The study evaluates whether a single large language model (LLM) can handle multiple customer‑support tasks or if separate specialist models are preferable. Using 13 models from five families and 200+ checkpoints across eight datasets, the authors find that multi‑task full fine‑tuning consistently outperforms other strategies. They also show that sequential LoRA and model merging can preserve earlier skills and improve off‑task robustness, offering practical guidelines for real‑world deployment.
By Md Tahmid Rahman Laskar, Xue-Yong Fu, Shashi Bhushan TN
PRISM‑VLM is a new benchmark for compact vision‑language models that evaluates each item across seven axes—task quality, behavioral robustness, and capability bottlenecks—rather than collapsing performance into a single accuracy score. It aggregates these axes into a single PScore while also providing per‑axis profiles, revealing differences that single‑axis benchmarks miss, such as sycophancy. The benchmark draws on items from fifteen public datasets and will be released with its full pipeline, prompts, and annotations.
By Sanghee Park, Kee-Eung Kim
The paper introduces a fixed‑budget revision protocol that uses deterministic verifiers to expose all remaining violations across exact‑length, lexical, and compositional constraints, thereby isolating model‑side revision behavior. Experiments on 19 open‑ and closed‑source LLMs show wide variability in controller‑level success, with some models achieving up to 99.8% success while others remain below 20%. Controlled studies reveal that post‑training and scale affect model responses to exact feedback, but do not consistently improve exact correction, and that recurrence of earlier outputs is linked to lower recoverability.
By Haitong Jiang, Chunlin Liu, Yile Wang, Yuhong Feng
The paper presents a new evaluation protocol for post‑training quantization of speech language models that separates lexical output, transcript‑insufficient endpoints, and packed implementations. In a Qwen2‑Audio case study, a 6‑bit allocation selected for translation improves chrF scores but degrades emotion recognition, while uniform and front‑layer controls perform better on emotion tasks. Similar patterns hold at 7 bits, and a 4‑bit study shows consistent emotion deficits across all low‑bit allocations, with no advantage for the selected scheme. The study highlights a precision‑dependent mismatch between lexical output, waveform‑dependent behavior, and nominal precision, without claiming a general failure of low‑bit models or a deployment benefit for the selected allocation.
By Mengzhe Geng, Jinxi Jin, Junhao Xu
ContraVis is a visual analytics system that assists human reviewers in detecting contradictions within legal contracts by modeling contracts as typed paragraph graphs. These graphs combine explicit contractual references with semantic relationships, enabling the system to condition large language model (LLM) reasoning and provide an interactive representation for analysts. In controlled experiments, graph-conditioned LLM reasoning identified more injected contradictions than standalone LLM analysis, and a formative study with lawyers showed that in-context evidence comparison helped validate contradictions.
By Luis Sante, Paula Lima, Mariana Rocha, Jorge Poco
Mizar is a 159.3‑million‑parameter audio‑language model designed for devices with limited memory and computation. It couples a compact CED‑Small audio encoder with SmolLM2‑135M via a frequency‑merging mapper and is trained in three stages—audio‑language alignment, audio‑dependent fine‑tuning, and post‑training—to improve performance on audio‑question tasks. Across five random seeds, Mizar outperforms all other sub‑200M‑parameter ALMs on MMAU, MMAR, and ADQA‑clean, achieving mean accuracies of 52.92%, 42.42%, and 36.02% respectively, while enabling local inference on a single CPU with an average latency of 1.09 seconds for MMAU questions.
By Kaiyang Li, Shaobo Han, Yue Tian, Shihao Ji
LiSeCo is a lightweight, gradient‑free method that controls language generation by directly intervening on the hidden activations of a token in embedding space. It uses control‑theoretic techniques to steer the generation trajectory away from undesired semantic regions and into a predefined allowed region, ensuring fine‑grained attribute control. The approach is computationally efficient, minimally impacts generation time, and is shown to be effective on tasks such as toxicity, sentiment, and bilingual language steering while preserving text quality.
By Emily Cheng, Carmen Amo Alonso
The paper introduces HIVE, a method for detecting hallucinations in diffusion large language models by leveraging hidden trajectory evidence. HIVE compresses hidden states, selects informative step-layer evidence, and uses continuous prefix embeddings to condition a verifier that outputs a hallucination score and diagnostics. Experiments on two diffusion LLMs and three QA benchmarks show that HIVE outperforms eight baselines and improves AUROC and AUPRC by several points over text-only verification.
By Guoshenghui Zhao, Tan Yu, Weijie Zhao
MetaHOPE is an error‑severity‑aware annotation framework designed to evaluate how well machine translation (MT) and large language models (LLMs) translate metaphors. The authors applied MetaHOPE to three state‑of‑the‑art systems—GoogleMT, GPT5.4, and Hunyuan‑7b—using two human‑annotated metaphor corpora (VUAMC and PSUCMC) for English‑to‑Chinese and Chinese‑to‑English translation. They also produced a bilingual post‑edited gold reference, creating a new resource for metaphor translation research.
By Jiahui Liang, Lifeng Han
The study investigates how multilingual training affects figurative language identification in proverbs, using 742 proverb concepts translated into seven languages. Five models—including multilingual encoders and instruction‑tuned LLMs—were evaluated with varying levels of multilingual supervision, and a new multidimensional annotation framework was introduced to classify proverbs into metaphorical, moral/advisory, cause‑effect, and culture‑specific forms. Results show that adding multilingual data beyond 50% yields limited gains, but the best supervision level depends on the model and language; combining diverse figurative forms improves overall performance, especially for the least frequent culture‑specific form.
By Rama Alomair, Remas Alsubaie, Walaa Saifalislam, Rima Alsonbul, Mona Alnajjar, Razan Aldossari, Haya Alibrahim, Abeer Aldayel
The study applies causal tracing to large audio language models (LALMs) to uncover how they fuse acoustic and textual information. Layer‑wise analysis reveals distinct fusion strategies—progressive integration in DeSTA versus abrupt late‑stage fusion in Qwen—while token‑wise analysis identifies the final sequence token as an informational bottleneck that decisively retrieves audio content. Additionally, an attention‑like query mechanism at intermediate tokens is observed, prompting the model to pull task‑relevant audio context.
By Wei-Chih Chen, Chien-yu Huang, Hung-yi Lee
nnFoundation introduces complementary convolutional and transformer-based 3D foundation models for radiology, trained on 2.1 million CT, MRI, and PET volumes from 125 datasets. The models are evaluated on 108 tasks—including segmentation, detection, classification, report generation, and image retrieval—under domain shift, low-data, and low-compute scenarios, consistently outperforming prior 3D foundation models and training from scratch. Performance varies by task type, with convolutional models excelling at spatially localized tasks and transformer models at global semantic reasoning, and dynamic alignment with dataset characteristics further enhances transferability.
By Constantin Ulrich Harsy, Tassilo Wald, Karol Gotkowski, Yannick Kirchhoff, Marcel Knopp, Maximilian Rokuss, Elisa Stegmeier, Philipp Schader, Dasha Trofimova, Raphael Stock, Kim-Celine Kahl, Stephen Schaumann, Selen Erkan, David Zimmerer, Stefan Denner, Moritz Langenberg, Sebastian Ziegler, Katharina Eckstein, Maximilian Fischer, Jonathan Suprijadi, B\'alint Kov\'acs, Benjamin Hamm, Anand Deshpande, Dimitrios Bounias, Nico Disch, Shuhan Xiao, Jessica K\"achele, Jan Sellner, Rajesh Baidya, Jeremias Traub, Lars Kr\"amer, Maximilian Zenk, Tim R\"adsch, Stefan Dvoretskii, Robin Peretzke, Jonathan Deissler, Alexandra Ertl, Partha Ghosh, Kris Dreher, Stefan Dinkelacker, Annika Reinke, Evangelia Christodoulou, Numan Saeed, Yoland Savriama, Santiago Estrada, David K\"ugler, Laura Alexandra Daza Barragan, Cristina Isabel Gonzalez Osorio, Jan Peeken, Michael Baumgartner, Marvin Teichmann, Guillaume Chabin, Matthias Kirchler, Valentin Koch, for the ALFA study, Markus Hohenhaus, Dimitri Koslov, Nina Decker, Mohammad Yaqub, Arnd Heuser, Martin Reuter, Julia A. Schnabel, Tobias Heimann, Florin Ghesu, Paul Brachmann, Claus P. Heu{\ss}el, Alexander Radbruch, Gianluca Brugnara, Aditya Rastogi, Martha Foltyn-Dumitru, Heinz-Peter Schlemmer, Ignaz Reicht, Julius C. Holzschuh, Michael Bach, Bram Stieltjes, Kai Schlamp, Lena Maier-Hein, Marco Nolden, Ralf Floca, Paul F. J\"ager, Philipp Vollmuth, Fabian Isensee, Klaus H. Maier-Hein
The paper introduces Motion Vision CAPTCHA (MVCAP), a new CAPTCHA framework that relies on motion-defined foreground structures to create challenges that are only solvable through temporal analysis of a dynamic background. MVCAP is implemented in three progressive levels—coherent motion, structural motion, and biological motion—and evaluated using the MVCAP-Bench, a browser-based benchmark with 600 live CAPTCHA instances. Human participants achieve 99.6% accuracy, whereas the best GUI agent scores only 16.8%, highlighting a significant human–agent perception gap and demonstrating that dynamic background camouflage is the key difficulty.
By Zeyu Zhang, Dingyi Rong, Zijian Chen, Zicheng Zhang, Xiongkuo Min, Guangtao Zhai
AWM‑VLA introduces a unified framework that embeds aligned world modeling directly into a diffusion‑transformer vision‑language‑action policy. By adding learnable future tokens aligned with vision‑language embeddings of future observations, the policy can anticipate long‑term consequences while generating actions. The method extends this with an object‑centric alignment objective and a principled weighting scheme, achieving up to 21% higher success rates on RoboCasa and humanoid tabletop benchmarks and producing object‑centric rationales preferred by human raters in 83% of cases.
By An Lanji, Dawei Liu, Jin Li, Haoran Xu, Mei Chen, Yu Tian
Groundbench is a new benchmark that evaluates vision‑language models on multi‑resolution polygon grounding, using the same 1,500 image‑expression‑referent triples but targeting exact‑N polygons with five different vertex budgets. It audits both filled‑region intersection‑over‑union (IoU) and legal‑polygon completion, revealing that the best models achieve 88.2 box IoU and 97.1 accuracy at IoU ≥ 0.5, while direct polygon predictions lag at 57.7 and 69.2. The study shows performance is non‑monotonic across budgets, collapses at the densest budget due to legality failures, and highlights that false spatial cues hurt more than false colour cues, underscoring an operational geometry gap beyond latent boundary perception.
By Zhonghan Bian, Zhenran Wang, Jinsong Li, Zhangyang Qi
UVU is a vision-language unified autoregressive framework that integrates visual supervision directly into the pre-training stage of multimodal large language models. By using continuous visual encoding and a large-scale iterative hierarchical clustering algorithm to build a pixel-level visual codebook, UVU enables lossless representation of visual inputs and autoregressive generation of pixel-level image tokens alongside textual tokens. This approach synergizes pixel-level visual perception with semantic-level visual understanding, allowing models to internalize visual reconstruction capabilities and improve multimodal understanding performance.
By Zhehan Kan, Xinghua Jiang, Yubo Zhu, Yanlin Liu, Xiaochen Yang, Zhixiang Wei, Shifeng Liu, Qingmin Liao, Wenming Yang, Xin Li, Yinsong Liu, Deqiang Jiang, Xing Sun
VIVAS is a new Vision‑Language Model pre‑training framework that addresses the lack of fine‑grained visual perception in existing VLMs. It introduces a unified token space and a dense‑structural‑semantic vision tokenizer that expands the textual vocabulary with visual tokens, enabling vision‑language unified autoregressive supervision over both visual details and linguistic content. Trained on 12.4 T tokens, VIVAS achieves state‑of‑the‑art results on 7 tasks and 39 multimodal benchmarks.
By Zhehan Kan, Yubo Zhu, Xinghua Jiang, Zhixiang Wei, Shifeng Liu, Wei Tong, Sheng Zhong, Qingmin Liao, Wenming Yang, Xin Li, Yinsong Liu, Deqiang Jiang, Xing Sun
The paper introduces O$^2$-VG, a unified framework for oriented object visual grounding in remote sensing images, comprising three complementary models: O$^2$-VG-Trans, a cross‑modality transformer; O$^2$-VG-Uni, which predicts universal oriented proposals; and O$^2$-VG-VLM, an autoregressive vision‑language model that generates oriented bounding boxes. It also presents DIOR‑R‑SVG, a new dataset containing image, expression, and oriented box triplets for training and evaluation. The framework demonstrates superior performance across multiple benchmarks and is supported by publicly available code.
By Zeyu Ding, Yong Zhou, Jiaqi Zhao, Wen-Liang Du, Xixi Li, Hancheng Zhu, Rui Yao, Abdulmotaleb El Saddik
The paper introduces an automated high‑throughput microscopy system for melissopalynology that combines H∞ robust mechanical control with deep learning pipelines. It uses U^2‑Net for salient object detection and a DINOv2 Vision Transformer trained via deep metric learning for pollen grain classification, augmented with Gradient‑Weighted Attention for interpretable texture annotations. The system reports a 95.8% classification recall and at least a six‑fold speedup over manual expert analysis.
By J. Staforelli-Vivanco, R. Jofr\'e, P. Coelho, I. Sanhueza, L. Viafora, C. Toro, J. Troncoso, M. Rondanelli-Reyes, I. Lamas, Andy Banegas-Medina, Isis-Yelena Montes, B. Mu\~noz-Cepeda, V. Salamanca-Levi, M. Gonz\'alez-Ortiz, E. Vera
The study evaluates how well pre‑trained models can classify the Bloom level of AI‑generated educational questions, a task that is crucial for ensuring pedagogical quality. Traditional machine‑learning models perform poorly on out‑of‑distribution data, whereas transformer and large‑language models achieve higher accuracy, especially after feature‑engineering techniques such as text splicing and appending learning objectives. Retraining the models yields the most significant performance gains across all datasets.
By Michael Lawrence Castanares, Princess Ventures, Allan Tan