Isolater - Feed

Ax Isaac Han, Sangyeon Park, Seungwon Oh, Donghu Kim, Hojoon Lee, Kyung-Joong Kim 3/3/2026

FIRE: Frobenius-Isometry Reinitialization for Balancing the Stability-Plasticity Tradeoff

arXiv research on FIRE reinitialization method balancing stability and plasticity in neural networks trained on nonstationary data.

Ax Shuo He, Lang Feng, Xin Cheng, Lei Feng, Bo An 3/3/2026

Online Causal Kalman Filtering for Stable and Effective Policy Optimization

arXiv paper on Online Causal Kalman Filtering for stabilizing policy optimization in LLM reinforcement learning with high-variance importance sampling.

Ax Danlong Yuan, Wei Wu, Zhengren Wang, Xueliang Zhao, Huishuai Zhang, Dongyan Zhao 3/3/2026

SWE-MiniSandbox: Container-Free Reinforcement Learning for Building Software Engineering Agents

arXiv research on SWE-MiniSandbox: container-free reinforcement learning method for training software engineering agents at scale.

Ax MiniCPM Team, Wenhao An, Yingfa Chen, Yewei Fang, Jiayi Li, Xin Li, Yaohui Li, Yishan Li, Yuxuan Li, Biyuan Lin, Chuan Liu, Hezi Liu, Siyuan Liu, Hongya Lyu, Yinxu Pan, Shixin Ren, Xingyu Shen, Zhou Su, Haojun Sun, Yangang Sun, Zhen Leng Thai, Xin Tian, Rui Wang, Xiaorong Wang, Yudong Wang, Bo Wu, Xiaoyue Xu, Dong Xu, Shuaikang Xue, Jiawei Yang, Bowen Zhang, Jinqian Zhang, Letian Zhang, Shengnan Zhang, Xinyu Zhang, Xinyuan Zhang, Zhu Zhang, Hengyu Zhao, Jiacheng Zhao, Zhi Zheng, Jie Zhou, Zihan Zhou, Shuo Wang, Chaojun Xiao, Xu Han, Zhiyuan Liu, Maosong Sun 3/3/2026

MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling

arXiv paper on MiniCPM-SALA: hybrid sparse and linear attention mechanism for 9B LLM enabling efficient ultra-long context modeling.

Ax Pengxiang Zhao, Hui-Ling Zhen, Xing Li, Han Bao, Weizhe Lin, Zhiyuan Yang, Manyi Zhang, Yuanyong Luo, Ziwei Yu, Xin Wang, Mingxuan Yuan, Xianzhi Yu, Zhenhua Dong 3/3/2026

Unleashing Low-Bit Inference on Ascend NPUs: A Comprehensive Evaluation of HiFloat Formats

arXiv research on low-bit floating-point formats (HiFloat) for efficient LLM inference on Ascend NPUs with rigorous performance evaluation.

Ax Stephan Vonschallen, Dominique Oberle, Theresa Schmiedel, Friederike Eyssel 3/3/2026

Knowledge-Based Design Requirements for Generative Social Robots in Higher Education

Framework for designing generative social robots in education, addressing hallucinations and safety concerns of LLM-powered tutoring systems.

Ax Jing Zhao, Ting Zhen, Junwei Bao, Hongfei Jiang, Yang Song 3/3/2026

Elo-Evolve: A Co-evolutionary Framework for Language Model Alignment

Elo-Evolve framework for LLM alignment using co-evolutionary multi-agent competition instead of static reward functions, improving training stability.

Ax Shaobin Zhuang, Yuang Ai, Jiaming Han, Weijia Mao, Xiaohui Li, Fangyikang Wang, Xiao Wang, Yan Li, Shanchuan Lin, Kun Xu, Zhenheng Yang, Huaibo Huang, Xiangyu Yue, Hao Chen, Yali Wang 3/3/2026

UniWeTok: An Unified Binary Tokenizer with Codebook Size $\mathit{2^{128}}$ for Unified Multimodal Large Language Model

UniWeTok: unified binary tokenizer for multimodal LLMs enabling high-fidelity visual reconstruction and semantic extraction with 2^128 codebook size.

Ax Jonathan Lys, Vincent Gripon, Bastien Pasdeloup, Axel Marmoret, Lukas Mauch, Fabien Cardinaux, Ghouthi Boukli Hacene 3/3/2026

Inner Loop Inference for Pretrained Transformers: Unlocking Latent Capabilities Without Training

Research on iterative refinement in Transformer architectures using inner loop inference to improve model capabilities without additional training.

Ax Jonathan Lys, Vincent Gripon, Bastien Pasdeloup, Axel Marmoret, Lukas Mauch, Fabien Cardinaux, Ghouthi Boukli Hacene 3/3/2026

Residual Connections and the Causal Shift: Uncovering a Structural Misalignment in Transformers

Research paper on structural misalignment in Transformers between residual connections (tied to current token) and causal supervision (targets next token).

Ax Ji Li, Bo Wang, Jing Xia, Mingyi Li, Shiyan Hu 3/3/2026

HIMM: Human-Inspired Long-Term Memory Modeling for Embodied Exploration and Question Answering

HIMM proposes non-parametric memory framework for embodied agents using MLLMs, disentangling episodic and semantic memory for long-horizon exploration.

Ax Md. Najib Hasan, Touseef Hasan, Souvika Sarkar 3/3/2026

Are LLMs Ready to Replace Bangla Annotators?

Study evaluates LLMs as zero-shot annotators for Bangla hate speech detection, examining bias and reliability in low-resource identity-sensitive annotation tasks.

Ax Guy Bar-Shalom, Ami Tavory, Itay Evron, Maya Bechler-Speicher, Ido Guy, Haggai Maron 3/3/2026

A Graph Meta-Network for Learning on Kolmogorov-Arnold Networks

Graph Meta-Network learns on Kolmogorov-Arnold networks using weight-space models to predict neural network accuracy on new datasets.

Ax Luke J. Huang, Zhuoyang Zhang, Qinghao Hu, Shang Yang, Song Han 3/3/2026

Stable Asynchrony: Variance-Controlled Off-Policy RL for LLMs

Stable Asynchrony proposes variance-controlled off-policy RL for LLM post-training that handles stale rollouts and heavy-tailed importance weights in asynchronous training.

Ax Bin Wang, Fan Wang, Pingping Wang, Jinyu Cong, Yang Yu, Yilong Yin, Zhongyi Han, Benzheng Wei 3/3/2026

Agentic Unlearning: When LLM Agent Meets Machine Unlearning

Agentic Unlearning removes sensitive information from LLM agent parameters and persistent memory in closed-loop systems, addressing parameter-memory backflow issues.

Ax Daniel Romero-Alvarado, Fernando Mart\'inez-Plumed, Lorenzo Pacchiardi, Hugo Save, Siddhesh Milind Pawar, Behzad Mehrbakhsh, Pablo Antonio Moreno Casares, Ben Slater, Paolo Bova, Peter Romero, Zachary R. Tidler, Jonathan Prunty, Luning Sun, Jose Hernandez-Orallo 3/3/2026

Capabilities Ain't All You Need: Measuring Propensities in AI

Framework measuring model propensities and behavioral tendencies alongside capabilities using Item Response Theory, relevant for safety evaluation.

Ax Wei Chen, Junle Chen, Yuqian Wu, Yuxuan Liang, Xiaofang Zhou 3/3/2026

Learning from Complexity: Exploring Dynamic Sample Pruning of Spatio-Temporal Training

Dynamic sample pruning method for spatio-temporal model training to reduce computational overhead on redundant large-scale datasets.

Ax Yousef Emami, Hao Zhou, Radha Reddy, Atefeh Hajijamali Arani, Biliang Wang, Kai Li, Luis Almeida, Zhu Han 3/3/2026

Large Language Model-Assisted UAV Operations and Communications: A Multifaceted Survey and Tutorial

Survey of LLM integration with UAV systems for environmental understanding, swarm coordination, and task planning applications.

Ax Yungang Yi, Weihua Li, Matthew Kuo, Catherine Shi, Quan Bai 3/3/2026

Depth-Structured Music Recurrence: Budgeted Recurrent Attention for Full-Piece Symbolic Music Modeling

Architecture using stateful recurrent attention for full-piece symbolic music modeling with constrained computational resources.

Ax Zanxi Ruan, Songqun Gao, Qiuyu Kong, Yiming Wang, Marco Cristani 3/3/2026

StructXLIP: Enhancing Vision-language Models with Multimodal Structural Cues

Fine-tuning approach enhancing vision-language models by aligning structural/edge-based cues across modalities for improved cross-modal retrieval.

Ax Chenyue Yu, Lingao Xiao, Jinhong Deng, Ivor W. Tsang, Yang He 3/3/2026

Dataset Color Quantization: A Training-Oriented Framework for Dataset-Level Compression

Dataset compression framework using color quantization to reduce image dataset storage demands while preserving model training effectiveness.

Ax Wei Chen, Yuqian Wu, Junle Chen, Xiaofang Zhou, Yuxuan Liang 3/3/2026

UrbanFM: Scaling Urban Spatio-Temporal Foundation Models

Foundation model for urban spatio-temporal forecasting addressing scenario-specificity in city-scale prediction tasks across multiple regions.

Ax Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, Anku Jaiswal 3/3/2026

Mitigating Structural Noise in Low-Resource S2TT: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration

Pipeline for low-resource speech-to-text translation with punctuation restoration to mitigate ASR structural noise in Nepali-English systems.

Ax Evangelia Christakopoulou, Vivekkumar Patel, Hemanth Velaga, Sandip Gaikwad 3/3/2026

Scaling Search Relevance: Augmenting App Store Ranking with LLM-Generated Judgments

Approach using LLM-generated textual relevance judgments to augment behavioral signals in large-scale app store search ranking systems.

Ax Hiroshi Sasaki 3/3/2026

Pseudo Contrastive Learning for Diagram Comprehension in Multimodal Models

Training paradigm using pseudo-contrastive learning to improve diagram comprehension and fine-grained visual understanding in multimodal models.

Ax Yongzhong Xu 3/3/2026

Optimizer-Induced Low-Dimensional Drift and Transverse Dynamics in Transformer Training

Analysis of transformer training dynamics under AdamW, identifying low-dimensional drift directions that capture 60-80% of parameter displacement.

Ax Zhicheng Fang, Jingjie Zheng, Chenxu Fu, Wei Xu 3/3/2026

Jailbreak Foundry: From Papers to Runnable Attacks for Reproducible Benchmarking

System translating jailbreak papers into executable modules via multi-agent workflow for reproducible benchmarking and evaluation of LLM robustness.

Ax Jintao Zhang, Zirui Liu, Mingyue Cheng, Xianquan Wang, Zhiding Liu, Qi Liu 3/3/2026

StaTS: Spectral Trajectory Schedule Learning for Adaptive Time Series Forecasting with Frequency Guided Denoiser

Diffusion model approach for time series forecasting using learnable spectral trajectory schedules to improve noise handling and structure recovery.

Ax Jitian Zhao, Changho Shin, Tzu-Heng Huang, Satya Sai Srinath Namburi GNVV, Frederic Sala 3/3/2026

CARE: Confounder-Aware Aggregation for Reliable LLM Evaluation

Method for improving LLM-as-a-judge evaluation by accounting for correlated errors and latent confounders in ensemble aggregation mechanisms.

Ax Peiyuan Zhang, Matthew Noto, Wenxuan Tan, Chengquan Jiang, Will Lin, Wei Zhou, Hao Zhang 3/3/2026

Attn-QAT: 4-Bit Attention With Quantization-Aware Training

Research on 4-bit quantization-aware training for attention mechanisms to enable FP4 computation on emerging GPUs, addressing challenges with dynamic range and heavy-tailed activations.

Ax Banseok Lee, Youngmin Kim 3/3/2026

Maximizing the Spectral Energy Gain in Sub-1-Bit LLMs via Latent Geometry Alignment

Addresses extreme LLM compression by identifying spectral energy gain in binary quantization and fixing latent geometry misalignment issues.

Ax Minghao Han, Lixian Zhang, Chenliang Liu, Zhipeng Zhou, Jun Wang, Wei Pan 3/3/2026

Reinforcement Learning for Control with Probabilistic Stability Guarantee: A Finite-Sample Approach

Novel RL approach for control systems with probabilistic stability guarantees derived from Lyapunov methods using finite trajectory samples.

Ax Sicong Che, Jiayi Yang, Sarfraz Khurshid, Wenxi Wang 3/3/2026

Property-Driven Evaluation of GNN Expressiveness at Scale: Datasets, Framework, and Study

Property-driven evaluation framework for assessing GNN expressiveness using formal specification and systematic testing across large-scale datasets.

Ax Ian Li, Zilei Shao, Benjie Wang, Rose Yu, Guy Van den Broeck, Anji Liu 3/3/2026

Breaking the Factorization Barrier in Diffusion Language Models

Proposes method to overcome factorization barrier in diffusion language models, enabling faster parallel token generation without sacrificing coherence.

Ax Chenwei Wu, Zitao Shuai, Liyue Shen 3/3/2026

REMIND: Rethinking Medical High-Modality Learning under Missingness--A Long-Tailed Distribution Perspective

REMIND addresses multi-modal learning with missing modalities in medical AI, treating missingness as a long-tailed distribution problem.

Ax Yongchao Huang 3/3/2026

BiJEPA: Bi-directional Joint Embedding Predictive Architecture for Symmetric Representation Learning

BiJEPA extends JEPA with bi-directional prediction for self-supervised learning, leveraging both forward and inverse relationships in representation learning.

Ax Bingran Wang, Seongha Jeong, Sebastiaan P. C. van Schie, Dongyeon Han, Jaeho Min, John T. Hwang 3/3/2026

Knowledge-guided generative surrogate modeling for high-dimensional design optimization under scarce data

Knowledge-guided generative surrogate modeling combines data-driven models with domain expert knowledge for design optimization under data scarcity.

Ax Jiaang Li, Haibin Chen, Langming Liu, Yujin Yuan, Yadao Wang, Yizhen Zhang, Chengting Yu, Xin Tong, Weidong Zhang, Shilei Liu, Wenbo Su, Bo Zheng 3/3/2026

Expert Divergence Learning for MoE-based Language Models

Expert Divergence Learning addresses expert homogenization in MoE language models by encouraging functional specialization through auxiliary loss during pre-training.

Ax Chao Huang, Yanhui Li, Yunkang Cao, Wei Wang, Hongxi Huang, Jie Wen, Wenqi Ren, Xiaochun Cao 3/3/2026

M3-AD: Reflection-aware Multi-modal, Multi-category, and Multi-dimensional Benchmark and Framework for Industrial Anomaly Detection

M3-AD is a multimodal framework for industrial anomaly detection using MLLMs with reflection-aware self-correction mechanisms for fine-grained scenarios.

Ax Maitri Krishna Sai 3/3/2026