RefDiT is a new framework for reference-guided image generation that addresses the shortcomings of previous methods when handling complex scenes with multiple objects. It introduces local region guidance by decomposing a single identifier token into attribute-level signals, allowing the model to learn correspondences between tokens and specific regions of a reference image. The approach incorporates a low-rank adapter (LoRA) within a diffusion transformer (DiT) to adjust the inference prompt based on user-provided guidance context, thereby enabling more precise local attribute control.
By Rameshwar Mishra, Srikrishna Karanam, A V Subramanyam
The paper introduces SEAL, a plug‑and‑play module that enhances single‑image sticker personalization in diffusion models by adding a semantic‑guided spatial attention loss, a split‑merge token strategy, and structure‑aware layer restriction. SEAL integrates without altering the U‑Net backbone and addresses overfitting issues such as visual entanglement and structural rigidity. Alongside SEAL, the authors release StickerBench, a large sticker dataset with six attribute tags to enable systematic evaluation of identity preservation and contextual controllability.
By Changhyun Roh, Yonghyun Jeong, Jonghyun Lee, Chanho Eom, Jihyong Oh
DART is a new RGB‑D pretraining method for surgical vision foundation models that incorporates pseudo‑labeled depth maps as a pixel‑space reconstruction target during training. By adding a depth reconstruction head to DINOv2’s masked iBOT framework, DART improves representation quality without affecting downstream RGB‑only fine‑tuning or inference. Across eight surgical benchmarks—including segmentation, depth estimation, and image‑level recognition—DART outperforms both natural‑image and in‑domain baselines, demonstrating that geometric pseudo‑labels can strengthen foundation model pretraining without extra labels or inference cost.
By John J. Han, Adam Schmidt, Muhammad Abdullah Jamal, Jie Ying Wu, Omid Mohareri
Mitra‑v2 is a tabular foundation model that achieves state‑of‑the‑art performance on a wide range of real‑world classification and regression tasks, including credit‑risk scoring, clinical prediction, equipment‑failure detection, and house‑price estimation. Trained solely on synthetic data with a larger and more diverse pretraining distribution than its predecessor, it uses a compact 2D Transformer backbone and improved optimization to handle longer contexts and larger feature spaces. On the TabArena and TALENT benchmarks, Mitra‑v2 outperforms leading models such as TabPFN‑3 and TabICLv2, matching the performance of a 1.6B‑parameter TabFM with only 77M parameters, and ranks first on multi‑class classification tasks with more than ten classes.
By Yefan Tao (Bernie), Xiyuan Zhang (Bernie), Xinyi Liu (Bernie), Boran Han (Bernie), Danielle Maddix (Bernie), Haoyang Fang (Bernie), Zhen Han (Bernie), Jiading Gai (Bernie), Xuanqing Liu (Bernie), Michael Bohlke-Schneider (Bernie), Yuyang (Bernie), Wang, Gerald Friedland, Kevan Mah, Chris Lee, Chris Kong
The paper investigates whether pretrained image models can generalize to unseen datasets by clustering their embeddings. Using encoders trained only on ImageNet‑1k, both supervised and self‑supervised, the authors evaluate clustering performance on out‑of‑domain images. They find that supervised encoders perform better within the training domain, while self‑supervised encoders excel far outside it, and that fine‑tuning self‑supervised models reverses this trend. Additionally, the study shows that the silhouette score in UMAP‑reduced space correlates strongly with clustering accuracy, offering a proxy metric when labels are unavailable.
By Scott C. Lowe, Joakim Bruslund Haurum, Sageev Oore, Thomas B. Moeslund, Graham W. Taylor
The paper introduces Ref-GeNVS, a training‑free, reflection‑aware approach for generative novel view synthesis in mirror scenes. It treats a mirror image as two complementary views, estimates the mirror plane and reflected camera poses, and uses a two‑stage generation process with Mirror‑gated attention and Reflection injection to produce reflection‑consistent novel views. The method leverages a multi‑view diffusion backbone without finetuning, outperforming recent generative NVS methods on synthetic and real mirror scenes.
By GeonU Kim, Shin Dong-Yeon, Tae-Hyun Oh
PLUME is a lightweight framework for personalizing large language models by learning a shared task-specific subspace from aggregated user data and then training only a small square matrix for each user within that subspace. The approach introduces cross-layer shared parameters and rank‑1 residual terms to reduce redundancy while preserving expressiveness. Experiments on personalized text generation benchmarks show that PLUME matches or outperforms strong baselines while cutting per‑user parameters by over 95%.
By Xinyu Li, Hao Zhou, Jianfeng Zhu, Julina Maharjan, Ruixin Guo, Feodor Dragan, Ruoming Jin
The paper explores using reinforcement learning to enhance automatic text simplification for low‑resource languages, focusing on Catalan. It introduces a new reward function that blends the SARI metric with penalty terms, and applies Group Relative Policy Optimization (GRPO) to fine‑tune the IberianLLM‑7B‑Instruct model on the ASSET dataset. Post‑training, the model shows improved simplification performance on two Catalan benchmarks and reduces prior negative behaviors, though cross‑lingual transfer from English, Spanish, and Catalan translations of ASSET does not yield significant gains on an out‑of‑domain benchmark.
By Arnau Ayguad\'e Domingo, Stefan Bott, Horacio Saggion
The paper introduces KoNA, a benchmark designed to evaluate selective non‑compliance in vision‑language models (VLMs) across five categories—False Premise, Visual Inaccessibility, Universal Unknown, Task Feasibility, and Safety. KoNA tests both query‑level and component‑level non‑compliance using paired single and compound queries, revealing that many VLMs struggle to refuse, correct, or abstain appropriately, especially when selective non‑compliance is required. Fine‑tuning VLMs on KoNA examples improves non‑compliance accuracy while preserving performance on fully answerable tasks, indicating that models can better distinguish answerable components from those needing non‑compliance.
By Minji Kim, Jihyoung Jang, Hyounghun Kim
The paper introduces Iris-mini and Iris-pro, two search agents trained at 35B and 397B parameter scales. They use a novel data pipeline that constructs reverse‑engineered multi‑hop queries from web hyperlinks, filters trajectories, and alternates supervised fine‑tuning with reinforcement learning in a process called SFT‑RL climbing. Evaluations on several benchmarks show that, with inference‑time context management, the agents achieve the best open‑source results in their parameter ranges.
By Ziyuan Liu, Hengqi Liu, Zichuan Wang, Yang Qin, Jiachen Liang, Xu Chu, Shaowei Chen, Yuantao Gu, Mu Chuan
The study investigates whether self‑supervised pretraining improves molecular graph neural networks by adapting the LeJEPA architecture to molecular graphs. While pretraining enhances learned representations and a frozen probe outperforms random initialization on tasks such as ogbg‑molhiv, it does not consistently boost finetuning performance across different data splits. Combining pretrained embeddings with 1024‑bit Morgan fingerprints yields modest gains, indicating that pretraining provides complementary information best exploited at the feature level.
By Micha{\l} Kulczykowski, Rafa{\l} {\L}ab\k{e}dzki
The paper explores using a language model’s low‑level symbolic skill—specifically SQL—to detect hallucinations without fine‑tuning. By having the model construct an SQL database from reference documents, it can reason over both the source and the model’s output, creating a neurosymbolic check. Experiments on RAGTruth and DiaHalu show this method outperforms direct prediction and rivals state‑of‑the‑art detectors, highlighting the value of leveraging inherent symbolic competences in LLMs.
By Renato Vukovic, Hsien-chin Lin, Carel van Niekerk, Benjamin Ruppik, Michael Heck, Shutong Feng, Nurul Lubis, Milica Gasic
The paper explores how to adapt tabular foundation models (TabFMs) for censored time‑to‑event prediction by linking them with CoxPH and DeepHit and revising training procedures. It evaluates zero‑shot, classification‑based fine‑tuning, and survival‑head adaptations across 74 single‑risk and 4 competing‑risk datasets, finding that zero‑shot works best on small datasets while supervised adaptation excels as data grows. The study shows that the choice of adaptation interface and data regime critically influences TabFM transfer performance.
By Minh-Khoi Pham, Luca Cotugno, Dan Cernei, Alina Sirbu, Stefano Masi, Giuseppe Prencipe, Alessandro Pingitore, Patrizia Landi, Working Group on Uric Acid, Cardiovascular Risk of the Italian Society of Hypertension, Tai Tan Mai, Martin Crane, Marija Bezbradica
The paper investigates how pause tokens affect fine‑tuning dynamics rather than just inference performance. Experiments on synthetic continual‑learning and math‑reasoning tasks show that masked pause tokens (MBP) reduce overwriting of learned modes and compress downstream‑step information. Across 1B‑8B Qwen and Llama models, MBP consistently improves reasoning scores while preserving language understanding, and the approach also benefits GRPO.
By Jaehyeon Kim, Suhwan Kim, Nakyung Lee, Yeongoon Kim, Jimin Seo, Giho Lee, Jungwoo Lee
The paper evaluates three open‑weight large language models—Gemma3‑12B (USA), Bielik‑11B‑v3 (Poland), and Qwen3‑4B (China)—against World Values Survey data for 63 demographic personas across three countries, using normalized Wasserstein distance to measure cultural misalignment. Surprisingly, none of the models shows a preference for its home country; Qwen3‑4B, built in China, has the highest misalignment for Chinese respondents. Targeted LoRA fine‑tuning on the five worst‑case personas, with fewer than 1,200 training pairs and under 15 minutes on a single GPU, reduces bias by 16.8% for Bielik‑11B, but the fine‑tuning redistributes bias rather than eliminating it, shifting worst‑case personas from American to Chinese elderly.
By Antoni Czolgowski, Abel Iyasele
The paper introduces Aplaud, a lightweight framework that personalizes large language models for survey response prediction. Aplaud builds on LoRA by separating adaptation into a shared low‑rank basis, a compact user‑specific correction, and a rank‑one residual for finer personalization. Experiments show that Aplaud outperforms existing LoRA‑based methods in both generalization and inference efficiency while keeping per‑user parameter costs low.
By Xinyu Li, Ruoming Jin, Jianfeng Zhu, Ruixin Guo, Zhi Liu
ConfRAG introduces a confidence-guided approach to reduce hallucinations in large language models and selectively trigger Retrieval-Augmented Generation (RAG) only when the model is uncertain. The ConfQA fine‑tuning strategy trains the model to answer correctly or respond with "I am unsure," achieving a drop in hallucination rates from 20‑40% to below 5% across factuality benchmarks. Building on ConfQA, ConfRAG limits external retrievals by more than 30% while maintaining over 95% accuracy in ideal scenarios.
By Yin Huang, Yifan Ethan Xu, Kai Sun, Vera Yan, Alicia Sun, Haidar Khan, Jimmy Nguyen, Jingxiang Chen, Mohammad Kachuee, Zhaojiang Lin, Yue Liu, Aaron Colak, Anuj Kumar, Wen-tau Yih, Xin Luna Dong
The paper introduces CARNIVAL, a model for protein annotation in cryo-electron tomography (cryo-ET) volumes that leverages simulated data and a forward model to generate domain‑specific augmented paired views for self‑supervised training. By incorporating simulation‑derived protein positions and identities into the architecture and loss function, the model localises semantic information at protein locations. CARNIVAL is evaluated on real tomograms without finetuning and outperforms a state‑of‑the‑art contrastive model that lacks forward‑model paired views or privileged information.
By Bogdan Toader, Kiarash Jamali, Tanmay A. M. Bharat, Sjors H. W. Scheres
The paper introduces PRICE, a systematic framework for adapting Large Language Models to short‑term Bitcoin price forecasting. PRICE combines parameter‑efficient fine‑tuning with LoRA, recursive multi‑step inference, integer‑rounded numerical representation, Context‑Task‑Format prompting, and exact zero‑temperature decoding, all built on a 4‑bit quantized LLaMA‑3 8B model. Ablation studies and comparative evaluations show that each component improves accuracy and reliability, enabling PRICE to achieve the lowest forecasting errors among eight transformer‑based and time‑series foundation models.
By Maryam Fakhari, Mehran Safayani
The study investigates whether the number of discrete class‑separability jumps (phase transitions) observed during ResNet fine‑tuning can predict final test accuracy. Across 75 experiments on four benchmarks (CIFAR‑10, CIFAR‑100, TinyImageNet, CIFAR‑10‑C) and three ResNet variants, a strong negative correlation is found on standard i.i.d. datasets (r = −0.84 on CIFAR‑10, r = −0.87 on CIFAR‑100), while the correlation weakens under distributional stress. Additional analyses show that the transition count retains predictive power after controlling for architecture depth and outperforms other training‑curve signals on in‑distribution benchmarks, though it is dominated by other signals on stressed datasets.
By Arunan J