The paper shows that post‑training quantization can introduce backdoors in large language models that are not detected by source‑precision checks. By formalizing the validation‑deployment gap with Quantization Behavioral Equivalence Classes (QBECs), the authors demonstrate that models can pass full‑precision tests yet exhibit malicious behavior after INT8 or 4‑bit compression. Experiments on machine translation and political stance classification reveal significant corruption and ideological shifts, and cross‑quantizer analysis indicates that attack persistence depends on the quantization scheme and architecture rather than just bit‑width.
By Jacopo Dardini, Claudio Stanzione, Giordano Col\`o, Giuseppe Fenza
The paper introduces a high‑resolution Rectified Flow Matching model trained on 256×256 chest CT images to serve as a generative prior for CT reconstruction. A two‑stage training strategy—initial strong anatomically informed augmentation followed by fine‑tuning—helps mitigate overfitting and improve structural fidelity. When evaluated on various CT inverse problems, Flow Matching‑based reconstruction methods outperform diffusion‑based algorithms in PSNR, SSIM, and perceptual quality while requiring fewer sampling steps.
By Davide Evangelista
The paper introduces Activation Boundary Matching for Low‑Rank Adaptation (ABM‑LoRA), a task‑informed initialization strategy that uses the signs of layer‑wise pre‑activations from a brief probe adapter as targets for a fresh adapter. By training with a margin‑based hinge objective on these activation boundaries, ABM‑LoRA captures useful adaptation directions that standard LoRA initializers miss, while requiring only a few forward passes. Experiments show that ABM‑LoRA outperforms or matches existing LoRA, SVD, and gradient‑based initializers across multiple models and benchmarks, including T5‑base/GLUE, ConvNeXt‑T, Swin‑T, Qwen2.5‑1.5B, and LLaMA2‑7B.
By Dongha Lee, Jinhee Park, Minjun Kim, Junseok Kwon
The paper examines OCR adaptation for low‑resource languages, noting that fine‑tuning often hits a performance ceiling in data‑scarce settings. It identifies that lower layers of language‑specific models learn redundant features while higher layers capture script nuances, leading to a structural inefficiency. To address this, the authors propose PSMC, a framework that pre‑trains a base model, specializes it per language, merges the experts via task arithmetic, and co‑trains a unified multilingual backbone, achieving about a 2% improvement in Word Recognition Rate across 10 Indian scripts without adding parameters.
By Achyuth P, Kahaan Shah, Chetan Arora
CommerceVibe is a system that generates e‑commerce creatives by synthesizing executable HTML/CSS code conditioned on product images, design requirements, and product information. It uses dual‑feedback reinforcement learning, combining rule‑based checks for text readability, product visibility, and layout validity with visual feedback from a vision‑language model that evaluates perceptual and commercial aspects. After fine‑tuning a large language model on 28,000 examples and applying dual‑feedback reinforcement learning, CommerceVibe achieves a weighted score of 94.0/100 on a 1,300‑case benchmark, outperforming both its SFT‑only counterpart and external models, and is validated by expert blind evaluations.
By Yajiao Xu, Jin Zhang, Jiangbo Ai, Tao Jiang, Mo Xu, Lina Huang, Chengfu Huo
The paper introduces QK Product Steering, a data‑free, training‑free method that edits the query‑key product in vision‑language models to reduce object hallucination. By suppressing a few dominant singular modes in selected middle layers and mapping the edited product back to query weights, the approach lowers hallucination rates without affecting inference cost. Experiments on three GQA‑based VLMs show a 4.0% average reduction in CHAIR$_s$, with the effect localized to symmetric mutual‑attention channels.
By Karn Tiwari, Varnith Chordia, Prathosh A P
The paper investigates why large reasoning language models struggle to transfer parametric knowledge across different scripts. Through observational data and regression analysis on ECLeKTic and MultiLoKo datasets, the authors find that script mismatch—not language family—is the main predictor of transfer failure when controlling for model capability and question difficulty. By providing key entities in the source language and training models to reason about transliteration ambiguities, they demonstrate a reduction in the cross‑script transfer gap, suggesting that post‑training improvements can enhance cross‑lingual knowledge transfer.
By Lucas Bandarkar, Alan Ansell, Trevor Cohn
The study investigates how large language models (LLMs) can assist humans in semantic memory search tasks. By using the semantic fluency task (SFT), the researchers evaluate whether LLMs can follow and enhance human mental trajectories during generative semantic retrieval. Results show that an LLM’s ability to track and predict human memory trajectories in this task surpasses that of other humans.
By Eric Lacosse, Mariana Duarte, Graham Todd, Peter M. Todd, Daniel C. McNamee
The paper introduces Mine and Refine, a two‑stage contrastive training framework designed to improve embedding‑based retrieval for large‑scale e‑commerce search. It tackles graded relevance, hard‑sample mining, and unstable similarity separability by using a lightweight LLM as a scalable labeler and a multi‑level circle loss to enforce margin‑controlled separation across relevance levels. The method has been deployed in production across multiple product verticals, yielding statistically significant increases in user engagement, gross order value, and retrieval relevance metrics.
By Jiaqi Xi, Raghav Saboo, Luming Chen, Johny Rufus, Aditya Dodda, Ved Sampath, Kenny Chi, Elyse Winer, Akshad Viswanathan, Martin Wang, Sudeep Das
The paper introduces Synth-JDoc, a synthetic Japanese document image dataset created by rendering text with HTML and CSS to produce multi‑column layouts that include both vertical and horizontal writing styles. Images generated by text‑to‑image models are embedded to enhance visual realism, and noise and degradation filters are applied to improve robustness. Experiments show that fine‑tuning Large Vision Language Models on Synth‑JDoc yields superior performance on reading vertically written Japanese text compared to prior synthetic datasets.
By Keito Sasagawa, Shuhei Kurita, Daisuke Kawahara
RegCL is a non‑replay continual learning framework that adapts the Segment Anything Model (SAM) for visual grounding across evolving multi‑sensorial media domains. It consolidates domain‑specific segmentation knowledge into a single lightweight SAM adapter by incrementally merging LoRA‑style AugModules and preserving compact historical feature statistics. Experiments on five heterogeneous datasets demonstrate that RegCL retains performance while adapting to new domains, outperforming other non‑replay continual learning and merging baselines.
By Yuan-Chen Shu, Zhiwei Lin, Xiaoyu Zhou, Yongtao Wang
The paper investigates cultural biases in large language models (LLMs) by introducing the Culture-Related Open Questions (CROQ) dataset, which contains 24‑language questions about generic culture. Experiments reveal that LLMs disproportionately favor Japan in their responses, especially when prompted in high‑resource languages, while low‑resource languages tend to highlight countries where the language is official. The study also finds that these biases emerge after supervised fine‑tuning rather than during pre‑training.
By Joseba Fernandez de Landa, Carla Perez-Almendros, Jose Camacho-Collados
The study investigates whether transferring relational structure from human mental representations to deep neural networks (DNNs) can improve fine‑grained alignment between the two. Using unsupervised Gromov‑Wasserstein optimal transport, the authors show that fine‑tuning pre‑trained DNNs with Relational Knowledge Distillation (RKD) brings the networks close enough to human representations to align at the individual‑object level on a test set of concepts not seen during training. The improvement is driven mainly by a more human‑like global structure of category distances, while local nearest‑neighbor overlap remains largely unchanged.
By Yuria Shimizu, Soh Takahashi, Takato Horii, Masafumi Oizumi
The paper investigates whether large language models can learn and reproduce annotator‑specific label‑explanation behavior, using two sentence‑pair tasks with four annotators each. It finds that individual annotator patterns are weak at the single‑annotation level but become detectable after reducing input‑content effects and aggregating across annotators. The authors propose cross‑annotator preference optimization (CAPO), which improves upon prompting and supervised fine‑tuning by better capturing annotator‑specific reasoning while maintaining stable attribution.
By Beiduo Chen, Pingjun Hong, Ziyun Zhang, Benjamin Roth, Anna Korhonen, Barbara Plank
arXiv:2510.25384v2 Announce Type: replace
Abstract: Large Language Models (LLMs) are promising tools for synthetic data generation in mental health. However, privacy policies and restrictions forced...
By Doan Nam Long Vu, Rui Tan, Lena Moench, Svenja Jule Francke, Daniel Woiwod, Florian Thomas-Odenthal, Sanna Stroth, Tilo Kircher, Christiane Hermann, Udo Dannlowski, Hamidreza Jamalabadi, Simone Balloccu, Shaoxiong Ji
The paper introduces "Destroy Me", a hybrid framework that generates realistic histopathology artifacts using Stable Diffusion and physics‑based modeling to create six common artifact types. Artifact realism is evaluated with KID and color Wasserstein metrics, and models trained on these augmented images outperform baselines on lung adenocarcinoma classification, achieving a 10.5% relative boost in macro F1‑score and a 15% increase in Cohen’s Kappa. The study highlights that selective, impact‑weighted augmentation is essential for enhancing robustness while preserving subtle diagnostic features.
By Zuzanna Krawczyk-Borysiak, Adam Krawczyk, Mateusz Miller, Gabriela Kaczmarek, S{\l}awomir Paku{\l}o, Ma{\l}gorzata Sok\'o{\l}, \.Zaneta Swiderska-Chadaj
arXiv:2607.00250v5 Announce Type: replace
Abstract: Maltese has substantial text corpora and pretrained language models, but paragraph-scale OCR training data remains scarce; NOMOCRAT provides 57 ver...
By Adam Darmanin
NL2AGBench is a benchmark that evaluates how well large language models can translate English geometry problems into the formal language required by AlphaGeometry’s theorem‑proving engine. The study tests ten state‑of‑the‑art LLMs, comparing executable translation accuracy, syntactic correctness, and error types, and finds a large gap between closed‑source and open‑source models. The authors also propose an error taxonomy and test mitigation strategies such as few‑shot prompting, fine‑tuning, and human‑guided hinting, which improve performance across model families.
By Samuel Xiao, Judy Song, Rory Hu, Ziliang Zong
The paper introduces MSCGC-KAN, a new EEG emotion recognition approach that builds on a pre‑trained CBraMod backbone. It incorporates a structured task head featuring multi‑scale causal graph convolution and Kolmogorov–Arnold feature mapping to better capture multi‑scale emotional dynamics, inter‑channel connectivity, and nonlinear discriminative patterns. Experiments on FACED and SEED‑VII show significant performance gains over a linear baseline, achieving balanced accuracies of 60.66% and 33.27% respectively.
By Haoliang Gong, Qingshan She, Jiale Xu, Yunyuan Gao, Xugang Xi
D-TAIA is a framework that adapts large language models for multi‑task predictive process monitoring, jointly predicting the next activity and remaining time of ongoing cases. It uses domain‑aware triplet loss pre‑training, FAISS‑based nearest‑neighbor retrieval for time estimation, and a TAIA inference strategy to preserve sequential reasoning while fine‑tuning a 10 M‑parameter backbone. Across four real‑world event logs, D‑TAIA achieves state‑of‑the‑art or competitive results compared to a fine‑tuned LLM and a recurrent neural network baseline, with ablation studies showing the effectiveness of NLP and computer‑vision techniques for this domain.
By Sjoerd van Straten, Christine Jacob, Marwan Hassani