Isolater - Feed

Ax Yaorui Shi, Yuxin Chen, Siyuan Wang, Sihang Li, Hengxing Cai, Qi Gu, Xiang Wang, An Zhang 3/3/2026

Look Back to Reason Forward: Revisitable Memory for Long-Context LLM Agents

Revisitable memory mechanism enabling long-context LLM agents to reference and reason over dispersed evidence across millions of tokens.

Ax Jonas Ngnaw\'e, Maxime Heuillet, Sabyasachi Sahoo, Yann Pequignot, Ola Ahmad, Audrey Durand, Fr\'ed\'eric Precioso, Christian Gagn\'e 3/3/2026

Robust Fine-Tuning from Non-Robust Pretrained Models: Mitigating Suboptimal Transfer With Epsilon-Scheduling

Epsilon-scheduling technique for robust fine-tuning from non-robust pretrained models addressing adversarial robustness transfer.

Ax Sebastian Bordt, Martin Pawelczyk 3/3/2026

Train Once, Answer All: Many Pretraining Experiments for the Cost of One

Method conducting multiple controlled pretraining experiments simultaneously in single training run to reduce computational costs.

Ax Chaorui Yao, Yanxi Chen, Yuchang Sun, Yushuo Chen, Wenhao Zhang, Xuchen Pan, Yaliang Li, Bolin Ding 3/3/2026

Group-Relative REINFORCE Is Secretly an Off-Policy Algorithm: Demystifying Some Myths About GRPO and Its Friends

Analysis showing Group Relative REINFORCE exhibits off-policy properties, demystifying GRPO and related LLM RL algorithms.

Ax Xiaohe Bo, Rui Li, Zexu Sun, Quanyu Dai, Zeyu Zhang, Zihang Tian, Xu Chen, Zhenhua Dong 3/3/2026

Prompt and Parameter Co-Optimization for Large Language Models

Joint optimization of prompts and parameters for LLMs exploring synergistic effects between prompt engineering and fine-tuning.

Ax Gyuhyeon Seo, Jungwoo Yang, Junseong Pyo, Nalim Kim, Jonggeun Lee, Yohan Jo 3/3/2026

SimuHome: A Temporal- and Environment-Aware Benchmark for Smart Home LLM Agents

SimuHome benchmark with 600 episodes for evaluating LLM-based smart home agents on temporal and environment-aware task execution.

Ax Jitai Hao, Hao Liu, Xinyan Xiao, Qiang Huang, Jun Yu 3/3/2026

Uni-X: Mitigating Modality Conflict with a Two-End-Separated Architecture for Unified Multimodal Models

Uni-X architecture mitigating gradient conflicts between vision and text modalities in unified multimodal models.

Ax Haijier Chen, Bo Xu, Shoujian Zhang, Haoze Liu, Jiaxuan Lin, Jingrong Wang 3/3/2026

Vid-LLM: A Compact Video-based 3D Multimodal LLM with Reconstruction-Reasoning Synergy

Vid-LLM video-based 3D multimodal LLM combining 3D reconstruction and reasoning capabilities for 3D scene understanding.

Ax Shane Bergsma, Bin Claire Zhang, Nolan Dey, Shaheer Muhammad, Gurpreet Gosal, Joel Hestness 3/3/2026

Scaling with Collapse: Efficient and Predictable Training of LLM Families

Study of loss curve collapse phenomenon during LLM family training under practical scaling conditions for predictable scaling laws.

Ax Haolei Xu, Xinyu Mei, Yuchen Yan, Rui Zhou, Wenqi Zhang, Weiming Lu, Yueting Zhuang, Yongliang Shen 3/3/2026

EasySteer: A Unified Framework for High-Performance and Extensible LLM Steering

EasySteer unified framework for efficient and extensible LLM steering through hidden state manipulation at inference time.

Ax Yuyou Zhang, Radu Corcodel, Chiori Hori, Anoop Cherian, Ding Zhao 3/3/2026

SpinBench: Perspective and Rotation as a Lens on Spatial Reasoning in VLMs

SpinBench diagnostic benchmark evaluating spatial reasoning and perspective-taking capabilities in vision language models.

Ax Victor Wang, Elias Stengel-Eskin 3/3/2026

Calibrating Verbalized Confidence with Self-Generated Distractors

Method to improve LLM confidence calibration using self-generated distractors to reduce overconfidence and improve safety.

Ax Yeongmin Kim, Donghyeok Shin, Mina Kang, Byeonghu Na, Il-Chul Moon 3/3/2026

Distillation of Large Language Models via Concrete Score Matching

Knowledge distillation method for efficient LLM inference using concrete score matching to preserve logit information beyond softmax smoothing.

Ax Ruiyang Wang, Hao-Lun Hsu, David Hunt, Jiwoo Kim, Shaocheng Luo, Miroslav Pajic 3/3/2026

COMRES-VLM: Coordinated Multi-Robot Exploration and Search using Vision Language Models

COMRES-VLM framework using vision language models for coordinated multi-robot exploration and object search in unknown indoor environments.

Ax Keming Wu, Sicong Jiang, Max Ku, Ping Nie, Minghao Liu, Wenhu Chen 3/3/2026

EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing

EditReward human-aligned reward model for scaling synthetic training data in open-source instruction-guided image editing models.

Ax Guanxi Lu, Hao Mark Chen, Yuto Karashima, Zhican Wang, Daichi Fujiki, Hongxiang Fan 3/3/2026

AdaBlock-dLLM: Semantic-Aware Diffusion LLM Inference via Adaptive Block Size

AdaBlock-dLLM improves diffusion-based LLM inference efficiency through adaptive block-size selection for semi-autoregressive decoding.

Ax Chenxi Whitehouse, Sebastian Ruder, Tony Lin, Oksana Kurylo, Haruka Takagi, Janice Lam, Nicol\`o Busetto, Denise Diaz, Francisco Guzm\'an 3/3/2026

MENLO: From Preferences to Proficiency -- Evaluating and Modeling Native-like Quality Across 47 Languages

MENLO framework for evaluating native-like LLM response quality across 47 languages using human-annotated preference pairs and audience design principles.

Ax Yuchen Song, Andong Chen, Wenxin Zhu, Kehai Chen, Xuefeng Bai, Muyun Yang, Tiejun Zhao 3/3/2026

Culture In a Frame: C$^3$B as a Comic-Based Benchmark for Multimodal Culturally Awareness

C³B benchmark for evaluating cultural awareness in multimodal LLMs using comics with cross-lingual tasks and varying difficulty levels.

Ax Amir Joudaki, Giulia Lanzillotta, Mohammad Samragh Razlighi, Iman Mirzadeh, Keivan Alizadeh, Thomas Hofmann, Mehrdad Farajtabar, Fartash Faghri 3/3/2026

Barriers for Learning in an Evolving World: Mathematical Understanding of Loss of Plasticity

arXiv paper analyzing loss of plasticity in deep learning on non-stationary data using dynamical systems theory.

Ax Luckeciano C. Melo, Alessandro Abate, Yarin Gal 3/3/2026

Stabilizing Policy Gradients for Sample-Efficient Reinforcement Learning in LLM Reasoning

arXiv paper on stabilizing policy gradients for sample-efficient RL in LLM reasoning to improve training efficiency.

Ax Zichen Liu, Anya Sims, Keyu Duan, Changyu Chen, Simon Yu, Xiangxin Zhou, Haotian Xu, Shaopan Xiong, Bo Liu, Chenmien Tan, Chuen Yang Beh, Weixun Wang, Hao Zhu, Weiyan Shi, Diyi Yang, Michael Shieh, Yee Whye Teh, Wee Sun Lee, Min Lin 3/3/2026

GEM: A Gym for Agentic LLMs

GEM is an open-source environment simulator for training agentic LLMs through experience-based learning, analogous to OpenAI Gym.

Ax Ali Hatamizadeh, Syeda Nahida Akter, Shrimai Prabhumoye, Jan Kautz, Mostofa Patwary, Mohammad Shoeybi, Bryan Catanzaro, Yejin Choi 3/3/2026

RLP: Reinforcement as a Pretraining Objective

arXiv paper proposing RLP, reinforcement learning as a pretraining objective for reasoning models instead of post-training only.

Ax Runzhe Zhan, Yafu Li, Zhi Wang, Xiaoye Qu, Dongrui Liu, Jing Shao, Derek F. Wong, Yu Cheng 3/3/2026

ExGRPO: Learning to Reason from Experience

arXiv paper on ExGRPO, reinforcement learning from verifiable rewards to improve LLM reasoning with experience reuse.

Ax Han Wang, Haoyu Li, Brian Ko, Huan Zhang 3/3/2026

On The Fragility of Benchmark Contamination Detection in Reasoning Models

arXiv paper showing benchmark contamination detection in reasoning models is easily evaded, highlighting evaluation vulnerabilities.

Ax Xinzhe Huang, Wenjing Hu, Tianhang Zheng, Kedong Xiu, Xiaojun Jia, Di Wang, Zhan Qin, Kui Ren 3/3/2026

Untargeted Jailbreak Attack

arXiv paper on untargeted jailbreak attacks against LLMs that optimize adversarial suffixes without fixed target responses.

Ax Heyang Gao, Zexu Sun, Erxue Min, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, Xu Chen 3/3/2026

Solving the Granularity Mismatch: Hierarchical Preference Learning for Long-Horizon LLM Agents

Hierarchical preference learning for long-horizon LLM agents addresses granularity mismatch between trajectory and step-level signals.

Ax Jialin Liu, Lisang Ding, Stanley Osher, Wotao Yin 3/3/2026

Expressive Power of Implicit Models: Rich Equilibria and Test-Time Scaling

Analysis of implicit models' expressive power showing infinite-depth weight-tied networks can match larger explicit models.

Ax Sahil Joshi, Agniva Chowdhury, Amar Kanakamedala, Ekam Singh, Evan Tu, Anshumali Shrivastava 3/3/2026

RACE Attention: A Strictly Linear-Time Attention for Long-Sequence Training

RACE Attention achieves linear-time complexity for long-sequence training, enabling multi-million token contexts.

Ax Junxi Yan, Zixi Wei, Qingyao Ai, Yiqun Liu, Jingtao Zhan 3/3/2026

What Scales in Cross-Entropy Scaling Law?

Analysis of cross-entropy scaling law breakdown at very large LLM scales and investigation of contributing factors.

Ax Chanjoo Jung, Jaehyung Kim 3/3/2026

TiTok: Transfer Token-level Knowledge via Contrastive Excess to Transplant LoRA

TiTok transfers token-level knowledge to enable LoRA parameters to work across different model architectures.

Ax Dachuan Shi, Abedelkadir Asi, Keying Li, Xiangchi Yuan, Leyan Pan, Wenke Lee, Wen Xiao 3/3/2026

SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs

SwiReasoning enables LLMs to switch between latent continuous and explicit discrete reasoning for token efficiency.

Ax Yilong Li, Shuai Zhang, Yijing Zeng, Hao Zhang, Xinmiao Xiong, Jingyu Liu, Pan Hu, Suman Banerjee 3/3/2026

Tiny but Mighty: A Software-Hardware Co-Design Approach for Efficient Multimodal Inference on Battery-Powered Small Devices

NANOMIND hardware-software co-design framework for efficient multimodal model inference on battery-powered mobile devices.

Ax Sheng Jia, Xiao Wang, Shiva Prasad Kasiviswanathan 3/3/2026