This study evaluates automatic speech recognition (ASR) for adolescent health communication in Twi, Dagbani, and Ewe by benchmarking five ASR systems on a Bible corpus and a domain-specific ASRH dataset, then performing supervised domain adaptation with a fine‑tuned Qwen3-ASR-0.6B model. Fine‑tuning significantly lowered word and character error rates, especially for Ewe, and the adapted model was deployed in the KasaHealth voice‑first application, which received high user approval and highlighted remaining domain gaps. The work demonstrates that in‑domain data, rather than model size or computational resources, is the primary limitation for effective ASR in these languages.
By Stephen E. Moore, Akwasi Asare, Mich-Seth Owusu, Paul Azunre, Joel Budu, Lawrence A. Adu-Gyamfi
The paper introduces a three-level evaluation framework—behavioral deployment, LM-head readout, and probe recoverability—to distinguish whether a language model fails a syntactic test by not encoding structure or by failing to use it. Using a trilingual control-dependency benchmark, the authors find that probe recoverability consistently exceeds LM-head readout, which in turn exceeds behavioral deployment across seven models and three languages, with the largest gap observed in Qwen3-0.6B Instruct. Layer-localized activation patching shows that instruction tuning shifts the decoded layer later, suggesting decoding favors surface shortcuts and that behavioral evaluation understates what models encode while probing alone overstates what they deploy.
By Zhenyan Lu, He Wang, Xiaohui Huang
The paper proposes a Diagnostic Evidence Network (DENet) that augments traditional fault‑diagnosis outputs with structured evidence, including classification, predicted characteristic frequency, and temporal localization of impulses. This evidence aligns with theoretical bearing physics and can be validated at inference time, achieving high AUROC for misclassification detection without sacrificing accuracy. A QLoRA‑adapted language model then translates DENet’s evidence into maintenance reports, markedly reducing unsupported claims.
By Yuntong Chen, Jianyu Liu, Yingqi Li, Guobin Zhao, Ziang Wang, Chao Chen, Xitian Tian, Lijiang Huang
The study evaluates large language model (LLM) graders on two computer‑science exams, testing 171 configurations of closed‑ and open‑weights models. While the best LLM configuration achieved a mean absolute error of 1.64/35—better than the 2.61/35 error between two human graders—its performance was highly sensitive to the prompt. A short "strict grader" preamble caused most open‑weight models to exceed acceptable error thresholds or stop grading entirely, whereas fine‑tuning with a single LoRA adapter restored parity with human graders and reduced sensitivity to harsh prompts.
By Ali Habibullah, Yazan Alshoibi, Mohammad Alshiekh, Salman Khan, Naeemullah Khan
S2Planner is a trajectory planner for autonomous driving that fuses data from three front-facing cameras, ego‑motion history, and the current driving command. It uses a fine‑tuned DINOv3 backbone with a Spatial Tuning Adapter to generate multi‑scale image features, which are refined by a coarse‑to‑fine decoder employing trajectory self‑attention and camera‑projected cross‑attention. The key contribution lies in integrating ego‑conditioned trajectory initialization with iterative, geometry‑guided sampling of multi‑scale image features, rather than introducing a new visual backbone or attention operator.
By Zhaowei Lu, Liguo Zhou, Yujie Guo, Lei Yu, Alois Knoll
The paper demonstrates that Large Language Models, despite their non‑linear components, exhibit a fundamental linearity property: when inputs from two distinct text streams are linearly combined, the model outputs a superposition of the individual next‑token distributions. This "Superposition Linearity Hypothesis" appears to be an intrinsic feature of the Transformer architecture, tends to weaken during pretraining, but can be largely restored with lightweight fine‑tuning. The authors also present a guided decoding method that separates the superposed outputs, allowing two coherent continuations to be generated from a single forward pass.
By Pavel Tikhonov, Anton Korznikov, Matvey Mikhalchuk, Nikita Dragunov, Temurbek Rahmatullaev, Polina Druzhinina, Anton Razzhigaev, Ivan Oseledets, Elena Tutubalina
World Action Agent (WAA) is a multi‑agent framework that lets vision‑language models (VLMs) directly pilot robots by operating within a visual action workspace. The workspace provides automatically selected contact views, editable action rehearsals, and in‑view correction to refine decisions before low‑level execution. WAA learns procedural skills from expert videos and human teaching, and its interaction traces can train smaller VLMs, achieving state‑of‑the‑art success on LIBERO‑Pro and improving out‑of‑domain performance on robosuite and Qwen3.5‑9B.
By Yehang Zhang, Haojian Huang, Yifan Chang, Jianchong Su, Bohan Zhou, Yingjie Xu, Wosong Chen, Tianhao Zhou, Chenxu Wang, Tianyi Zhang, Yangkai Wei, Wenqian Li, Shiyuan Deng, Yinchuan Li, Ying-Cong Chen, Zexi Li
The paper presents a retrieval‑augmented generation pipeline for answering regulatory compliance questions in finance. It builds a three‑stage retriever on LegalBERT and a compact 2B–12B generator served with 4‑bit quantization, achieving a Recall@10 of 0.774 on the ObliQA benchmark and improving answer quality via RAFT‑LoRA fine‑tuning. However, the adapted models fail to transfer to Australian case‑law questions, and a closed‑book model performs almost as well while lacking verifiable grounding.
By Tobias Deu{\ss}er, Abhishek Pillai, Aurelio F. Bariviera, Dhananjay Bhardwaj, Lorenz Sparrenberg, David Berghaus, Christian Bauckhage, Rafet Sifa
AV‑GRPO introduces a modality‑anchored diffusion reinforcement learning framework for joint audio‑video generation, addressing limitations in fidelity, text‑modality alignment, and cross‑modal synchronization. It decouples learning signals through modality‑anchored rollouts, employs trajectory‑locked frozen‑tower optimization to reduce computational cost, and adapts objectives to each modality’s dynamics. The accompanying 5DAV dataset provides difficulty‑controllable, decoupled training samples, and experiments on JavisBench and VABench show AV‑GRPO surpasses LTX‑2.3 in generation quality, semantic alignment, and synchronization.
By Zhiyu Xu, Weilong Yan, Yufei Shi, Shiyang Li, Yihao Liu, Kin-Man Lam, Yuewen Cao
The paper presents Baszta, a Polish multi‑label content‑safety classifier trained by fine‑tuning the 124M‑parameter allegro/herbert‑base‑cased model on five categories (hate, vulgarity, sexual content, crime, self‑harm) using a Focal + R‑Drop objective. In out‑of‑distribution evaluation on the Gadzi Język benchmark, Baszta achieves a small but statistically significant improvement in micro‑F1 over the Bielik Guard system, though the macro‑F1 advantage disappears when both models are properly tuned. The study also explores calibration techniques, showing that per‑category temperature scaling can recover performance lost by Platt scaling or isotonic regression, and discusses the trade‑offs between robust calibration and adversarial recall.
By Adam G\'orski, Mateusz J\k{a}kalak, Rafa{\l} Jakubowski
The paper introduces ρ_p-LoRA, a rank-allocation strategy for low-rank adaptation (LoRA) in large language models that uses ρ_p regularization (0 < p < 1) to induce sparsity in rank-one components. By regularizing the energy of each component, redundant parts are encouraged to vanish while important ones are retained, leading to an implicit thresholding criterion derived from a two-dimensional proximal subproblem. Experiments on natural language understanding and question-answering tasks show that ρ_p-LoRA achieves performance comparable to existing LoRA baselines.
By Zebang Xie, Chuanyang Zheng, Yik-Chung Wu, Yihang Gao
The paper introduces FROST, an online framework that filters synthetic training data by estimating its utility through gradient feedback anchored in real data. FROST calibrates batch utility against recent history to decide when to filter, removing 20–30% of synthetic samples while improving performance on image classification and LLM fine-tuning tasks. The method is also applied to a large‑scale industrial ads re‑ranking system, yielding significant gains over an optimized production baseline.
By Yanran Wu, Sana Lakdawala, Renzo Tassara Miller, Chongyang Bai, Sharath Ciddu, Shivendra Pratap Singh, Kungang Li, Sandeep Pandey, Chunwei Liu
The paper investigates task collapse—a failure mode where online RL fine‑tuning of a pretrained flow‑matching vision‑language‑action policy erodes performance on individual tasks—using a 450M‑parameter SmolVLA policy on LIBERO‑10. Three exploration‑noise strategies are compared: a fixed noise scale, a learned noise network, and an uncertainty‑gated controller that reallocates exploration based on novelty and competence signals without task labels. The uncertainty‑gated controller prevents task collapse across all tested seeds, whereas the other two approaches consistently cause collapse, demonstrating its effectiveness in preserving task performance during fine‑tuning.
By Mehmet Turan Yard{\i}mc{\i}, Yunus Emre \c{C}o\u{g}urcu
The study investigates how phenotypic changes can be encoded as low‑dimensional modulations of a self‑organizing developmental system, using neural cellular automata (NCAs) as a model. By applying low‑rank adaptation (LoRA) to pretrained NCAs, the authors show that simple rank‑one adjustments can control horizontal and vertical scaling of a 2D emoji phenotype, and that these adaptations generalize across diverse phenotypes sharing the same regulatory scaffold. Analysis of thousands of phenotype‑specific NCA adapters reveals latent low‑dimensional directions that govern scaling, style, and symmetrical fission, offering a computational realization of D’Arcy Thompson’s grid transformations in a minimal cybernetic tissue.
By Benedikt Hartl, Milton L. Montero, Marcello Barylli, Sebastian Risi, Michael Levin
The paper introduces SCOPE, a method for reconstructing high‑resolution digital elevation models (DEMs) from coarser‑resolution training pairs. SCOPE learns a continuous terrain representation by predicting a latent coefficient field on a low‑resolution grid and reusing local Fourier residual functions, thereby decoupling coefficient prediction from output‑grid construction. Experiments on diverse land–ocean datasets show that SCOPE outperforms competing methods across six metrics, reduces reconstruction error by about 12 % at three‑times‑unseen scale, and achieves these gains with only a modest increase in computational cost.
By Zekai Shi, Meng Zhang, Haokun Zhang, Bo Zhang
BanglaTurn is a new corpus of 35,374 Bangla podcast speech samples, each 3 to 15 seconds long, labeled for end‑of‑turn detection through speaker diarization, an LLM pass, and human verification. A Whisper‑based model with task‑specific classification heads achieves 84.33 % accuracy on a balanced test set, outperforming the Smart‑Turn v3 baseline (69.28 %) and reducing the false‑negative rate from 51.57 % to 7.55 %, though with a higher false‑positive rate. The study also details the contributions of encoder‑layer fine‑tuning, multi‑scale pooling, INT8 quantization, and reports inference latency of 165–191 ms on CPU.
By Mizbaul Haque Maruf
The paper investigates how to close the quality gap in low‑resource text‑to‑speech for Khmer and Korean using the VoxCPM2 model. By training a single low‑rank adaptation (LoRA) adapter on a shared 25.5‑hour corpus, the authors improve Khmer’s mean opinion score from 3.85 to 4.23 with a rank‑64 adapter, while Korean shows no significant gain. The study highlights that adaptation benefits mainly when the base model is weak and that training loss does not always align with human ratings.
By Phannet Pov, Hyun Woo Park, Voneat Pen, Sovandara Chhoun, Wan-Sup Cho, Saksonita Khoeurn
MEVL-STP introduces a two‑stage pipeline for spotting arbitrarily shaped scene text. The detection stage fuses features from six frozen vision encoders via a hierarchical Feature Pyramid Network and a Progressive Scale Expansion network to produce precise polygon masks. The recognition stage then crops these masks and feeds them to a fine‑tuned Qwen3‑VL‑8B‑Instruct model, achieving state‑of‑the‑art detection and end‑to‑end performance on CTW1500, Total‑Text, and ICDAR 2015 without synthetic pretraining.
By Aman Anand, Partha Pratim Roy, Shivakumara Palaiahnakote
The paper introduces an observation‑Gram matrix that captures how each Gaussian in a 3D Gaussian Splatting model is viewed from training camera directions. This matrix serves as a distortion metric, enabling closed‑form degree reduction, Lagrangian rate‑distortion degree allocation, and matrix‑weighted vector quantisation. When applied to the Compressed3D framework, the metric improves PSNR by 0.49 dB before fine‑tuning and still yields a 0.32 dB gain at matched bitrate without any training images, while a training‑free stack built on the metric is 15% smaller than the image‑free GSICO at equal quality on Mip‑NeRF 360.
By Krzysztof Pietroszek
The paper presents a hybrid CNN–state‑space–attention backbone designed for 12‑lead ECG classification, combining early waveform tokenization, mixed temporal dynamics modeling, and late global attention. It introduces an ECG‑oriented Joint‑Embedding Predictive Pretraining (JEPA) that samples span masks at latent resolution and predicts clean latent targets via a momentum encoder, avoiding waveform reconstruction. Experiments on CPSC2018, Chapman‑Shaoxing, and PTB‑XL, with pretraining on ~350K unlabeled CODE‑15 recordings, demonstrate strong supervised baselines and improved transfer, especially in low‑label scenarios and with LoRA adaptation.
By Yakoub Bazi, Sarah Aljuhani, Mohamad M. Al Rahhal, Mansour Zuair, Naif Alajlan