The paper investigates how different estimators of the reverse Kullback–Leibler (KL) divergence used as a regularization term in reinforcement learning (RL) training of large language models (LLMs) affect training stability and downstream performance. By analyzing gradient bias across various estimator configurations, the authors demonstrate that biased gradients can cause training instabilities, while unbiased configurations improve performance on both in‑domain and out‑of‑domain tasks. Experiments on Qwen2.5‑7B, Llama‑3.1‑8B‑Instruct, and Qwen3‑4B‑Instruct‑2507 confirm these findings and show that KL regularization also stabilizes off‑policy RL training in asynchronous setups.
By Vedant Shah, Johan Obando-Ceron, Vineet Jain, Brian Bartoldson, Bhavya Kailkhura, Sarthak Mittal, Glen Berseth, Pablo Samuel Castro, Yoshua Bengio, Esmeralda S. Whitammer, Moksh Jain, Siddarth Venkatraman, Aaron Courville
The paper introduces PLUS, a framework that uses reinforcement learning to generate text-based summaries of individual users’ preferences, characteristics, and past conversations. These summaries condition a reward model, allowing it to predict personalized response preferences and improving reward accuracy by 11–77 % over the standard Bradley‑Terry model. PLUS demonstrates robust performance with new users and topics, achieves a 25 % improvement over existing personalized RLHF techniques, and enables zero‑shot personalization for state‑of‑the‑art models like GPT‑4.
By Hyunji Nam, Yanming Wan, Mickel Liu, Peter Ahnn, Jianxun Lian, Natasha Jaques
MyoMechanix is a multimodal dataset and framework for action quality assessment that incorporates muscle activity and other physiological signals alongside visual data. It contains over 7,500 samples of 20 weight‑loaded actions from 38 subjects, with synchronized RGB video, 3D pose, sEMG, and additional signals. The accompanying Fitness Knowledge Graph structures expert annotations into relationships among actions, phases, key steps, errors, and corrective feedback, enabling compositional scoring and interpretable assessment through the CUBIST engine. The project also introduces MyoMechanix‑AQA, MyoMechanix‑VideoQA, and a novel MyoMechanix‑Video2EMG task, demonstrating that multimodal sensing and structured representations improve performance, interpretability, and error attribution.
By Hao Yin, Paritosh Parmar, Lijun Gu, Lin Xu, Tianxiao Guo, Xiujin Liu, Tianyou Zheng, Yang Zhang, Weiwei Fu
The paper introduces a hierarchical multimodal mixture-of-experts (MoE) model for interstitial lung disease (ILD) classification. It combines a frozen, pre‑trained imaging expert with structured electronic health records (EHR) through a two‑stage gating system: a modality‑level gate weights imaging and EHR predictions, while a sub‑gating module further decomposes the EHR branch into clinically defined feature groups with learned, group‑specific contributions. The approach preserves stable imaging representations, allows input‑dependent clinical weighting, and enhances interpretability across anatomical regions, imaging–EHR utilization, and EHR feature groups, achieving the highest mean AUC (0.8750 ± 0.0443) under strict patient‑level cross‑validation.
By Alec K. Peltekian, Gorkem Durak, Halil Ertugrul Aktas, Carrie Lynn Richardson, Mary Carns, Kathleen Aren, GR Scott Budinger, Anthony J. Esposito, Alexander Misharin, Alok Nidhi Choudhary, Ankit Agrawal, Ulas Bagci
The paper presents a lightweight monocular vision pipeline for extracting sidewalk paths on low‑power embedded micromobility platforms. It evolves through three design iterations—from a skeleton‑graph baseline to a distance‑transform corridor planner and finally to a compact image‑space architecture—using a SegFormer‑B0 student model trained with semi‑supervised pseudo‑labels. The final system achieves high segmentation accuracy (IoU 0.946) and fast planning (under 50 ms per frame) while reducing temporal instability and increasing template‑path availability across real campus sequences.
By Lkhanaajav Mijiddorj, Yang Yan, Tyler Beringer, Bilguunzaya Mijiddorj, Alex N. Ho, Bin Xu, Binbin Weng
The paper examines how supervised fine-tuning (SFT) of large language models can leak personally identifiable information (PII) when the fine-tuning data contains user-provided sensitive details. It introduces COVA, a coverage-aware decoding algorithm that improves targeted PII reconstruction from SFT models, especially when an adversary has limited contextual knowledge about a target. Experiments on medical and legal Q&A datasets show that even small proprietary SFT datasets can lead to significant privacy leakage via PII reconstruction.
By Sae Furukawa, Alina Oprea
The paper introduces Functional Entropy, a new uncertainty quantification technique for assessing the functional correctness of code generated by large language models. It evaluates token‑probability and sampling‑based methods across three programming languages and five LLMs, finding that token‑probability approaches generalize well while NLI‑based sampling fails due to semantic clustering. Functional equivalence methods, which replace NLI with an LLM‑based functional assessment, achieve superior AUROC and calibration in most model‑benchmark combinations.
By Dylan Bouchard, Mohit Singh Chauhan, Zeya Ahmad, Ho-Kyeong Ra
The paper introduces continuous adversarial flow models, a continuous-time flow framework trained with an adversarial objective that replaces the fixed mean-squared-error criterion of flow matching. By incorporating a learned discriminator, the method guides training toward a different generalized distribution, yielding samples more closely aligned with the target data distribution. Applied as a post‑training step, it markedly improves ImageNet 256px generation metrics—reducing the guidance‑free FID of latent‑space SiT from 8.26 to 3.63 and of pixel‑space JiT from 7.17 to 3.57—and also enhances guided generation and text‑to‑image benchmarks.
By Shanchuan Lin, Ceyuan Yang, Zhijie Lin, Hao Chen, Haoqi Fan
The paper introduces a new task called Multimodal Unsupervised Continual Post-Training (MU‑CPT), which allows multimodal large language models (MLLMs) to continuously learn from streaming unlabeled data. It identifies token‑level visual dependence (VD) as essential for MU‑CPT, using its structural distortion to detect cross‑modal forgetting and its heterogeneity to guide new‑task learning. The proposed Visual Dependence‑Aware (VDA) framework includes Visually Constrained Optimal Transport (VC‑OT) to mitigate forgetting and Visually Modulated Adaptation (VMA) to enhance new‑task plasticity, achieving a balance between stability and adaptability in MU‑CPT.
By Kaichen Li, Zhilin Zhu, Jianhao Huang, Zhengqin Lai, Baochen Xiong, Zibo Shao, Yaguang Song, Linhui Xiao, Xiaoshan Yang, Changsheng Xu
PANDA (Prototype‑Anchored Data Alignment) is a two‑stage framework that enables a primary‑modality model to benefit from auxiliary modalities even when those modalities are only partially paired or absent at inference. In Stage 1, a shared embedding is learned from the paired subset and class prototypes are estimated from the auxiliary data; in Stage 2, the primary encoder is trained on all subjects using cross‑entropy and alignment to the frozen prototypes. PANDA was evaluated on Alzheimer’s MRI and TCGA‑Lung pathology, achieving significant AUC gains and improved survival prediction while requiring no auxiliary inputs during deployment.
By Sheethal Bhat, Mahfuzur Rahman Chowdhury, Paula Andrea Perez-Toro, Stephan Wunderlich, Rose Dawn Bharat, Siming Bayer, Andreas Maier
The paper introduces Semantic Boundary Predictor (SBP), an inference‑time framework that improves demographic fairness in synthetic face generation by applying a single, one‑shot intervention during reverse denoising. SBP learns linear semantic boundaries from late‑stage latent representations and applies them only at the initial noisy latent, leaving the rest of the diffusion process unchanged. Experiments on CelebA‑HQ show significant reductions in fairness disparity—98% for gender, 95% for binary race, and 15% for four‑class race—while preserving image quality across demographic groups.
By Subir Kumar Parida, Rajbabu Velmurugan, Ketan Kotwal, R. S. Sengar, Swati Hiremath
The paper introduces Socialized Detector Learning (SDL), a framework for improving heterogeneous, category-specialized object detectors by exchanging knowledge within a detector society. It proposes Trajectory-Guided and Reciprocal Distillation (TGRD), which estimates transfer difficulty, constructs a carrier trajectory, and performs progressive consolidation and reciprocal transfer. Experiments on MS COCO show that carriers outperform simultaneous aggregation by 2.6 AP, and reciprocal detectors achieve high AP on previously unsupported categories while staying close to expert performance.
By Weihao Li, Yunqi Zhu, Zhihe Fan, Ruipu Zhao, Boan Tao, Xinjie Yao, Yan Fan, Pengfei Zhu
TDFNet introduces a Tri-projection Deformable Fusion Network that uses equirectangular, cube map, and tangent projections to mitigate geometric distortions in panoramic salient object detection. It incorporates a cross-projection deformable attention module for geometry-aware sampling and a latitude-guided fusion module that balances ERP and CMP features using spherical latitude priors. The network’s three-branch encoding preserves global continuity, local detail, and boundary precision, improving detection performance over existing projection-based methods.
By Qiangqiang Zhou, Jiacong Yu, Jiawei Xu, Yong Chen, Xin Huang, Ping Li
The paper surveys Multi‑Modal Anomaly Detection (MMAD), a field that identifies rare abnormal events across heterogeneous data sources used in safety‑critical domains like industrial inspection and cybersecurity. It formalizes MMAD, outlines five core characteristics, and categorizes existing methods into normality‑assumption and anomaly‑assumption paradigms, highlighting how foundation models are reshaping the field. The survey also compiles benchmarks, evaluation protocols, and identifies open problems for developing robust, adaptive, and interpretable MMAD systems.
By Xudong Mou, Zexin Wu, Chuan Luo, Shiru Chen, Xudong Liu, Chunming Hu, Renyu Yang
The paper investigates how adversarial examples transfer between client models in federated learning and explores the relationship between these examples and client data distributions. It proposes a defense strategy based on adversarial training that leverages the transferability of model robustness. Experiments on real-life datasets demonstrate that the new attack and defense methods outperform existing state‑of‑the‑art approaches.
By Zuobin Xiong, Deval Mukherjee, Homook Cho, Wei Li
The paper applies sparse autoencoders to a neutrino foundation model trained on IceCube data, uncovering a validated atlas of physical concepts within the model’s internal representation. Causal analysis shows the direction reconstruction head largely ignores this atlas, whereas an uncertainty head trained on the same representation effectively uses quality and brightness features, improving angular resolution from 20.2° to 3.2° at 20% efficiency. These findings demonstrate that mechanistic interpretability can expose latent physics and guide the design of downstream tasks.
By Rapha\"el Bonnet-Guerrini, Johann Ioannou-Nikolaides, Inar Timiryasov, Vincenzo Piuri
The paper examines how large language models (LLMs) can be biased by irrelevant social contexts when evaluating teachers, using a large U.S. classroom transcript dataset. It shows that spurious contexts can shift model ratings by up to 1.48 points on a 7‑point scale and that standard mitigation methods like SFT and DPO are insufficient. The authors introduce Debiasing‑DPO, a method that combines contrastive reasoning‑augmented DPO with SFT, which reduces bias by 84% and improves predictive accuracy by 52% on Llama and Qwen Instruct models.
By Hyunji Nam, Dorottya Demszky
RefLAM is a pipeline that converts manuscript page images and clean transcriptions into validated, line-level ground truth for Arabic handwritten text recognition. It combines a deep‑learning page‑segmentation model, a multimodal large language model for structured OCR, and a diacritic‑agnostic fuzzy alignment engine that assigns a confidence score to each line, with a provable correctness guarantee for perfect scores. Using RefLAM, the authors achieved a 75× speedup over manual annotation and released AraMS‑28k, a dataset of 14 historical Arabic manuscripts with detailed annotations.
By Mohamed Guechaoui, Mohamed Diaa Zellagui, Souleyman Chaib, Sahraoui Dhelim
The paper introduces GRIP, an algorithm‑agnostic framework for machine unlearning in Mixture‑of‑Experts large language models. GRIP enforces hard geometric constraints on router updates, projecting gradient changes into the null space of the retain set’s routing matrix to prevent routing manipulation. Two variants—training‑time stochastic projection and post‑training analytical correction—show significant improvements in routing stability, retain accuracy, and resistance to white‑box adversarial recovery across two MoE models.
By Andy Zhu, Rongzhe Wei, Yupu Gu, Pan Li
Physics‑Informed Neural Networks (PINNs) often fail on stiff or advection‑dominated partial differential equations. This study compares two proposed fixes—switching from FP32 to FP64 precision to address an L‑BFGS stopping artifact, and replacing the MLP with a state‑space‑model (SSM) backbone plus sub‑sequence alignment to mitigate architectural simplicity bias—using a pre‑registered, seed‑paired 144‑run experiment across convection, reaction, and wave problems, plus an independent 85‑run study. The results show that the remedies act on disjoint subsets of regimes and seeds: precision changes affect some seeds in opposite directions, alignment improves success in hard convection cases, and the SSM backbone alone succeeds on many reaction seeds, but none of the remedies fully substitutes for the other, and all must be evaluated jointly and reported per seed.
By Jinyuan Zhang, Peng He, He Hu, Yin Yuan, ShengShuo Jiao