Isolater - Feed

Ax Jaechang Kim, Yotaro Shimose, Zhao Wang, Kuang-Da Wang, Jungseul Ok, Shingo Takamatsu 7/1/2026

Visual Prompt Discovery via Semantic Exploration

Method for generating visual prompts via semantic exploration to improve LLM perception and visual reasoning capabilities.

Ax Ruijia Yang, Zeyi Wen 7/1/2026

An Efficient Heterogeneous Co-Design for Fine-Tuning on a Single GPU

SlideFormer system for efficient LLM fine-tuning on single GPUs using asynchronous computation overlapping and CPU-GPU co-design.

Ax Weixin Chen, Antonio Vergari, Han Zhao 7/1/2026

Can VLMs Reason Robustly? A Neuro-Symbolic Investigation

Study of vision-language model robustness under distribution shifts in visual deductive reasoning tasks.

Ax Juhan Park, Taerim Yoon, Seungmin Kim, Joong-Gil Kim, Wontae Ye, Jeongeun Park, Yoonbyung Chai, Geonwoo Cho, Geunwoo Cho, Dohyeong Kim, Kyungjae Lee, Yong-Jae Kim, Sungjoon Choi 7/1/2026

Learning Dexterous Grasping from Sparse Taxonomy Guidance

Research on dexterous robotic manipulation using sparse taxonomy guidance for grasp planning with reinforcement learning and multi-finger control.

Ax Lifu Wang, Pan Zhou 7/1/2026

SMART: When is it Actually Worth Expanding a Speculative Tree?

Optimization method for tree-based speculative decoding that balances computational overhead with token acceptance in LLM generation, improving inference efficiency.

Ax Hanjun Cho, Jay-Yoon Lee 7/1/2026

RARE: Redundancy-Aware Retrieval Evaluation Framework for High-Similarity Corpora

Evaluation framework for retrieval-augmented generation systems on high-redundancy corpora like financial reports and legal documents, addressing real-world RAG limitations.

Ax Hanjun Cho, Gahyun Yoo, Hanseong Kim, Jay-Yoon Lee 7/1/2026

Generalizing Numerical Reasoning in Table Data through Operation Sketches and Self-Supervised Learning

arXiv paper on TaNOS framework improving numerical reasoning over tables via operation sketches and self-supervised learning.

Ax Zhuonan Yang, Jacob Xiaochen Li, Francisco Piedrahita Velez, Eric Todd, David Bau, Michael L. Littman, Stephen H. Bach, Ellie Pavlick 7/1/2026

Shared Lexical Task Representations Explain Behavioral Variability In LLMs

arXiv paper analyzing LLM prompt sensitivity, showing shared lexical task representations explain behavioral variability across prompting styles.

Ax Zhen Liu, Yuhan Liu, Jinjun Wang, Wei Song, Jianyi Liu, Jingwen Fu 7/1/2026

Structured Progressive Knowledge Activation for LLM-Driven Neural Architecture Search

arXiv paper using LLMs to guide neural architecture search by translating architectural knowledge into executable code edits.

Ax Zhiqing Zhong, Zhijing Ye, Jian Zhang, Weijian Zheng, Bolun Sun, Xiaodong Yu 7/1/2026

KV-RM: Regularizing KV-Cache Movement for Static-Graph LLM Serving

arXiv paper on KV-cache optimization for static-graph LLM serving. Infrastructure-focused, limited LLM application relevance.

Ax Zhiqing Zhong, Zhijing Ye, Jiamin Wang, Xiaodong Yu 7/1/2026

An Executable Benchmarking Suite for Tool-Using Agents

arXiv paper on executable benchmarking suite standardizing evaluation of tool-using agents across web, code, and task environments.

Ax Xiaochong Jiang, Shiqi Yang, Ziwei Li, Lifei Liu, Haoran Yu, Yichen Liu 7/1/2026

ChainCaps: Composition-Safe Tool-Using Agents via Monotonic Capability Attenuation

arXiv paper on safety mechanisms for tool-composing agents via monotonic capability attenuation to prevent unsafe end-to-end effects.

Ax Madhulatha Mandarapu, Sandeep Kunkunuru 7/1/2026

Knowledge Graphs as the Missing Data Layer for LLM-Based Industrial Asset Operations

arXiv paper showing knowledge graphs improve LLM agent accuracy in industrial asset operations vs flat document stores.

Ax Sebastian Nagl, Ann-Kristin Mayrhofer, Martin Heidebach, Aleyna Ko\c{c}ak, Anne Zettelmeier, Elly Breu, Angelina Greiner, Sofija Milijas, Matthias Grabmair 7/1/2026

BenGER: Benchmarking LLM Systems on Subsumption-Based Legal Reasoning in German Law

arXiv benchmark dataset (BenGER) for evaluating LLM reasoning on German legal subsumption tasks.

Ax Siyu Chen, Miao Lu, Beining Wu, Heejune Sheen, Fengzhuo Zhang, Shuangning Li, Zhiyuan Li, Jose Blanchet, Tianhao Wang, Zhuoran Yang 7/1/2026

INFUSER: Influence-Guided Self-Evolution Improves Reasoning

arXiv paper presenting INFUSER, iterative co-training framework for self-improving LLM reasoning without heavy supervision.

Ax Tianyu Ding, Jianhong Xin, Juan Pablo De la Cruz Weinstein 7/1/2026

Keep Policy Gradient in Charge: Sibling-Guided Credit Distillation for Long-Horizon Tool-Use Agents

arXiv paper on credit distillation method for training long-horizon tool-use agents with reinforcement learning.

Ax Jiaxin Ai, Tao Hu, Xuemeng Yang, Shu Zou, Hairong Zhang, Daocheng Fu, Yu Yang, Hongbin Zhou, Nianchen Deng, Pinlong Cai, Zhongyuan Wang, Botian Shi, Kaipeng Zhang, Licheng Wen 7/1/2026

ComAct: Reframing Professional Software Manipulation via COM-as-Action Paradigm

arXiv paper proposing COM-as-Action paradigm for AI agents manipulating professional software via component objects.

Ax Xilong Wang, Xiaoxing Chen, Patrick Li, Dawn Song, Neil Gong 7/1/2026

Same-Origin Policy for Agentic Browsers

arXiv paper studying security of agentic browsers, examining same-origin policy effectiveness with AI agents.

Ax Soham Bhattacharjee, Dushyant Singh Chauhan, Salem Lahlou, Martin Takac, Nils Lukas 7/1/2026

Entropy-Gated Latent Recursion

arXiv paper proposing inference-time scaling method for LLM reasoning via deterministic layer recursion.

Ax Rome Thorstenson 7/1/2026

Protein contacts are already in the attention: a single-forward-pass alternative to the Categorical Jacobian

Shows protein contact prediction signals are concentrated in attention heads of language models, enabling efficient single-pass prediction.

Ax Meher Bhaskar Madiraju, Meher Sai Preetam Madiraju 7/1/2026

RigorBench: Benchmarking Engineering Process Discipline in Autonomous AI Coding Agents

RigorBench evaluates autonomous coding agents beyond correctness, measuring engineering discipline and process quality in problem-solving approaches.

Ax Shivam Ratnakar, Kartikeya Vats 7/1/2026

The Geometry of Refusal: Linear Instability in Safety-Aligned LLMs

Uses contrastive logit steering to identify linear refusal mechanisms in safety-aligned LLMs, showing safety compliance as manipulable linear feature.

Ax Arthur Wuhrmann, Gaetan Stein, Daniel Brunner, Andrei Kucharavy 7/1/2026

Measuring & Mitigating Over-Alignment for LLMs in Multilingual Criminal Law Courts

Studies over-alignment issues in multilingual LLMs for criminal law translation and summarization in Swiss courts.

Ax Annika Marie Schoene, Cansu Canca, Gautham Vijay Kumar, Anson Antony 7/1/2026

One Year Later...The Harms Persist, But So Do We!

Evaluates safety guardrails of eight LLMs across 16 psychiatric conditions using adversarial attacks and introduces harm taxonomy framework.

Ax Kwok Chun Au, Adam Block 7/1/2026

Training for the Model You Return: Improving Optimization for Iterate-Averaged Language Models

Improves language model training when using iterate-averaged models by reformulating optimizer design as optimal control problem.

Ax DatologyAI, :, Matthew L. Leavitt, Siddharth Joshi, Haoli Yin, Rishabh Adiga, Haakon Mongstad, Alvin Deng, David Schwab, Bogdan Gaza, Ari Morcos 7/1/2026

Brevity is the Soul of Inference Efficiency: Inducing Concision in VLMs via Data Curation

Reduces VLM inference costs by curating pretraining data to induce concise outputs rather than using traditional model compression techniques.

Ax Shuimu Chen, Yuteng Chen, Yuanshen Guan, Zebang Cheng, Zeyu Zhang, Shengqian Qin, Bin Xia, Jiaran Li, Wenming Yang, Fei Ma 7/1/2026

Reflect-R1: Evidence-Driven Reflection for Self-Correction in Long Video Understanding

Reflect-R1 improves long video understanding through evidence-driven self-correction using external verification rather than internal reflection.

Ax Sergey Kurilenko 7/1/2026

SHARD: cell-keyed residual splitting for alignment-resistant private dense retrieval

SHARD protects dense vector stores from inversion attacks by using cell-keyed residual splitting to resist alignment-based attacks on RAG systems.

Ax Yujin Wang, Junli Chen, Yixuan Li, Shunan Dong, Huazhong Yang, Yongpan Liu, Hongyang Jia 7/1/2026

The Speedup Paradox: Rethinking Inference Speed-Quality Trade-off in Embodied Tasks

Studies how inference speed optimizations in embodied AI models affect action quality differently than static ML tasks through empirical analysis.

Ax Bal\'azs Szalontai, \'Abel Szauter, Bal\'azs M\'arton, P\'eter Verebics, Bal\'azs Pint\'er, Tibor Gregorics 7/1/2026

Diff-Based Code Corruption using LLMs for Large-Scale Bugfix Benchmarking

Generates diverse, realistic bugs in code using LLM-based diffs to create larger-scale, more representative bugfix benchmarks for evaluating LLM capabilities.

Ax Kan Zhu, Mathew Jacob, Chenxi Ma, Yi Pan, Stephanie Wang, Arvind Krishnamurthy, Baris Kasikci 7/1/2026

TraceLab: Characterizing Coding Agent Workloads for LLM Serving

TraceLab provides a public dataset and analysis of real coding agent workloads across multiple LLM-based agents and models to improve serving efficiency.

Ax Gan Luo, Zihan Qin, Bin Dong, Wotao Yin 7/1/2026

From Search to Synthesis: Training LLMs as Zero-Shot Workflow Generators

Training LLMs to generate executable workflows as zero-shot structured solutions encoding task-level algorithmic patterns for reproducible deployment.

Ax Zhongyao Wang 7/1/2026

Why Do Few-Step Text Latents Fail When Image Latents Work? Non-Commitment at Sharp Categorical Readouts

Analysis of why deterministic few-step generation fails for text latents due to geometric inability to resolve discrete choices before sharp readouts.

Ax Woernle Frank, Fedosov Vladimir, Grinenko Artemiy 7/1/2026

Hierarchical Global Attention (HGA)

Hierarchical Global Attention drop-in replacement for dense causal attention enabling 64K-token context without retraining or parameter changes.

Ax John Sweeney 7/1/2026

Revocable Learned State via Process Sidecars

Process sidecar method for revoking learned memory in adapted language models through two-coefficient edits after safety fine-tuning.

Ax Rajat Ghosh, Datta Nimmaturi, Aryan Singhal, Vaishnavi Bhargava, Henry Wong, Johnu George, Debojyoti Dutta 7/1/2026

Predictable GRPO: A Closed-Form Model of Training Dynamics

First-principles reduced-order model of GRPO training dynamics for LLM reasoning with closed-form analysis replacing empirical hyperparameter tuning.

Ax Haoming Meng, Anton Sugolov, Vardan Papyan 7/1/2026

Gradient Smoothing: Coupling Layer-wise Updates for Improved Optimization

Depth-wise gradient augmentation optimization paradigm for transformers leveraging structured layer relationships to improve training dynamics.

Ax Ved Sriraman, Peihan Liu, Daniel Hsu, Adam Block 7/1/2026

Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback

Imitation learning theory explaining why on-policy distillation outperforms offline supervised fine-tuning under noisy expert feedback using theoretical model.

Ax Luke Budny, Yuhong Guo, Kevin Cheung 7/1/2026

Quality-Aware Modulation for Diffusion Transformers

Quality-aware modulation for diffusion transformers improving image generation by incorporating quality information into denoising process modulation.

Ax Huanlin Gao, Fang Zhao, Qiang Hui, Fuyuan Shi, Shaoan Zhao, Yantao Li, Chao Tan, Ting Lu, Yuren You, Kai Wang, Shiguo Lian 7/1/2026

OTCache: Optimal Transport for Geometry-Aware Caching in Diffusion Models

Training-free diffusion model acceleration using optimal transport for geometry-aware caching schedule prediction across inference budgets.

Ax Shreyas Rajesh, Kartik Sharma, Tonmoy Monsoor, Mehmet Yigit Turali, Richard Idro, Juliana Kayaga, Robert Sebunya, Tracy Tushabe Namata, Jessica Nichole Pasqua, Vwani Roychowdhury, Rajarshi Mazumder 7/1/2026

Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care

LLM-based clinical decision support for epilepsy medication prediction in resource-constrained settings, adapted to local practice with deferral capability.

Ax Gaurab Baral, Aaditya Khanal, Yangyang Tao, Junxiu Zhou 7/1/2026

Knowledge Distillation from Large Reasoning Models to Compact Student Models: A Case Study on the John O Bryan Mathematics Competition

Knowledge distillation from reasoning model to compact student via dual-agent Chain-of-Thought generation, fine-tuned with LoRA on edge hardware.

Ax Rui Zhou, Tianci Xie 7/1/2026

TDGT: A Tabular Data Generation Toolkit supporting adaptive GPU-accelerated Bayesian mixture models, diffusion-based models, and latent-space generative modeling

Web-based toolkit for synthetic tabular data generation using Bayesian mixtures, diffusion models, and latent-space methods with privacy preservation.