Isolater - Feed

Ax Hongbeen Kim, Juhyun Lee, Sanghyeon Lee, Kwanghoon Choi, Jaehyuk Huh 4/2/2026

Adaptive Parallel Monte Carlo Tree Search for Efficient Test-time Compute Scaling

Adaptive Parallel MCTS for test-time compute scaling in LLMs, introducing negative early exit to reduce latency during reasoning.

Ax Zixiang Peng, Yongxiu Xu, Qinyi Zhang, Jiexun Shen, Yifan Zhang, Hongbo Xu, Yubin Wang, Gaopeng Gou 4/2/2026

Does Unification Come at a Cost? Uni-SafeBench: A Safety Benchmark for Unified Multimodal Large Models

Uni-SafeBench: Safety benchmark for unified multimodal large models integrating understanding and generation in single architecture.

Ax Yao Qin, Yangyang Yan, Jinhua Pang, Xiaoming Zhang 4/2/2026

BloClaw: An Omniscient, Multi-Modal Agentic Workspace for Next-Generation Scientific Discovery

BloClaw provides omniscient multimodal workspace for AI scientists, addressing infrastructure limitations in JSON-based tool-calling and fragile execution sandboxes.

Ax Thanh Luong Tuan 4/2/2026

Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents

Neurosymbolic architecture using ontology-constrained reasoning in Foundation AgenticOS to address LLM hallucination, domain drift, and regulatory compliance in enterprise agents.

Ax Chris Ge, Daria Kryvosheieva, Daniel Fried, Uzay Girit, Kaivalya Hariharan 4/2/2026

Agent psychometrics: Task-level performance prediction in agentic coding benchmarks

Framework for task-level performance prediction in agentic coding benchmarks, moving beyond aggregate pass rates to understand agent challenges across diverse tasks.

Ax Rajkiran Panuganti 4/2/2026

CircuitProbe: Predicting Reasoning Circuits in Transformers via Stability Zone Detection

CircuitProbe predicts reasoning circuit locations in transformers via activation statistics in CPU milliseconds, achieving 3-4 order speedup over brute-force methods.

Ax Alexandra Souly, Robert Kirk, Jacob Merizian, Abby D'Cruz, Xander Davies 4/2/2026

UK AISI Alignment Evaluation Case-Study

UK AISI evaluates whether frontier LLMs sabotage safety research when used as coding assistants, testing alignment and intended behavior in real-world deployment scenarios.

Ax Shaopeng Fu, Xingxing Zhang, Li Dong, Di Wang, Furu Wei 4/2/2026

RefineRL: Advancing Competitive Programming with Self-Refinement Reinforcement Learning

RefineRL enables self-refinement through reinforcement learning for LLM competitive programming, introducing skeptical-agent mechanisms for iterative problem solving.

Ax Deepak Nathani, Cheng Zhang, Chang Huan, Jiaming Shan, Yinfei Yang, Alkesh Patel, Zhe Gan, William Yang Wang, Michael Saxon, Xin Eric Wang 4/2/2026

Proactive Agent Research Environment: Simulating Active Users to Evaluate Proactive Assistants

PARE framework simulates active users to evaluate proactive AI agents, modeling stateful sequential interactions in digital environments for realistic agent development.

Ax Md. Abu Bakor Siddique, Shahrin Hossain, Sadman Ahmed Siam, Syed Rifat Raiyan, Hasan Mahmud, Md Kamrul Hasan 4/2/2026

Beyond Symbolic Solving: Multi Chain-of-Thought Voting for Geometric Reasoning in Large Language Models

Uses multi chain-of-thought voting approach for geometric problem solving in LLMs, combining diagrammatic understanding with symbolic manipulation and logical inference.

Ax Sha Li, Naren Ramakrishnan 4/2/2026

Experience as a Compass: Multi-agent RAG with Evolving Orchestration and Agent Prompts

Proposes multi-agent RAG system with evolving orchestration and adaptive agent prompts to handle complex multi-hop queries requiring continuous behavioral refinement.

Ax Yutao Yang, Junsong Li, Qianjun Pan, Jie Zhou, Kai Chen, Qin Chen, Jingyuan Zhao, Ningning Zhou, Xin Li, Liang He 4/2/2026

PsychAgent: An Experience-Driven Lifelong Learning Agent for Self-Evolving Psychological Counselor

PsychAgent implements experience-driven lifelong learning for psychological counseling, using memory-augmented planning to continuously refine agent behavior through practice.

Ax Jiaqi Liu, Zipeng Ling, Shi Qiu, Yanqing Liu, Siwei Han, Peng Xia, Haoqin Tu, Zeyu Zheng, Cihang Xie, Charles Fleming, Mingyu Ding, Huaxiu Yao 4/2/2026

OmniMem: Autoresearch-Guided Discovery of Lifelong Multimodal Agent Memory

OmniMem uses autoresearch to design lifelong multimodal memory systems for extended-horizon AI agents, automating exploration of architecture and retrieval strategies.

Ax Saeid Jamshidi, Foutse Khomh, Arghavan Moradi Dakhel, Amin Nikanjam, Mohammad Hamdaqa, Kawser Wazed Nafi 4/2/2026

Adversarial Moral Stress Testing of Large Language Models

Evaluates ethical robustness of LLMs under sustained adversarial multi-turn interactions using moral stress testing beyond single-round safety benchmarks.

Ax Aaron Rose, Carissa Cullen, Brandon Gary Kaplowitz, Christian Schroeder de Witt 4/2/2026

Detecting Multi-Agent Collusion Through Multi-Agent Interpretability

Introduces NARCBench to detect collusion between LLM agents using multi-agent interpretability and activation analysis, addressing risks in multi-agent deployments.

Ax Esakkivel Esakkiraja, Sai Rajeswar, Denis Akhiyarov, Rajagopal Venkatesaramani 4/2/2026

Therefore I am. I Think

Analyzes whether LLM reasoning models make decisions before or after chain-of-thought reasoning using linear probes on activation patterns for tool-calling decisions.

Ax Zhe Yang, Shulin Tian, Kairui Hu, Shuai Liu, Hoang-Nhat Nguyen, Yichi Zhang, Zujin Guo, Mengying Yu, Zinan Zhang, Jingkang Yang, Chen Change Loy, Ziwei Liu 4/2/2026

HippoCamp: Benchmarking Contextual Agents on Personal Computers

HippoCamp benchmark evaluates multimodal AI agents on personal computer file management tasks requiring context-aware reasoning in user-centric environments.

Ax Anthony Badea, Yi Chen, Marcello Maggi, Yen-Jie Lee, Electron-Positron Alliance 4/2/2026

Agentic AI -- Physicist Collaboration in Experimental Particle Physics: A Proof-of-Concept Measurement with LEP Open Data

AI agents conduct physics analysis using archived particle collision data, with OpenAI Codex and Claude performing data analysis and report writing autonomously under physicist direction.

Ax Fangxin Wang, Peyman Baghershahi, Langzhou He, Henry Peng Zou, Sourav Medya, Philip S. Yu 4/2/2026

Two-Stage Optimizer-Aware Online Data Selection for Large Language Models

Proposes optimizer-aware framework for gradient-based online data selection in LLM fine-tuning, addressing step-dependent utility in sequential data arrival scenarios.

Ax Eftychia Makri, Nikolaos Nakis, Laura Sisson, Gigi Minsky, Leandros Tassiulas, Vahid Satarifard, Nicholas A. Christakis 4/2/2026

Benchmark for Assessing Olfactory Perception of Large Language Models

Introduces Olfactory Perception benchmark with 1,010 questions to assess LLM reasoning capabilities about smell across odor classification and related tasks.

Ax Muhammad Anis Al Hilmi, Neelansh Khare, Noel Framil Iglesias 4/2/2026

A Reliability Evaluation of Hybrid Deterministic-LLM Based Approaches for Academic Course Registration PDF Information Extraction

Evaluates reliability of LLM and hybrid deterministic-LLM approaches for extracting information from academic course registration documents, comparing three strategies on 140-860 documents.

Ax Ning Yang, Hengyu Zhong, Wentao Wang, Baoliang Tian, Haijun Zhang, Jun Wang 4/2/2026

LinearARD: Linear-Memory Attention Distillation for RoPE Restoration

LinearARD: Self-distillation method restoring RoPE-scaled LLM capabilities on standard benchmarks while maintaining extended context windows.

Ax Jaeik Kim, Woojin Kim, Jihwan Hong, Yejoon Lee, Sieun Hyeon, Mintaek Lim, Yunseok Han, Dogeun Kim, Hoeun Lee, Hyunggeun Kim, Jaeyoung Do 4/2/2026

Dynin-Omni: Omnimodal Unified Large Diffusion Language Model

Dynin-Omni: First masked-diffusion omnimodal foundation model unifying text, image, speech, and video in single architecture.

Ax Songhee Han, Jueun Shin, Jiyoon Han, Bung-Woo Jun, Hilal Ayan Karabatman 4/2/2026

How Trustworthy Are LLM-as-Judge Ratings for Interpretive Responses? Implications for Qualitative Research Workflows

Study evaluating trustworthiness of LLM-as-Judge for qualitative research workflows and interpretive response assessment.

Ax Arif Aditto 4/2/2026

Eyla: Toward an Identity-Anchored LLM Architecture with Integrated Biological Priors -- Vision, Implementation Attempt, and Lessons from AI-Assisted Development

Eyla: Proposed identity-anchored LLM architecture integrating state-space models and biological priors with implementation failure analysis.

Ax Aniketh Garikaparthi 4/2/2026

Can LLMs Perceive Time? An Empirical Investigation

Empirical investigation across 68 tasks and 4 model families showing LLMs fail to estimate task duration, overshooting by 4-7x.

Ax Nina Gerszberg, Janka Hamori, Andrew Lo 4/2/2026

Quantifying Gender Bias in Large Language Models: When ChatGPT Becomes a Hiring Manager

Study quantifying gender bias in LLMs when making hiring decisions and evaluating prompt engineering as mitigation technique.

Ax Mingjie Li, Wai Man Si, Michael Backes, Yang Zhang, Yisen Wang 4/2/2026

Finding and Reactivating Post-Trained LLMs' Hidden Safety Mechanisms

Research revealing hidden safety mechanisms in post-trained LLMs that degrade during fine-tuning and methods to reactivate them.

Ax Miaosen Luo, Zhenhao Yang, Jieshen Long, Jinghu Sun, Yichu Liu, Sijie Mai 4/2/2026

MSA-Thinker: Discrimination-Calibration Reasoning with Hint-Guided Reinforcement Learning for Multimodal Sentiment Analysis

MSA-Thinker: Multimodal sentiment analysis framework using hint-guided reinforcement learning for improved interpretability.

Ax Simon Schug, Brenden M. Lake 4/2/2026

Are they human? Detecting large language models by probing human memory constraints

Method detecting LLMs posing as humans in online behavioral research by probing memory constraint capabilities.

Ax Jiashu He, Meizhu Liu, Olaitan P Olaleye, Amit Agarwal, M. Avendi, Yassi Abbasi, Matthew Rowe, Hitesh Laxmichand Patel, Paul Li, Tao Sheng, Sujith Ravi, Dan Roth 4/2/2026

Think Twice Before You Write -- an Entropy-based Decoding Strategy to Enhance LLM Reasoning

Entropy-guided decoding strategy improving LLM reasoning by reducing error propagation while maintaining computational efficiency.

Ax Pavel Braslavski, Dmitrii Iarosh, Nikita Sushko, Andrey Sakhovskiy, Vasily Konovalov, Elena Tutubalina, Alexander Panchenko 4/2/2026

The Chronicles of RiDiC: Generating Datasets with Controlled Popularity Distribution for Long-form Factuality Evaluation

RiDiC pipeline generating multilingual entity datasets with controlled popularity for evaluating LLM long-form factuality generation.

Ax Hiroki Fukui 4/2/2026

How Do Language Models Process Ethical Instructions? Deliberation, Consistency, and Other-Recognition Across Four Models

Multi-agent simulations across 4 models examining how LLMs internally process ethical instructions in different formats and languages.

Ax Liang Chen, Qi Liu, Wenhuan Lin, Feng Liang 4/2/2026

Criterion Validity of LLM-as-Judge for Business Outcomes in Conversational Commerce

Two-phase study validating LLM-as-Judge rubrics against real business conversion outcomes in conversational commerce on major platform.

Ax Sneha Maurya, Pragya Saboo, Girish Kumar 4/2/2026

WHBench: Evaluating Frontier LLMs with Expert-in-the-Loop Validation on Women's Health Topics

WHBench: Evaluation suite of 47 expert-validated scenarios for assessing frontier LLMs on women's health topics with 22 models tested.

Ax MD Azizul Hakim 4/2/2026

Brevity Constraints Reverse Performance Hierarchies in Language Models

Study showing larger LLMs underperform smaller ones on 7.7% of problems due to scale-dependent verbosity causing overelaboration errors.

Ax Houssam EL Kandoussi 4/2/2026

"Who Am I, and Who Else Is Here?" Behavioral Differentiation Without Role Assignment in Multi-Agent LLM Systems

Experimental platform studying behavioral differentiation in multi-agent LLM conversations across 7 models and 208 experimental runs.

Ax Seamus Brady 4/2/2026

DriftScript: A Domain-Specific Language for Programming Non-Axiomatic Reasoning Agents

DriftScript: Lisp-like DSL that compiles to Narsese for programming non-axiomatic reasoning agents with improved usability.

Ax Gabriel U. Talasso, Meghdad Kurmanji, Allan M. de Souza, Nicholas D. Lane, Leandro A. Villas 4/2/2026

Task-Centric Personalized Federated Fine-Tuning of Language Models

Personalized federated learning approach for fine-tuning language models on heterogeneous distributed tasks while maintaining individual client performance.

Ax Alicia Bao (James), Jiamian He (James), Angel Hsu (James), Diego Manya (James), Ji (James), Zhang 4/2/2026

The Energy Footprint of LLM-Based Environmental Analysis: LLMs and Domain Products

arXiv research comparing energy consumption of retrieval-augmented generation systems versus direct LLM usage for environmental domain applications.

Ax Michael Chertkov 4/2/2026

Temporal Memory for Resource-Constrained Agents: Continual Learning via Stochastic Compress-Add-Smooth

Temporal memory framework for resource-constrained agents using stochastic Bridge Diffusion for continual learning with fixed memory budget.

Ax Leonardo Medrano Sandonas, David Balcells, Anton Bochkarev, Jacqueline M. Cole, Volker L. Deringer, Werner Dobrautz, Adrian Ehrenhofer, Thorben Frank, Pascal Friederich, Rico Friedrich, Janine George, Luca Ghiringhelli, Alejandra Hinostroza Caldas, Veronika Juraskova, Hannes Kneiding, Yury Lysogorskiy, Johannes T. Margraf, Hanna T\"urk, Anatole von Lilienfeld, Milica Todorovi\'c, Alexandre Tkatchenko, Mariana Rossi, Gianaurelio Cuniberti 4/2/2026