frozen vision encoder, and compresses retrieved visual tokens into as few as 16 tokens before passing them to an LLM. It uses Encoder-Bridged Retrieval and Query-Driven Distillation to work without expensive cross-modal training. On LongBench-v2, BAMBOO, and LooGLE-v2, it achieves the best compression-aware accuracy while injecting only 16 tokens.
In briefA computational framework for understanding embodied plans: GenEA, a model that infers people's goals and multi-step plans from their actions
SummaryThis paper introduces GenEA, an agent representation for reasoning about multi-step, long-horizon embodied plans. It extends task-and-motion planning with simulation-based intuitive physics, coupling an articulated body's coarse biomechanics with its interactions with nearby objects. Embedding this state-space in a generative model yields a joint distribution over goals and embodied plans. In object manipulation and navigation experiments, GenEA recapitulates human reasoning about embodied plans and outperforms heuristics, standard task-and-motion planning, and a state-of-the-art vision-language model.
WhyThe authors build a generative model over goals and embodied plans, comparing it to heuristics, task-and-motion planning, and a vision-language model in manipulation and navigation experiments.
Aalap D. Shah, Basil Saeed, Joshua B. Tenenbaum et al.
In briefBenchmarking LLM Next-Step Recommendations Using Knowledge Space Theory
SummaryThis paper presents a benchmarking method that uses Knowledge Space Theory as a formal reference for evaluating large language models on next-step learning recommendations. Given a task catalogue and prerequisite relations, it derives admissible knowledge states and their immediate next steps, then compares model outputs against these reference sets. In a sample study of 2,400 responses from six models across four mathematics domains, exact-set accuracy ranged from 61.3% with descriptions only to 98.2% with the complete knowledge space. The benchmark measures structural recommendation accuracy under the tested conditions, not learning benefit.
In briefSepsisFlow: An Event-Stream Foundation Model for Continuous ICU Sepsis Risk Warning
SummaryThis paper presents SepsisFlow, an event-stream foundation model for sepsis in intensive care units that updates hourly. It predicts pre-onset risk, defines onset contexts from patient states, and models 72-hour post-onset organ-support and physiological trajectories. Evaluated on MIMIC-IV and eICU, it reportedly detected more cases than NEWS2, qSOFA, and SIRS at equal alarm volumes and showed mortality differences across onset contexts.
In briefEgoHMP: Precise 3D Human Motion Prediction Using First-Person Cues
SummaryThis paper proposes EgoHMP, a framework that uses egocentric views and head motion to infer a person's interaction intent for predicting human motion in 3D scenes. It combines a vision-motion alignment encoder, an intent predictor that suppresses invalid interaction cues, and a contact-aware motion decoder with a global trajectory self-prompt. The authors report that EgoHMP achieves state-of-the-art performance.
In briefDrugCoding: A scientific programming agent that automates drug discovery computing tasks
SummaryThis paper presents DrugCoding, a scientific programming agent system for drug discovery that plans tasks, adaptively routes to specialized models, and generates and executes code. It uses multi-agent collaboration to carry out multi-stage scientific computations. In multi-turn tool-calling experiments, it achieved an average session-level Pass@1 of 97.35% on the full benchmark, outperforming the strongest baseline, Codex, by 9.24%. On the challenging subset it reached 97.92%, also exceeding the strongest baseline.
In briefPredicting single-cell gene activity from routine H&E tissue slide images
SummaryThis paper studies how to predict gene expression from ordinary hematoxylin and eosin (H&E) stained tissue images to obtain spatial molecular information. The authors adapt pathology foundation models to single-cell prediction and evaluate them on paired H&E and Xenium 5K data. Their approach raises the mean gene-wise Pearson correlation by 2.67 times over the standard zero-shot baseline and reaches state-of-the-art performance on this task.
In briefCP-MLPs: A Tensor-Rank Theory of Tied and Untied MLP Blocks
SummaryThis paper introduces CP-MLPs, a tensor framework that treats each multiplicative unit as a rank-1 interaction with detector, value, and output roles. It proves exact tensor-rank separations between tied gated blocks such as ReLU² and untied gated blocks such as ReGLU and SwiGLU, showing the width advantage of untying depends on the target structure. Experiments indicate that loss plateaus alone do not establish an approximation gap, since arbitrarily accurate approximation is governed by border rank.
WhyThe authors use CP tensor rank to show exact separations between tied and untied gated MLP blocks, with width advantages depending on target structure.
Md Rifat Arefin, Farzaneh Heidari, Irina Rish et al.
In briefMouldIt: Data-efficient multi-task time-series representation learning for injection moulding
SummaryThis paper examines representation learning for modelling the injection moulding process, comparing its data efficiency and multi-modal benefits with classical machine learning methods. The authors synthesise a process-informed dataset based on a real reference cycle of an Erlanger Träger part. They introduce MouldIt, a temporal convolution and attention based variational autoencoder that reconstructs in-mould signals while classifying parts as acceptable or as synthetic failure modes. MouldIt outperforms classical methods from 4000 cycles in the single-modal setting and from 2000 cycles in the multi-modal setting, with part sorting losses of 6.0 ± 0.7% and 1.0 ± 0.3% respectively.
In briefpicoEmergencyNet: An Ultra-Small Attention-Enhanced Model for Disaster Recognition in Aerial Images
SummaryThis paper introduces picoEmergencyNet, an ultra-lightweight convolutional neural network for classifying disasters in aerial images. It combines depthwise dilated convolutions, efficient channel attention, and channel shuffling to extract multi-scale features at low computational cost. The model has 23k parameters and 7M FLOPs, and reaches up to 96.1% accuracy on the AIDER and AIDERv2 datasets. On an NVIDIA Jetson AGX Orin without hardware-specific optimization, it runs at 258 FPS on GPU and 46 FPS on CPU, with about 7 times lower activation memory than competing models.
In briefA Readable Method for Checking Who Translated Early Chinese Buddhist Texts
SummaryThis paper presents a stylometric method for attributing early Chinese Buddhist translations, using features chosen by Buddhist philologists such as function characters, sentence formulas, and part-of-speech n-grams. The authors build style profiles from 31 authenticated translations by seven translators and compare eight disputed texts with An Shigao's profile. Texts T1557 and T0101 show the strongest similarity to An Shigao's authenticated corpus, with scores of 0.783 and 0.687, and these results agree with a published BERT-based approach.
In briefDepthGS: Depth-guided Gaussian splatting for high-quality surface reconstruction.
SummaryThis paper proposes DepthGS, a depth-guided Gaussian splatting framework for 3D surface reconstruction. The authors derive theoretical relations between depth discrepancies and reconstruction error, and use them to design adaptive regularization for single-view and multi-view optimization. Experiments on DTU, Mip-NeRF360, and Tanks and Temples show competitive performance in both surface reconstruction and novel view synthesis.
In briefCurriculum-based proof repair using learned counterexamples
SummaryThis paper presents TopoVerus, a language-model framework for repairing Verus formal proofs. It generates a structured source-level counterexample, a repair rationale, and a repaired proof in a single pass rather than through iterative refinement. Training combines reinforcement learning from Verus feedback with a proof-code safety gate, and a curriculum uses a topological reward based on invariant dependencies to order invariant repairs.
In briefFlashMotion: Ultra-Fast Long-Term 3D Motion Generation
SummaryFlashMotion is a framework for generating high-quality long-term 3D human motion under strict inference budgets, covering text-to-motion, music-to-dance, and speech-to-gesture tasks. It combines a 3D motion-space MeanFlow formulation for few-step generation with an Adaptive Soft masking method for motion editing, and uses a lightweight non-autoregressive BiMamba architecture to improve per-step efficiency. The authors report state-of-the-art overall performance in both quality and efficiency.
In briefMultifidelity Operator Inference for model reduction when high-fidelity data is scarce
SummaryThis paper addresses the high cost of simulating PDE-governed models and the scarcity of high-fidelity training data for Operator Inference. It proposes Multifidelity Operator Inference (MF-OpInf), which supplements scarce high-fidelity snapshots with abundant low-fidelity data from coarser discretizations. The method estimates reduced operators using a multifidelity linear regression framework. Numerical experiments on model PDE problems show that, under the same computational budget, MF-OpInf predicts solution fields more accurately than high-fidelity-only OpInf.
In briefPUMA: A Method That Locates Image Forgeries by Combining Meaning and Noise Clues in a Learned Order
SummaryThis paper presents PUMA, a method for localizing manipulated regions in images by combining semantic structure with noise-based forensic evidence. It uses spatially constrained unbalanced optimal transport to match features across branches and adaptively chooses the order of two sequential operators. The authors report strong localization results on four test sets, with higher scores than listed representative methods on COVERAGE and NIST16.
In briefWhen to Start Denoising? Adaptive Denoising Entry for Generative Video Compression
SummaryThis paper studies diffusion-based generative video compression, where the timestep at which the compressed latent enters the diffusion process shapes the balance between fidelity and realism. The authors formulate denoising start-timestep selection as a geometric optimization problem and propose ADE-GVC, which uses a lightweight predictor to estimate an adaptive start point and a temporal adapter to inject compression-aware conditions into a pretrained video diffusion model. The authors report state-of-the-art perceptual performance and bitrate savings relative to the open-source baseline GLC-Video.
In briefSynapse-Control: a lightweight runtime controller that decides when agents get past-experience advice
SummaryThis paper presents Synapse-Control, an agent runtime with a lightweight control layer that decides when to send retrieved advice to the task-executing model. Rules shortlist relevant advice, and a 1.7B classifier scores the candidates instead of generating diagnoses with a large model. On AppWorld and WebShop, with 1,308 attempts per method, it achieves 270 successes (29.8% more than Base), replaces 4,987 generative diagnosis calls, and cuts auxiliary control time relative to Sentry by 50.2% and 51.0%.
In briefMegrez: A Knowledge-Guided Agent for Formal Mathematics over MK Set Theory
SummaryThis paper presents Megrez, an agent that guides formalization of mathematics in Rocq libraries built on Morse–Kelley (MK) set theory. It uses a knowledge graph to retrieve declarations and concepts, a short-burst proving loop with branch tracking, and a three-check pipeline for statement fidelity, axiom auditing, and full builds. Deployment and evaluation on a point-set topology corpus show effective cross-chapter formalization and scalable premise retrieval.
In briefEarth Embeddings for Environmental Remote Sensing: A Review and Evaluation
SummaryThis paper introduces the theoretical foundations of Earth embeddings, general-purpose representations derived from multi-source Earth observations, and reviews several representative products. It examines their applications across six major domains of environmental remote sensing. The authors systematically evaluate them on four downstream tasks: hyperspectral image classification, individual tree crown segmentation, crop yield prediction, and aboveground biomass mapping. They conclude that current Earth embeddings do not always outperform imagery or hand-crafted features and should be used selectively rather than as direct replacements.
In briefCorrecting OCR Errors in Mixed Cyrillic-Latin Uzbek Text with a Byte-Level Model
SummaryThis paper addresses OCR errors in Uzbek documents where Cyrillic and Latin characters that look alike are mixed on the same page. The authors propose a post-OCR correction system built on the byte-level ByT5 model, using a noise generator whose weights come from 359 manually classified error events, plus a conservative decoding rule bounded by Levenshtein distance. On a gold set of 5,204 line pairs, the character error rate drops from 0.0578 to 0.0369, 91.3% of homoglyph errors are restored, and only 84 new errors are introduced.
In briefOne Byte, One Token: Converting a 27B Subword LLM into a Byte-Level Model
SummaryThis paper examines whether an existing 27B subword language model can be converted directly into a pure byte-level model, with one byte per token and no added patching or encoder modules. The authors shrink the vocabulary to 384 entries, keep all other weights, and continue training in two stages using data generated by the original model. The byte model stays close to the original on most benchmarks and improves on character-level tasks, but loses some Chinese character lookup knowledge and runs about four times slower in Chinese output throughput.
In briefReal-time joint optimization of traffic signals and lane markings at an intersection via reinforcement learning
SummaryThis paper proposes a reinforcement learning method to jointly optimize traffic signal control and lane channelization at a single isolated intersection in real time. A deep Q-network outputs joint actions for signal phases and lane markings, with a rule-based filter that blocks abrupt, unsafe marking changes. Experimental results show the method increases intersection throughput.
In briefComputing Interactions, Not Attributes: What a Fact Layer Should Provide in the Chinese Sexagenary System
SummaryThis paper studies what a deterministic fact layer should supply when language models are used in rule-governed domains. It uses the Chinese sexagenary system underlying BaZi chart analysis as a testbed, comparing model performance closed-book, with a reference table, and with computed facts at increasing levels. Directly answering models err most where attributes must be combined, and the authors conclude the fact layer must supply computed interactions.
In briefToward Reliable VLM Judges: Stable Verdicts and Causal Sensitivity Diagnostics
SummaryVision-language models are increasingly used as automated judges for multimodal and embodied tasks, but aggregate accuracy does not show whether their verdicts change for the right reasons. The paper proposes State-Conditional Invariance, which requires a judge to stay stable under presentation changes that preserve meaning while responding to interventions that alter the task outcome. It instantiates this in GridWM-Judge, a diagnostic benchmark built from deterministic MiniGrid trajectories with Full, evidence-removal, and counterfactual variants. Evaluations across five diagnostic tasks reveal heterogeneous failure modes, motivating joint reporting of correctness, presentation stability, and causal sensitivity.
In briefIFRS Compliance Verification: A Multi-LLM Benchmark and Ensemble Classifier
SummaryThis paper addresses verifying corporate financial disclosures against IFRS, a laborious auditing task. It introduces a benchmark of 19,998 IFRS regulation-disclosure pairs and a multi-stage framework combining domain-adapted sentence embeddings with an ensemble classifier that fuses probability outputs from GPT-4.1, Claude Sonnet 4, and Gemini 2.5 Pro. On items where the LLMs disagreed, the ensemble reaches 81.7% weighted F1, outperforming the strongest single LLM, majority voting, and eleven unsupervised aggregation baselines.
In briefProving Transformers Can Learn Graph Search: Training Dynamics and Depth's Exponential Power
SummaryThis paper studies path-finding on directed graphs as a symbolic model of multi-step reasoning in transformers. It analyzes how many layers are needed to implement depth-first search (DFS), showing that two layers suffice for near-linear size. The authors also show that with the right training data, a two-layer transformer can provably learn DFS through gradient flow and generalize to unseen graphs.
In briefA lightweight deep learning tool for reconstructing 3D astrocyte shapes in normal and stressed brain tissue
SummaryThis paper presents AstroSeg-vEM, a lightweight framework for segmenting, reconstructing in 3D, and quantitatively analyzing astrocyte morphology from volume electron microscopy images. It combines transformer-based 2D segmentation with dedicated post-processing and works directly on high-resolution EM sections. Tested on four mouse hippocampal datasets covering normal and stress-associated states, it achieved average Dice and weighted Dice scores of 0.92 and 0.95 on a held-out test set. The architecture runs efficiently on consumer-grade hardware.
In briefPhyRubric: Rubric-Based Diagnostic Evaluation of Physical Common Sense in Text-to-Video Generation
SummaryThis paper introduces PhyRubric, a framework for evaluating whether text-to-video models produce videos that follow physical common sense. For each caption, it builds a rubric before viewing any video, splitting the caption into semantic and physical items and scoring each item using timestamped evidence. The authors release PhyRubric-Bench, covering 34 categories across five physical domains, and PhyRubric-Train with 3,788 videos, used to fine-tune the open-source judge PhyRubric-Judge. Experiments on ten text-to-video models show that physical common sense remains unsolved, and the judge correlates with human ratings more closely than existing open-source judges.
In briefMarket-Conditioned JEPA: Learning from Unpredictable Market Shocks
SummaryThe paper argues that much of financial time-series variation comes from market-wide shocks that an asset's own history cannot predict. It introduces Market-Conditioned JEPA (MC-JEPA), which conditions prediction on the realized market factor and builds a leave-one-out market condition from other assets' future target embeddings during training. On 11 years of minute-level Chinese A-share data, it outperforms per-asset JEPA across nine targets covering returns, exposure, risk, and liquidity, and delivers the best portfolio performance in a historical backtest.
In briefProof that pair-merge tokenisation is hard to approximate on binary strings and strongly NP-hard on unary datasets
SummaryThis paper studies how tokenisers such as BPE shorten text by repeatedly merging adjacent symbol pairs. The authors prove that, unless P = NP, no polynomial-time algorithm can achieve an approximation within a factor of about n/log n, even for a single binary string. For merges that may skip occurrences, they give a simple algorithm achieving O(n/log log n). Over a one-letter alphabet, the related tokenisation problems are strongly NP-complete on datasets with repeated strings.
WhyThe authors reduce BPE-style merge selection to grammar size, proving approximation hardness on binary strings and strong NP-completeness for unary datasets.
In briefTesting learned AC optimal power flow models shows results depend on data encoding
SummaryThe paper applies thirteen metamorphic relations to audit 17 learned AC optimal power flow surrogate configurations, covering two released checkpoints and random-weight controls. Every configuration violates at least six relations, even when reporting tolerances are changed tenfold. For GridSFM-Open v1.1 the authors identify an indexing bug and several preprocessing issues, and in a case500 experiment augmentation benefits depend on the training protocol while all predictions remain AC infeasible.
In briefStudies whether using intermediate measurements helps multi-objective Bayesian optimization in scientific design.
SummaryThis paper examines composite modeling in multi-objective Bayesian optimization, a method for finding good designs when several goals must be balanced. Composite modeling learns from intermediate responses that are combined into final objectives, and it is compared with approaches that use only black-box objective observations. The authors run controlled experiments with four multi-objective methods across eight synthetic and application-inspired scientific design benchmarks. They report that composite modeling tends to help more when the mapping from intermediate responses to final objectives is non-bijective and non-affine.
In briefInterpretable Clinical Text Prediction Without Concept Labels Using Token-Concept Attention
SummaryThis paper proposes a framework that explains language model predictions on clinical text through human-understandable concepts, without requiring annotated concept labels. A cross-attention module estimates how strongly each medical concept appears at each text token, and these scores feed a linear layer that predicts clinical outcomes. On chest trauma severity stratification and in-hospital mortality prediction using MIMIC-III, the framework's AUROC came within 0.02 of blackbox models, and it supports both cohort- and patient-level interpretation.
In briefBenchmarking vendor-guided optimizations on edge NPUs, from NPU-only to end-to-end performance
SummaryThis report proposes a benchmarking method that compares three configurations: a baseline, vendor-guided optimization applied only to the NPU (VG-NPU), and vendor-guided optimization across the full deployment pipeline (VG-E2E). The authors test four edge NPUs on three vision workloads. Their results show that the two optimization levels do not improve accuracy and latency consistently, and NPU-only rankings do not always match end-to-end rankings, so both levels should be evaluated separately.
In briefA benchmark that traces clinical reasoning in LLMs back to source evidence
SummaryThe authors built Med Evidence Bench (MEB), with 1,536 physician-written items across guideline knowledge, standardized clinical scenarios, and real-world laboratory scenarios, where each medical claim links to a versioned source-text span. Ten large language models scored high on guideline knowledge but markedly lower on real-world scenarios, with rankings varying by level. Automated grading showed strong agreement with the majority vote of three physicians, though agreement was lower for evidence application.
Zhiqun Zhao (1), Xiaodong Yang (2), Bruce Gu (3 et al.
In briefUsing LLMs to Discover Structurally Sound, Interpretable Cognitive Models
SummaryThis paper presents DisCo, a framework that uses large language model guided evolutionary search to automatically discover interpretable symbolic cognitive models. It includes a Validator that filters structurally invalid models and a Cleaner that removes redundant parameters to improve parameter identifiability. Across three behavioral tasks, DisCo finds models with performance competitive with task-specific baselines, and ablations show the Validator is important for recovering psychologically meaningful structure.
In briefDo Dense Rewards Justify Their Token Cost in GRPO Training?
SummaryThis paper examines whether dense process rewards, which score intermediate reasoning steps, improve reinforcement learning for language model reasoning. The authors propose a matched-compute protocol that equalizes generated tokens and training GPU-hours across 1,668 runs on six models. Under matched compute, a Monte-Carlo dense reward underperforms standard GRPO, while training on prompts with predicted success near one half yields modest accuracy gains.
Vikash Singh, Weicong Chen, Debargha Ganguly et al.
In briefDarmok: A system that solves British cryptic crosswords by cross-checking clues and using backup methods
SummaryThis paper presents Darmok, an automated system that solved British cryptic crosswords at a superhuman level. At the 55th Times Crossword Championship in London, it solved all seven puzzles without errors and finished faster than every human competitor in each round. The system cross-checks clues across three independent models, uses confirmed crossing letters to unlock harder clues, and fills remaining gaps with a whole-puzzle model and dictionary lookups. The authors also show that backup paths were essential for reliability when the primary model slowed down or failed.
Siddhartha Chatterjee (The University of Texas at Austin)
In briefHow post-training changes verbalized algorithmic computation in language models
SummaryThis paper studies how SFT, DPO, and RLVR post-training affect language models' use of internal information and written chain-of-thought (CoT), using Olmo-3-7B-Instruct on multi-digit addition with carry values as intermediate state. Probes recover carry information from internal activations with similar accuracy for SFT and RLVR, but manipulating that information does not reliably control outputs. CoTs produced by RLVR give higher accuracy across checkpoints than those from SFT or DPO, and carry-related statements receive higher importance scores than random controls after RLVR, while DPO shows similar sensitivity without comparable accuracy gains.
In briefMassive Activations Distort Sparse Autoencoder Evaluation (Paper)
SummaryThis paper shows that massive activations in some models concentrate most activation variance on the first principal component, which distorts standard sparse autoencoder evaluations based on variance explained and ablation damage. The authors propose removing the first principal component before training the autoencoder and adding it back to the reconstruction. In affected models, top-50 feature sharing across runs rises from 4-39.5% to 73-99.5%, and at the same TopK sparsity, model degradation falls by up to 44% and content reconstruction error by up to 33.9%.
In briefUsing stochastic rounding to close the BF16–FP16 convergence speed gap in RL post-training
SummaryBF16 is widely used for pretraining large language models, but in reinforcement learning with verifiable rewards at low learning rates, it converges more slowly than FP16. The authors attribute this to weight sticking, where small updates to FP32 shadow weights are lost when rounded to BF16. Applying stochastic rounding matches FP16's convergence speed while yielding lower KL divergence than BF16 training with a larger learning rate, and it keeps BF16's dynamic range without requiring loss scaling.
Mahdi Nazemi, Oleg Rybakov, Tijmen Blankevoort et al.
In briefSanity checks for removing attention components: null models and controls
SummaryThis paper proposes three reusable checks for architectural changes to Transformers: whether a measured statistic exceeds its noise floor, whether it survives a geometric null model, and whether the training gain depends on the specific direction removed. The authors apply these to self-attention removal, which deletes the part of each attention head's output aligned with its own value vector. Across nine pretrained multi-head models, a causal earlier-position null accounts for 48% to 63% of this alignment. In a pre-registered factorial experiment, self-directed removal lowers validation loss and outperforms a fixed arbitrary direction.
In briefAgent benchmarks lose their ability to rank models before running out of tasks
SummaryThe paper argues that saturation of data-science and ML-engineering agent benchmarks largely stems from how results are measured, since most reduce each run to a pass/fail bit. By re-scoring released runs of eight benchmarks, the authors show this binary grading is too coarse and capped at a reference human. They propose a clipped log-ratio score with dated anchors and cost as a separate axis.
In briefFrom Long Traffic Videos to Verifiable Conflict Evidence (AgenticConflictLens)
SummaryThis paper presents AgenticConflictLens, an agentic framework that extracts traffic conflicts from long fixed-camera videos based on natural-language safety queries. It uses a Traffic-Aware Controller with structured traffic memory and conflict-analysis tools to decide where to search, which road users to examine, and which evidence to verify or reject. On 96 annotated hours from Bellevue and NYC DOT, the method achieves the highest conflict-existence F1 and temporal IoU on both datasets while inspecting only about 8.7% and 11.9% of the video.
Boshra Khalili, Sonia Moshfeghi, Jinwoo Jang et al.
In briefEfficient A/B testing for recommender systems using predictable scaling patterns
SummaryFull-scale A/B experiments for recommendation systems are computationally expensive, and small-scale results can be misleading. The paper finds that recommendation models' performance changes follow a predictable scaling law even when different feature changes are applied. It proposes a method that uses this law and statistical inference to predict full-scale effects with confidence intervals. Tested on Wukong, HSTU, and LLM-based recommenders across more than 40 changes, the method distinguishes beneficial from neutral or regressive changes using only a fraction of the compute.
In briefAdaptive catalyst discovery via composable AI for catalysis
SummaryThe paper presents Composable Catalytic Intelligence (CatAI), a modular AI framework that learns 1.8 million structure-property relations and organizes them into 30 physicochemical encoders reusable by task-specific models. On benchmark tasks it achieves a 7.8% lower task-averaged prediction error than the next-best models. Experimentally, it identified a new confined-passable pocket catalyst for industrial homogeneous polymer synthesis with 9 times the activity of a commercial benchmark, and guided closed-loop optimization of a Cu-based geminal-atom photocatalyst for C-O coupling that outperformed traditional Ni-based systems.
WhyThe authors propose a modular catalyst AI framework with reusable physicochemical encoders, reporting benchmark gains and two experimental validations in catalyst discovery.
In briefBeyond the Log-Linear Frontier: Finite-Budget Neural Scaling Laws
SummaryClassical neural scaling laws model error as a separable power law in model size and dataset size. This paper studies how spectral and domain heterogeneity cause learning curves and compute-optimal allocations to deviate from this log-linear frontier at finite budgets. It proposes estimators, including a closed-form initializer and a water-filling approach to domain mixing. The authors report reduced holdout error and allocation regret across multiple experimental settings.
In briefHeadline Framing Sensitivity in LLMs: Evidence from Chinese News
SummaryThis paper examines whether large language models judge news headlines by the reported fact or by how it is framed. The authors audit 63,926 Chinese news headlines and run a controlled experiment on 30 real events, 30 synthetic events, and 29 natural headline pairs across four Chinese models. Changing only the headline frame shifted perceived misleadingness and event severity in the same direction across all models. Supplying the article body reduced this severity bias, and a small human reference study showed the same direction of effect.
In briefUniWeaver: Using Unified Multimodal Models to Generate Interleaved Text-Image Content
SummaryUnified multimodal models can generate rich mixed text-image content, but they lack native support for coordinating the two and need explicit planning to turn rough requests into coherent works. The paper proposes UniWeaver, a model-agnostic framework that combines multimodal planning, generation, and visual feedback. The authors also build UniWeaver-10K, a dataset of 10,000 tasks across 20 task families, and UniWeaver-Bench, a benchmark for evaluating interleaved outputs. Experiments report state-of-the-art performance across multiple model backbones.
In briefGEMOR-GS: A 3D Gaussian Splatting SLAM Method for Accurate, Globally Refined, and Re-enterable Robot Maps
SummaryGEMOR-GS is a LiDAR-inertial-visual SLAM method that builds a 3D Gaussian splatting map intended to be geometrically accurate, globally refined, and re-enterable for robots. It first estimates the LiDAR-camera extrinsic calibration offline, then alternately refines keyframe poses and Gaussian parameters, and rebuilds the global reconstruction from the refined keyframes. After tracking loss, it recovers pose from the 3DGS map using an XFeat-based map, RANSAC-PnP, and Gaussian RGB-D warp refinement. Experiments report that the offline reconstruction outperforms the online map by 14.3 to 19.9 dB and improves PSNR by about 9 dB over Gaussian-LIC2, with 86 to 100% relocalization success on held-out data.
In briefEJAI: A Civic AI System That Preserves Judgment on Data Center Environmental Justice
SummaryThis paper presents EJAI, a system that helps residents examine environmental justice issues raised by data centers. It combines an interactive map, retrieved evidence, statements about missing data, and two stakeholder personas so users can inspect the evidence and form their own policy recommendations. In a small functional benchmark, retrieved evidence supported 98.4% of factual claims. A within-subjects study with 15 participants preparing for a simulated town hall found that some users made conditional recommendations, while others judged the personas to be simplified or biased.
Evan Pa\v{c}es-Wiles, Bryen Alvarez, Sara Gulabani et al.
In briefThe RSI Tax: When Self-Improvement Costs Outweigh Gains
SummaryThis paper examines whether recursive self-improvement (RSI), where an AI agent improves itself using its own task experience, is worth its expense in continual task streams. It defines the "RSI Tax" as the token cost of the self-improvement process and evaluates it alongside task-execution costs. Across six methods and two language models, a bounded-note baseline matched or exceeded the accuracy of three RSI methods while saving 65.6-71.0% of total tokens. The authors find that verification makes up most of the tax, and that much of the associated cost goes to states or artifacts never used later.
In briefParallel Broyden methods for faster evaluation of nonlinear state space models
SummaryNonlinear state space models are usually evaluated one step at a time, which limits use of parallel hardware. This paper proposes a quasi-Newton method based on Broyden's method that approximates Jacobian blocks with diagonal-plus-low-rank terms built from trajectory secants, avoiding automatic differentiation and remaining compatible with parallel scan through rank compression. Experiments on low-rank RNNs, parallel MCMC algorithms, and denoising diffusion models show fewer iterations than other quasi-Newton methods, and it is the fastest parallel method when Jacobian evaluation is expensive or state dimensions are large.
In briefOperational model for a trusted data space for vocational education
SummaryData sharing in vocational education has long been limited by unclear ownership, weak trust, and insufficient incentives. Building on an Internet of Data based trusted data space for vocational education, this paper develops a multi-stakeholder operational model. The model covers a value network, a product portfolio, revenue allocation, and governance rules. It emphasizes data sovereignty protection and incentive compatibility, offering a reference for sustainable regional vocational education data ecosystems.
In briefDifferentiable Filtering for Inertial Odometry: a learning method that embeds Kalman filter updates into training to reduce drift in inertial positioning
SummaryInertial odometry is important for autonomous systems but suffers unbounded drift from integrated sensor noise and biases. The paper proposes DFIO, which embeds the update stages of an error-state Kalman filter into a computational graph and, after pre-training, jointly optimizes displacement and covariance outputs using posterior-position supervision through differentiable measurement updates. On a public pedestrian dataset, DFIO reduces mean absolute and one-second relative translation errors by 13.2% and 39.6% versus the TLIO baseline. Tests on noisy uncalibrated smartphone data indicate reduced translational drift and maintained trajectory alignment.
In briefDetecting Tiny Lesions: A Deformable Mamba Model for Retinal Segmentation
SummaryThis paper proposes D²Mamba, a model for segmenting diabetic retinopathy lesions that vary in shape and are often only a few pixels in size. It combines non-causal state space duality with deformable Mamba scanning to capture global context and adapt to lesion shapes. Evaluated with five-fold cross-validation on the DDR dataset, it achieved 33.87% mIoU, 2.99 points above the strongest baseline, with microaneurysm Dice improving by 10.06 points.
In briefGlaucoma detection with conditional-entropy evidential deep learning
SummaryThis paper proposes a conditional entropy function to quantify predictive uncertainty in Evidential Deep Learning for glaucoma detection from retinal fundus images. Validated on the EyePACS-AIROGS-light-V2 and RIM-ONE-DL datasets, the method produced uncertainty estimates with substantially lower calibration errors than Deep Ensembles, MC Dropout, post-hoc calibration, and conventional Dempster-based EDL. The authors also analyze Grad-CAM heatmaps with a focus index, finding that false negatives showed activations displaced from the optic disc.
In briefNepali-English code-mixed speech recognition: data preparation and adapting a speech foundation model
SummaryThis paper adapts SraVaani, a 449M-parameter hybrid TDT-CTC speech model, to recognise speech that mixes Nepali and English, using a 172.8-hour corpus across six Nepali speech domains. Because the pretrained tokenizer lacks digits and several symbols, the authors generate Nepali and English readings for numbers and use zero-shot ASR output to choose a plausible form. Full fine-tuning lowers test word error rate from 48.6% to 17.1%, and a 12-epoch run reaches 16.3%, slightly better than fine-tuned Whisper large-v3-turbo at 16.4%.
Jeevan Neupane, Janardan Bhetwal, Kshitiz Paudel et al.
In briefTMAL: A method for segmenting thin, curved structures while preserving their connectivity
SummaryThis paper proposes TMAL, a method for segmenting thin, curved structures such as retinal vessels and surface cracks while keeping both pixel accuracy and structural continuity. It uses soft affinities that emphasize thin regions near boundaries and predicts directional affinities at several offsets, combined with weights from decoder features. The authors report that TMAL achieves the highest Dice, IoU, and clDice scores on the DRIVE and DeepCrack datasets and greatly reduces the connected-component error Δβ0.
