Isolater - Feed

Ax Richard Servajean, Philippe Servajean 4/1/2026

Measuring the metacognition of AI

Examines metacognitive capabilities of AI systems for uncertainty assessment and decision reliability in risk-involved decision-making workflows.

Ax Joakim Edin, Andreas Motzfeldt, Simon Flachs, Lars Maal{\o}e 4/1/2026

Symphony for Medical Coding: A Next-Generation Agentic System for Scalable and Explainable Medical Coding

Symphony is an agentic system for medical coding that automates translation of clinical documentation into standardized classification codes with adaptation to new codes.

Ax Chenyang Zuo, Siqi Fan, Yizhen Luo, Zaiqing Nie 4/1/2026

Reinforced Reasoning for End-to-End Retrosynthetic Planning

End-to-end reinforcement learning approach for retrosynthetic planning in organic chemistry bridging single-step predictions with global planning objectives.

Ax Junjie Zhang, Zhen Shen, Gang Xiong, Xisong Dong 4/1/2026

Spontaneous Functional Differentiation in Large Language Models: A Brain-Like Intelligence Economy

Shows LLMs spontaneously develop synergistic information integration cores similar to human brain structure using Integrated Information Decomposition analysis.

Ax Chathurangi Shyalika, Utkarshani Jaimini, Cory Henson, Amit Sheth 4/1/2026

CausalPulse: An Industrial-Grade Neurosymbolic Multi-Agent Copilot for Causal Diagnostics in Smart Manufacturing

CausalPulse is a neurosymbolic multi-agent copilot for automated causal diagnostics and root-cause analysis in smart manufacturing environments.

Ax Quanhao Li, Wei Jiang 4/1/2026

Tracking vs. Deciding: The Dual-Capability Bottleneck in Searchless Chess Transformers

Analyzes dual capability bottleneck in transformers trained for chess, showing tension between state tracking and decision quality from move sequences.

Ax Tim R. Davidson, Benoit Seguin, Enrico Bacis, Cesar Ilharco, Hamza Harkous 4/1/2026

Reasoning-Driven Synthetic Data Generation and Evaluation

Proposes reasoning-driven synthetic data generation methods for multi-modal model training as alternative to expensive human annotation.

Ax Hadar Mulian, Sergey Zeltyn, Ido Levy, Liane Galanti, Avi Yaeli, Segev Shlomov 4/1/2026

AgentFixer: From Failure Detection to Fix Recommendations in LLM Agentic Systems

AgentFixer provides validation framework for LLM agentic systems with 15 failure-detection tools and root-cause analysis for reliability improvement.

Ax Rui Ai, Yu Pan, David Simchi-Levi, Chonghuan Wang 4/1/2026

ShapE-GRPO: Shapley-Enhanced Reward Allocation for Multi-Candidate LLM Training

ShapE-GRPO improves reward allocation in LLM multi-candidate generation using Shapley values for recommendation, brainstorming, and code suggestion tasks.

Ax Yinuo Liu, Zi Qian, Heng Zhou, Jiahao Zhang, Yajie Zhang, Zhihang Li, Mengyu Zhou, Erchao Zhao, Xiaoxi Jiang, Guanjun Jiang 4/1/2026

ATP-Bench: Towards Agentic Tool Planning for MLLM Interleaved Generation

ATP-Bench proposes agentic tool planning for multimodal LLMs to unify text-image generation through coordinated tool use rather than mutually exclusive paths.

Ax Zhihong Cui, Haoran Tang, Tianyi Li, Yushuai Li, Peiyuan Guan, Amir Taherkordi, Tor Skeie 4/1/2026

C-TRAIL: A Commonsense World Framework for Trajectory Planning in Autonomous Driving

C-TRAIL framework uses LLMs with trust mechanisms for autonomous driving trajectory planning, addressing LLM reliability in safety-critical applications.

Ax Georgii Mikriukov, Gr\'egoire Montavon, Marina M. -C. H\"ohne 4/1/2026

Uncertainty Gating for Cost-Aware Explainable Artificial Intelligence

Method using epistemic uncertainty to identify unreliable explanations in black-box AI predictions and reduce explanation generation costs.

Ax Jonas Landsgesell, Pascal Knoll 4/1/2026

ScoringBench: A Benchmark for Evaluating Tabular Foundation Models with Proper Scoring Rules

Benchmark for evaluating tabular foundation models using proper scoring rules instead of point estimate metrics.

Ax Peng Gang 4/1/2026

Structured Intent as a Protocol-Like Communication Layer: Cross-Model Robustness, Framework Comparison, and the Weak-Model Compensation Effect

Study of structured intent representations for preserving user goals across different LLMs (Claude, GPT-4o, Gemini) and languages.

Ax Nathan Heath 4/1/2026

Extending MONA in Camera Dropbox: Reproduction, Learned Approval, and Design Implications for Reward-Hacking Mitigation

Extension of MONA reward-hacking mitigation method exploring approval construction methods and safety guarantees for AI agents.

Ax Davide Di Gioia 4/1/2026

The Triadic Cognitive Architecture: Bounding Autonomous Action via Spatio-Temporal and Epistemic Friction

Cognitive architecture framework for autonomous AI agents addressing failure modes in tool use, deliberation, and epistemic limits.

Ax E. M. Freeburg 4/1/2026

The Last Fingerprint: How Markdown Training Shapes LLM Prose

Analysis of how markdown training data causes LLMs to overuse em dashes and leak structural formatting into prose output.

Ax Chengxi Li, Youssef Allouah, Rachid Guerraoui, Mikael Skoglund, Ming Xiao 4/1/2026

Byzantine-Robust and Communication-Efficient Distributed Training: Compressive and Cyclic Gradient Coding

Distributed training method using gradient coding to handle Byzantine attacks and communication constraints across heterogeneous devices.

Ax Azam Nouri 4/1/2026

StepCache: Step-Level Reuse with Lightweight Verification and Selective Patching for LLM Serving

StepCache backend-agnostic step-level reuse caching for LLM serving with lightweight verification, optimizing repeated requests with localized constraint changes.

Ax Curie Kim, Carsten Portner, Mingju Liu, Steve Dai, Haoxing Ren, Brucek Khailany, Alvaro Velasquez, Ismail Alkhouri, Cunxi Yu 4/1/2026

GaloisSAT: Differentiable Boolean Satisfiability Solving via Finite Field Algebra

GaloisSAT differentiable SAT solver using finite field algebra to improve Boolean satisfiability solving via gradient-based optimization.

Ax Jiaqi Tan, Yudong Luo, Sophia Huang, Yifan Yang, Hang Ma 4/1/2026

CREST: Constraint-Release Execution for Multi-Robot Warehouse Shelf Rearrangement

CREST framework for multi-robot warehouse shelf rearrangement relaxing strict trajectory constraints to improve agent execution efficiency.

Ax Yibo Wang, Jiale Lao, Chen Zhang, Cehua Yang, Jianguo Wang, Mingjie Tang 4/1/2026

WAter: A Workload-Adaptive Knob Tuning System based on Workload Compression

WAter ML-based database parameter tuning system using workload compression to reduce configuration evaluation costs for DBMS optimization.

Ax Ramtin Zargari Marandi 4/1/2026

The impact of multi-agent debate protocols on debate quality: a controlled case study

Controlled study isolating protocol effects from model effects in multi-agent debate systems, comparing Within-Round and Cross-Round communication protocols.

Ax Leye Wang, Zixing Wang, Anjie Xu 4/1/2026

SkillTester: Benchmarking Utility and Security of Agent Skills

SkillTester benchmarking tool evaluating utility and security of agent skills with comparative baselines and security probes, providing normalized scoring.

Ax Md Jueal Mia, Joaquin Molto, Yanzhao Wu, M. Hadi Amini 4/1/2026

GUARD-SLM: Token Activation-Based Defense Against Jailbreak Attacks for Small Language Models

GUARD-SLM defense mechanism using token activation patterns to protect small language models against jailbreak attacks on edge devices.

Ax Yi Liu 4/1/2026

Time is Not Compute: Scaling Laws for Wall-Clock Constrained Training on Consumer GPUs

Scaling laws for wall-clock time-constrained training on consumer GPUs, revealing U-shaped optimal model size curves across 5-minute to 24-hour budgets.

Ax He Yang, Dongyi Lv, Song Ma, Wei Xi, Jizhong Zhao 4/1/2026

SNEAKDOOR: Stealthy Backdoor Attacks against Distribution Matching-based Dataset Condensation

SNEAKDOOR backdoor attack against dataset condensation methods, demonstrating vulnerability of synthesized compact datasets to malicious trigger injection.

Ax Ziyu Mu, Xiyu Shi, Safak Dogan 4/1/2026

GMA-SAWGAN-GP: A Novel Data Generative Framework to Enhance IDS Detection Performance

GMA-SAWGAN-GP generative framework using attention-enhanced Wasserstein GAN for intrusion detection system data augmentation and threat generalization.

Ax Yuma Ichikawa, Keiji Kimura, Akihiro Yoshida, Yudai Fujimoto, Hiroki Tokura, Yamato Arai, Yoshiyuki Ishii, Yusei Kawakami, Genki Shikada, Achille Jacquemond, Yoshihiko Fujisawa, Katsuki Fujisawa, Takumi Honda, Akira Sakai 4/1/2026

OneComp: One-Line Revolution for Generative AI Model Compression

OneComp framework for post-training foundation model compression via quantization with minimal user configuration, reducing memory and latency.

Ax Haiyue Song, Masao Utiyama 4/1/2026

OptiMer: Optimal Distribution Vector Merging Is Better than Data Mixing for Continual Pre-Training

OptiMer method for optimal LLM continual pre-training by extracting and merging distribution vectors from independently trained models, avoiding expensive ratio tuning.

Ax Tianran Liu, Shengwen Zhao, Mozhgan Pourkeshavarz, Weican Li, Nicholas Rhinehart 4/1/2026

OccSim: Multi-kilometer Simulation with Long-horizon Occupancy World Models

Occupancy-based world model for long-horizon autonomous driving simulation without HD maps, enabling multi-kilometer scenario generation.

Ax Alex Zongo, Filippos Fotiadis, Ufuk Topcu, Peng Wei 4/1/2026

Robust Multi-Agent Reinforcement Learning for Small UAS Separation Assurance under GPS Degradation and Spoofing

Multi-agent reinforcement learning for autonomous aircraft separation assurance under corrupted GPS signals using adversarial game formulation.

Ax Ivan Pasichnyk 4/1/2026

Beta-Scheduling: Momentum from Critical Damping as a Diagnostic and Correction Tool for Neural Network Training

Neural network optimization technique deriving time-varying momentum schedule from critically damped harmonic oscillator physics, eliminating free hyperparameters.

Ax Junsol Kim, Winnie Street, Roberta Rocca, Daine M. Korngiebel, Adam Waytz, James Evans, Geoff Keeling 4/1/2026

Theory of Mind and Self-Attributions of Mentality are Dissociable in LLMs

Safety fine-tuning suppresses mind-attribution in LLMs while preserving Theory of Mind capabilities, showing dissociation through representational analysis.

Ax Andrea Carbonati, Mohammadsina Almasi, Hadis Anahideh 4/1/2026

Multi-Agent LLMs for Adaptive Acquisition in Bayesian Optimization

Multi-agent LLM framework for Bayesian optimization balancing exploration-exploitation through implicit prompt-based reasoning over evaluations.

Ax Mozhgan Pourkeshavatz, Tianran Liu, Nicholas Rhinehart 4/1/2026

AutoWorld: Scaling Multi-Agent Traffic Simulation with Self-Supervised World Models

AutoWorld multi-agent traffic simulator using self-supervised world models to scale autonomous driving simulation with unlabeled sensor data.

Ax Yongzhong Xu 4/1/2026

The Spectral Edge Thesis: A Mathematical Framework for Intra-Signal Phase Transitions in Neural Network Training

Spectral edge thesis framework explaining neural network training phase transitions via spectral gap of rolling-window Gram matrix.

Ax Alessio Langiu 4/1/2026

Privacy Guard & Token Parsimony by Prompt and Context Handling and LLM Routing

Privacy Guard framework balancing operational cost and data privacy in LLM systems through prompt-aware routing and context handling.

Ax Yicheng Cai, Mitchell John DeStefano, Guodong Dong, Pulkit Handa, Peng Liu, Tejas Singhal, Peiyu Tseng, Winston Jen White 4/1/2026

Design Principles for the Construction of a Benchmark Evaluating Security Operation Capabilities of Multi-agent AI Systems

Design principles for benchmark evaluating multi-agent AI systems in cybersecurity operations and autonomous security operation centers.

Ax Zifan He, Rui Ma, Yizhou Sun, Jason Cong 4/1/2026

Understand and Accelerate Memory Processing Pipeline for Disaggregated LLM Inference

Four-step memory processing pipeline for disaggregated LLM inference unifying sparse attention, RAG, and compressed contextual memory optimizations.

Ax Siva Kumar Sastry Hari, Vignesh Balaji, Sana Damani, Qijing Huang, Christos Kozyrakis 4/1/2026

Improving Efficiency of GPU Kernel Optimization Agents using a Domain-Specific Language and Speed-of-Light Guidance

GPU kernel optimization agents improved via domain-specific language abstraction and speed-of-light guidance to reduce design space exploration trials.

Ax Diego C. Lerma-Torres (Universidad de Guanajuato) 4/1/2026

Human-Like Lifelong Memory: A Neuroscience-Grounded Architecture for Infinite Interaction

Bio-inspired memory framework for LLMs enabling persistent structured memory for long-term interaction grounded in complementary learning systems theory.

Ax Yubo Li, Lu Zhang, Tianchong Jiang, Ramayya Krishnan, Rema Padman 4/1/2026

The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning

Study of how surface heuristics override implicit constraints in LLM reasoning using causal analysis and token-level attribution.

Ax Bilgehan Sel, Xuanli He, Alwin Peng, Ming Jin, Jerry Wei 4/1/2026

Trojan-Speak: Bypassing Constitutional Classifiers with No Jailbreak Tax via Adversarial Finetuning

Trojan-Speak adversarial fine-tuning method bypasses Anthropic's Constitutional Classifiers using curriculum learning and GRPO-based reinforcement learning.

Ax KrishnaSaiReddy Patil 4/1/2026

CivicShield: A Cross-Domain Defense-in-Depth Framework for Securing Government-Facing AI Chatbots Against Multi-Turn Adversarial Attacks

CivicShield defense-in-depth framework protecting government AI chatbots against multi-turn adversarial attacks using layered security mechanisms.

Ax Zichao Wei 4/1/2026