The study evaluates multi-harness reinforcement learning (RL) for coding agents by training two Qwen3-8B RL variants—one with rewards normalized within each harness and one across harnesses—using identical replay data. Results show modest gains over the supervised fine-tuning baseline (+0.48 and +0.77 percentage points for within and cross normalization, respectively), but no consistent advantage of cross normalization on unseen harnesses. Performance differences across harnesses are much larger than the training gains, indicating that harness dependence remains a significant challenge.
By Chenqian Le, Jiayi Cheng, Qijia He, Runhao Li, Yinghao Li, Xupeng Chen
JMT-YOLO is a new joint multi-task learning framework for domain adaptive object detection that builds on the Mean Teacher architecture. It adds an auxiliary semantic segmentation task and a masking mechanism to better model foreground–background boundaries and improve instance-level alignment. Experiments on Pascal VOC→Clipart1k, Cityscapes→Foggy Cityscapes, and KITTI→Cityscapes show that JMT‑YOLO outperforms existing state‑of‑the‑art methods, achieving 46.9%, 57.3%, and 61.0% detection performance respectively.
By Zhenhua Yu, Jiahao Zhang, Wenchao Zhang, Ou Ye, Hengxiang Zhao
The paper examines how fine‑tuned defenses against prompt injection in large language models can be audited before deployment. By leveraging intermediate checkpoints from the fine‑tuning process, the authors create strong optimization‑based attacks (e.g., GCG and BEAST) that dramatically increase attack success rates—from near zero to up to 96%—against state‑of‑the‑art defenses. This demonstrates that current empirical evaluations often overestimate robustness and that pre‑deployment stress‑testing can uncover real vulnerabilities.
By Xiaoxue Yang, Bozhidar Stevanoski, Matthieu Meeus, Yves-Alexandre de Montjoye
The paper introduces Multi-Modal Typography, a systematic investigation into how perturbations applied to speech and on-screen text can compromise audio‑visual reasoning in large language models. Experiments on the WorldSense benchmark show that such attacks can raise the attack success rate against Gemini 3.7 Flash from 9.53% to 43.41%, and fine‑tuning can reduce the success rate for Qwen2.5‑Omni‑7B from 60.6% to 52.0%. The study highlights that cross‑modal typography attacks are a significant, yet underexplored, vulnerability in multi‑modal reasoning systems.
By Tianle Chen, Deepti Ghadiyaram
The paper introduces LHSDet, a detector for AI‑generated images that treats detection as a visual question answering task. It employs a triple‑branch architecture: a low‑level visual branch for local texture cues, a high‑level visual branch using SigLIP2 for global perception, and a semantic‑level textual branch that generates captions with BLIP‑2. Experiments show that LHSDet achieves strong accuracy and robustness across various generative models.
By Qian Yao, Luming Yang, Jun-Jie Huang, Yongjun Wang, Xinwang Liu, Meng Wang
DeFM is a self‑supervised foundation model trained solely on depth images for robotics. It uses a DINO‑style self‑distillation objective on 60 M curated depth images to learn geometric and semantic representations that generalize across environments, tasks, and sensors. The model incorporates a novel input normalization strategy for metric awareness, can be distilled into compact versions for resource‑constrained systems, and achieves state‑of‑the‑art performance on depth‑based classification, segmentation, navigation, locomotion, and manipulation benchmarks, with strong sim‑to‑real transfer.
By Manthan Patel, Jonas Frey, Mayank Mittal, Fan Yang, Alexander Hansson, Amir Bar, Cesar Cadena, Marco Hutter
Optimizable Model Averaging (OMA) is a black‑box method that assigns nonnegative weights to each candidate model’s predictive probabilities, keeping all model parameters frozen. By minimizing regularized predictive entropy, OMA learns a single scalar coefficient per model to average outputs from heterogeneous architectures. The authors show that appropriate regularization reduces expected calibration error and improves the trade‑off between performance and calibration compared to simple output averaging.
By Mijung Park
StereoEngine introduces two methods for generating realistic synthetic stereo data for robotic manipulation. The first method warps a monocular video into a right view using depth estimation and a video diffusion model to inpaint disoccluded regions, while the second method renders stereo pairs directly in simulation. These approaches outperform existing baselines in stereo quality and enable improved bimanual manipulation policies when used for mid‑training or task‑specific fine‑tuning.
By Jason Chen, I-Chun Arthur Liu, Gaurav Sukhatme, Daniel Seita
BLEND is a framework for federated vision‑language models that balances personalization and generalization. It uses selective fine‑tuning with global vision and text adapters shared across clients and a personalized vision projection adapter for each client. A novel loss function blends outputs from global and personalized adapters, while an anchor design in the projection heads promotes generalization, leading to superior performance over existing baselines.
By Payam Abdisarabshali, Fardis Nadimi, Kasra Borazjani, Naji Khosravan, Richeng Jin, Seyyedali Hosseinalipour
Vision2Code is a new benchmark for image-to-code generation that covers 15 domains and 539 test examples without relying on reference source code. It evaluates generated programs by executing them and comparing rendered outputs to the original images using domain‑specific rubrics, yielding scores that correlate better with human judgments than generic metrics. The study shows domain variability in model performance and demonstrates that the scoring method can improve fine‑tuning of models like Qwen3.5‑9B and Phi‑4‑MM via rejection sampling.
By Ajay Vikram Periasami, Junlin Wang, Bhuwan Dhingra
The paper investigates whether modern video‑language models can reliably understand temporal dynamics in driving scenarios. Using controlled ego‑motion tasks, the authors find that various prompting strategies are inconsistent and heavily biased, while newer models only show modest few‑shot improvements. However, lightweight LoRA fine‑tuning and probes on intermediate visual features can recover strong motion discrimination, suggesting that motion cues exist within the models but are not effectively leveraged by the final video‑to‑language pathway.
By Shiva Sreeram, Erfan Aasi, Tsun-Hsuan Wang, Alaa Maalouf, Daniela Rus
The paper investigates how the operational value of deep‑learning PV disaggregation depends on deployment regimes—whether the model serves known customers, new customers, future periods, or new regions. Using Ausgrid smart‑meter data and cross‑dataset adaptation to Pecan Street, the authors benchmark several deep sequence architectures and find that deployment regime explains more error variation than model choice. Fine‑tuning on a small number of target‑region households can significantly reduce transfer error, and the benefits of long input windows vary by use case.
By Jack Zheng, Hao Wang
The paper presents a comprehensive framework for comparing parameter‑efficient fine‑tuning (PEFT) methods by treating each as a smooth map from a small parameter space into weight space, with a pretrained checkpoint as the base point. It identifies three key invariants—image near the base point, fibre dynamics, and base change—that explain expressivity, training dynamics, and initialization behavior. The authors prove numerous results about common PEFT techniques (e.g., LoRA, Householder adaptation) and classify 382 methods from 138 papers using seven coordinate descriptors.
By Xiaoyu Li, Zhizhou Sha, Chiwun Yang, Dai Shi
The paper introduces Cross-Behavior Policy Optimization (CBPO), a method that unifies offline demonstrations and policy-generated rollouts into a single regression loss using a fixed-reference target. CBPO eliminates the need for separate supervised fine‑tuning and reinforcement learning objectives, allowing positive and negative offline completions, reference responses, and current and historical rollouts to share a common reward‑labeled format and replay buffer. Experiments on two model scales and four mathematical reasoning benchmarks show that CBPO, especially the CBPO‑Unified variant, outperforms baseline methods in avg@32 and pass@32 metrics, with historical rollouts providing additional gains.
By Du Cheng, Junyu Lu, Jingfan Zhang, Yuzhuo Zhang, Haoxuan Jia, Shichao Weng
The paper introduces Prompteina, an all‑atom protein structure generation model that accepts natural‑language descriptions of desired structural and functional properties, similar to text‑to‑image systems. It is trained on a large paired dataset of 1.8 million protein structures and over 50 million captions, and fine‑tuned from the Proteina‑Complexa model to align residues with textual cues. Prompteina outperforms existing baselines on most structural metrics, improves binder design performance, and experimentally validates nine new binders for the GM2A target.
By Joseph Lee, Yanjing Li, Bo Qiang, Gregory Lau, Allyson Ettinger, Noah A. Smith, Rory Donovan-Maiye, David Baker, Hanna Hajishirzi, Pang Wei Koh
Zebra3D introduces a layerwise sparse‑linear hybrid attention framework that enhances the efficiency of pretrained 3D foundation models without requiring training from scratch. By assigning linear‑centered attention to tolerant early stages and tile‑sparse attention to sensitive middle‑to‑late stages, it preserves context through landmark summaries and enables training‑free conversion via closed‑form affine fitting. Across multiple backbones, Zebra3D achieves strong reconstruction and camera‑pose accuracy with high attention sparsity, delivering up to 7.1× speedups at 1,024 views and maintaining performance close to dense counterparts after optional post‑conversion training.
By Utkarsh Saxena, Ruijing Yang, Mehdi Rezagholizadeh, Akash Haridas, Vikram Appia, Emad Barsoum
The paper introduces Golden Anchor Learning (GAL), a framework that optimizes text-to-image (T2I) generators using feedback from downstream image‑to‑video (I2V) models to produce ‘golden anchors’ that improve video quality. Experiments show that GAL consistently raises VBench scores by 2.97%–8.46% and enhances motion quality across four I2V models, even when trained only on HunyuanVideo‑1.5 renders.
By Jinyuan Xu, Shitong Shao, Lichen Bai, Zikai Zhou, Guoxi Huang, Nantheera Anantrasirichai, Zeke Xie
The paper "One Thing Leads to Another: The Essentials of Time Series Causal Discovery" offers a unified tutorial on detecting causation between time series, tracing its origins from signal processing and outlining its evolution across statistics, dynamical systems, and machine learning. It reviews key principles, limitations, assumptions, and evaluation methods of modern discovery techniques, and highlights their growing importance in fields such as agriculture, Earth science, healthcare, and AI. A companion repository provides code packages and resources to support readers.
By Kurt Butler, Damian Machlanski, Michael P. J. Camilleri, Zhihua Liu, Panagiotis Dimitrakopoulos, Feng Chen, Ricardo Silva, Sotirios A. Tsaftaris
RARA (Retrieval‑Anchored Representation Alignment) is a two‑stage method for fine‑tuning pre‑trained robot policies on new tasks with few demonstrations. It first retrieves and aligns representations from pre‑training that match the target task’s intent and motion, then fine‑tunes while constraining changes to those retrieved representations. Experiments on Diffusion Policy and GR00T across 13 tasks in simulation and on a real robot show RARA improves in‑distribution success by 9.5 points and out‑of‑distribution success by 13.6 points, outperforming all baselines.
By Xuhui Kang, Sung-Wook Lee, Yan Zhuang, Tianle Zhong, Yen-Ling Kuo
PPG-Bench is an open‑source benchmark that evaluates five publicly available photoplethysmography (PPG) foundation models across nine datasets and 26 tasks, covering cardiovascular state estimation, activity, sleep, screening, and identification. The benchmark tests models under four downstream regimes—linear probing, random forests, frozen MLP probes, and end‑to‑end finetuning—while also comparing them to three controls: a random/constant baseline, a deterministic pulse‑rate‑variability feature set, and a generic time‑series foundation model (MOMENT). Results show that over a third of task‑regime combinations yield no advantage for any PPG foundation model over the baseline, and in most cases no PPG‑specific model outperforms the simple controls, indicating that current PPG foundation models lack generalised, transferable representations.
By Arvasu Kulkarni, Mahir Jain, Aditya Ray Mishra, Sandeep Singh, Siddharth Panwar