The study evaluates self‑supervised learning (SSL) models pretrained on ImageNet‑1k and the Human Protein Atlas (HPA) Field‑of‑View (FOV) for protein localization in microscopy images. DINO‑based Vision Transformer backbones pretrained on either dataset transfer well to the OpenCell dataset, achieving strong performance even without fine‑tuning and improving further when fine‑tuned (0.704 ± 0.027 macro F1 on 17 classes). At the single‑cell level, the HPA‑pretrained model outperforms others in k‑nearest‑neighbor classification across all neighborhood sizes (macro F1 ≥ 0.515).
By Ben Isselmann, Dilara G\"oksu, Heinz Neumann, Andreas Weinmann
Jina-OCR-v1 is an end‑to‑end document parsing model designed for low‑budget GPUs, combining a compressed‑vision encoder with a 3B mixture‑of‑experts decoder that activates about 570 M parameters per token. It uses a FastMTP speculative decoding head that shares a single draft block across three prediction steps, with greedy verification ensuring lossless decoding. Post‑training includes instruction alignment, robustness fine‑tuning on difficult documents, and GRPO with dense verifiable rewards, achieving 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR‑Bench while delivering the highest page throughput at 2.57 pages per second on an NVIDIA L4 GPU.
By Alejandro Bar\'on Garc\'ia, Feng Wang, Emilia Garcia Casademont, Han Xiao
The paper introduces a novel compression framework for image-to-shape Diffusion Transformers (DiTs) that significantly reduces model size while preserving geometric fidelity. By exploiting the non-uniform importance of 3D DiT layers, the authors combine structured pruning, adaptive quantization, and targeted fine‑tuning into a vitality‑guided approach. The method achieves up to a 66% reduction in model size across state‑of‑the‑art image‑to‑3D models without compromising synthesis quality, offering a plug‑and‑play solution for efficient 3D shape generation.
By Jaeah Lee, Hyunjin Kim, Jaewoong Cho, Gihyun Kwon
The paper presents a two‑pipeline framework for retinal fundus analysis that combines four‑class disease classification with vessel segmentation. It fine‑tunes eight ImageNet‑pretrained CNNs on the FIVES dataset, applies five gradient‑based explanation methods to assess model interpretability, and benchmarks ten U‑Net variants—including transformer‑based and attention‑enhanced architectures—on the FIVES and DRIVE datasets. The best classification results come from ResNet101 (94.17% accuracy), while the strongest segmentation performance is achieved by Attention U‑Net with a ResNet101V2 backbone, improving DRIVE IoU from 60.80% to 64.83%.
By Fatema Tuj Johora Faria, Mukaffi Bin Moin, Pronay Debnath, Asif Iftekher Fahim, Faisal Muhammad Shah
SimCast‑S2S is a generative latent‑diffusion framework designed for probabilistic subseasonal‑to‑seasonal precipitation forecasting. It tackles three key challenges: it uses a diffusion‑based generative pipeline for uncertainty quantification, operates in a compact latent space learned by VAEs for efficient large‑ensemble generation, and employs transfer learning with LoRA to overcome limited training data. On reanalysis data, it outperforms deep‑learning baselines and competes with or surpasses state‑of‑the‑art operational systems such as ECMWF‑S2S.
By Hiep V. Dang, Antonios Mamalakis
The paper investigates whether the rank of latent matrices in matrix‑chain‑of‑thought (Matrix‑CODI) models influences performance on reasoning tasks. Across multiple training regimes on ProsQA and GSM8K‑Aug, rank‑k projection ablations show flat accuracy curves, indicating that truncating the latent matrix to low rank does not hurt performance. Experiments with various readout architectures—bilinear, bilinear‑plus‑GELU, SVD‑augmented, and quadratic—confirm that rank‑indifference persists even for nonlinear readouts, and a linear probe on the latent matrix underperforms a raw pretrained hidden state.
By Samuel Larson (Pebble ML)
TRACE is a graph‑network simulator that stores granular contact history on edges using attention‑based message passing and a gated recurrent unit, with an edge‑identity dictionary to preserve memory as contacts change. It predicts normal and tangential forces while enforcing Coulomb friction and internal force balance, and is trained via single‑step pretraining followed by autoregressive fine‑tuning. On 2D and 3D granular column‑collapse benchmarks, TRACE achieves significantly lower long‑rollout position and deposit errors than existing simulators, uses fewer parameters, maintains near‑zero particle interpenetration, and outperforms the material point method by 12.2× (2D) and 8.9× (3D).
By Changjian Zhou, Negin Yousefpour, Jie Qi, Junfeng Fang, Guillermo A. Narsilio, Hans Petter Jostad
LightEMMA is a longitudinal evaluation framework that tests the autonomous driving performance of vision‑language models (VLMs) without fine‑tuning or prompt engineering. Using this protocol, the authors evaluated 15 VLMs from five major families on the nuScenes prediction benchmark and found that larger, more capable models do not consistently outperform earlier generations. The study identifies common failure modes such as overreliance on historical actions and difficulty reconciling conflicting visual cues, underscoring the need for domain‑specific adaptation to enhance VLM safety in autonomous driving.
By Zhijie Qiao, Haowei Li, Zhong Cao, Henry X. Liu
EmoDistill is an offline framework that distills emotional negotiation skills from large language model interactions into smaller agents. It separates emotion selection, handled by an Implicit Q‑Learning selector, from emotion‑conditioned expression, learned by a LoRA‑adapted 7B policy via supervised fine‑tuning and judge policy optimization. Experiments across four negotiation domains show that the full EmoDistill policy outperforms vanilla and IQL‑only baselines, while removing the explicit emotion channel markedly reduces negotiation utility and reveals partial, domain‑dependent transfer to unseen counterparties.
By Yunbo Long, Haolang Zhao, Lukas Beckenbauer, Liming Xu, Alexandra Brintrup
The paper investigates weight‑space merging of independently fine‑tuned multilingual machine translation models. Experiments show that merging is more successful when models share a target language, yet it still cannot match the peak performance of language‑specific checkpoints. When target languages differ, performance drops sharply, and analysis reveals that overlapping neuron activation and incompatible upper‑layer geometries cause these failures.
By Baban Gain, Trilok Nath Singh, Asif Ekbal
The paper investigates how retrieval‑augmented generation (RAG) affects single‑turn mental‑health question answering. It introduces a lightweight selective retrieval policy that activates retrieval only when user queries exhibit psychoeducational, coping, or specificity needs, guided by a rule‑based safety trigger. Experiments show that unconditional retrieval improves specificity but degrades overall quality and safety, whereas selective retrieval maintains closed‑book performance for low‑need cases while avoiding these negative effects.
By Hyunseo Oh, Chong-Kwon Kim, Yoonhyuk Choi
The paper introduces a benchmark and evaluation system for measuring how well generative image models preserve the identity of a subject across generation, editing, restoration, and multi‑subject scenarios. It compares three paradigms—input context, trainable subject‑specific parameters, and a persistent identity layer—showing that persistent identity consistently improves fidelity while keeping image quality and instruction adherence high. The study finds that identity preservation remains a distinct limitation of current foundation models, especially under iterative edits, small scales, severe degradation, and multi‑subject composition.
By Mengwei Ren, Xuaner Zhang, Zhihao Xia
PrivateHub is a contrastive diffusion model designed to generate synthetic multi‑sensor data that protects private user activities while keeping non‑private applications detectable. It operates in two stages: App‑Conditioned Pre‑training, which conditions the model on application embeddings, and App‑Aware Fine‑tuning, which uses contrastive learning to separate private from non‑private data. Experiments on three real‑world datasets demonstrate that PrivateHub reduces private‑application inference accuracy by 40–50% without harming non‑private performance and remains robust even when attackers retrain on the synthetic data.
By Jiechao Gao, Yuandong Pan, Jie Wang, Michael Lepech, Bradford Campbell
The paper introduces VoxPrivacy, a benchmark for assessing interactional privacy in Speech Language Models (SLMs). It evaluates models on a 32‑hour bilingual dataset across three difficulty tiers, revealing that most open‑source SLMs perform near random on conditional privacy decisions and even strong closed‑source systems struggle with proactive privacy inference. The authors also validate these findings on a real‑speech subset and show that fine‑tuning on a 4,000‑hour training set can improve privacy‑preserving capabilities while maintaining robustness.
By Yuxiang Wang, Hongyu Liu, Dekun Chen, Xueyao Zhang, Zhizheng Wu
The paper reviews how Large Language Models (LLMs) are being adapted for medical reasoning, moving beyond single-step answers to systems that can systematically, transparently, and verifiably reason. It introduces a taxonomy of enhancement techniques, split into training-time methods such as supervised fine‑tuning and reinforcement learning, and test-time methods like prompt engineering and multi‑agent systems. The review examines their application across text, image, and code modalities in key clinical areas—diagnosis, education, and treatment planning—and tracks the shift in evaluation benchmarks from simple accuracy to more nuanced assessments of reasoning quality and visual interpretability.
By Zizhan Ma, Wenxuan Wang, Meidan Ding, Shiyi Zheng, Shengyuan Liu, Jie Liu, Jiaming Ji, Linlin Shen, Yixuan Yuan, Wenting Chen
The paper investigates how on-policy distillation (OPD) and reinforcement learning with verifiable rewards (RLVR) can be combined for post‑training reasoning in large language models. It shows that a two‑stage approach—first applying OPD, then RL—outperforms single‑signal methods and other joint baselines on logic and math reasoning benchmarks. The authors explain this advantage through pass@k analysis, learning dynamics, and parameter updates, concluding that OPD expands solution coverage while RL sharpens performance within that support, and that the OPD validation score is the key trigger for switching to RL.
By Boyan Li, Bingsen Chen, Chenghao Yang, Ping Nie, Chen Zhao, Xi Ye
The paper investigates how different post‑training methods—supervised fine‑tuning, reasoning‑augmented fine‑tuning, and preference optimization (ORPO)—affect the internal computation of refusal behavior in language models. Experiments on Llama‑3.1‑8B, Gemma‑2‑9B, and Qwen3‑8B show that reasoning‑augmented training consistently creates a distinct refusal computation across models, while the architecture influences the internal structure and steerability of refusal. None of the studied methods simultaneously achieve a distributed refusal mechanism, preserve general capability, and allow easy corrective edits, indicating that current post‑training approaches are not a fully reliable defense for safety-critical applications.
By Hoang Cuong Nguyen, Mark Dras, Usman Naseem
The paper introduces Hybrid Search, a method that refines warm-initialized large language model (LLM) based automatic speech recognition (ASR) systems by exploiting interactions between ASR hidden states and the base LLM’s hidden states. By identifying tokens with high semantic dependence and selectively correcting them, the approach surpasses traditional global LLM‑correction techniques such as rescoring and late fusion. The study demonstrates that even after warm initialization, LLM‑based ASR models can further benefit from their base LLM during inference.
By Chan-Jan Hsu, Jaeyeon Kim, Chao-Han Huck Yang, Shinji Watanabe, Hung-yi Lee, Carlos Busso
Para-Pipe is a hierarchical mapping framework that integrates intra- and inter-stage operator parallelism within a pipelined architecture for machine‑learning computational graphs on heterogeneous System‑on‑Chip (SoC) platforms. By selectively fine‑tuning parallelism levels across pipeline stages, it navigates the trade‑off between throughput and latency, reducing inter‑processor communication overhead and improving energy efficiency. Evaluation on Amlogic and Black Sesame SoCs shows multiple Pareto‑optimal configurations, with throughput‑optimized setups achieving up to 11.0% better energy efficiency than purely pipelined strategies and 23.3% better than non‑pipelined parallel execution.
By Yujie Zhang, Huiying Lan, Ehsan Aghapour, Zhiyuan Ning, Peng Zan, Weidong Shao, Anuj Pathania, Tulika Mitra
The paper presents an end‑to‑end sequence‑to‑sequence approach for correcting Tamil spelling and grammar errors, leveraging progressively fine‑tuned transformer models (mT5‑small and mBART‑50). Using a synthetic corpus of 657,720 noisy‑clean sentence pairs across ten error categories, the authors introduce a four‑stage training schedule that targets surface noise, contextual grammar, single‑site sandhi, and multi‑site cross‑word sandhi. The best model, mBART‑50 v5, achieves 69.3% exact‑match accuracy on a balanced diagnostic set, with notable gains in sandhi (87.5%) and subject‑verb agreement (43.5%) accuracy, while also revealing a precision‑recall trade‑off for sandhi corrections.
By Karthikeyan A, Jaya Nirmala S, Sangeetha Sivanesan, Indhu R, Pranav Kumar, Bharat Jude Johnson, Vishnu Ram