Isolater - Feed

Ax Gergely Szilvasy (Meta FAIR), Manuel Faysse (Meta FAIR, MICS, CentraleSup\'elec), Maria Lomeli (Meta FAIR), Matthijs Douze (Meta FAIR), Pierre-Emmanuel Mazar\'e (Meta FAIR), Lo\"ic Cabannes (Meta FAIR), Wen-tau Yih (Meta FAIR), Herv\'e J\'egou (Meta FAIR) 5/15/2026

Self-Pruned Key-Value Attention: Learning When to Write by Predicting Future Utility

Self-Pruned Key-Value Attention mechanism reduces KV cache size in transformers by predicting future utility for efficient long-sequence generation.

Ax Karan Sehgal, Khawar Naveed Bhatti 5/15/2026

Comparative Evaluation of Machine Learning Approaches for Minority-Class Financial Distress Prediction Under Class Imbalance Constraints

Comparative study of ML approaches for financial distress prediction under class imbalance constraints using classical and neural methods.

Ax Mohammad R. Rezaei, Tejas Balaji, Rahul G. Krishnan 5/15/2026

SurF: A Generative Model for Multivariate Irregular Time Series Forecasting

SurF generative model for forecasting irregular multivariate time series using Time Rescaling Theorem as learnable bijection.

Ax Mokshit Surana 5/15/2026

Fair and Calibrated Toxicity Detection with Robust Training and Abstention

Study on fairness and calibration in toxicity detection models using training interventions and safety mechanisms.

Ax Cristian Hinostroza, Rodrigo Toro Icarte, Christ Devia, Andres Carvallo De Ferari, Eugenio Herrera-Berg, Denis Parra, Jorge F Silva 5/15/2026

Rethinking Layer Relevance in Large Language Models Beyond Cosine Similarity

Research challenging cosine similarity as a metric for assessing layer relevance in LLMs, proposing alternatives for mechanistic interpretability.

Ax Mashrekur Rahman 5/15/2026

Mini-JEPA Foundation Model Fleet Enables Agentic Hydrologic Intelligence

Fleet of small foundation models for geospatial hydrologic analysis enabling accessible agentic environmental reasoning systems.

Ax Marius S. Knorr, Robert M\"uller, Jan P. Bremer, Nils Schweingruber 5/15/2026

Reinforcement Learning for Tool-Calling Agents in Fast Healthcare Interoperability Resources (FHIR)

Reinforcement learning approach for tool-calling LLM agents reasoning over healthcare FHIR resource graphs for clinical queries.

Ax Qi Cao, Yufan Wang, Peijia Qin, Shuhao Zhang, Pengtao Xie 5/15/2026

LLMs Know When They Know, but Do Not Act on It: A Metacognitive Harness for Test-time Scaling

Metacognitive framework leveraging LLM self-monitoring signals to control test-time inference and improve problem-solving.

Ax Leena Chennuru Vankadara, Moritz Haas, Luke Hayward, Sebastian Bordt, Alessandro Breccia 5/15/2026

How to Scale Mixture-of-Experts: From muP to the Maximally Scale-Stable Parameterization

Principled scaling rules for mixture-of-experts architectures analyzing hyperparameter relationships with network and expert dimensions.

Ax Andrew Lanpouthakoun, Aryaman Arora, Zhengxuan Wu, Dhruv Pai, Ben Keigwin, Dan Jurafsky, Christopher Potts 5/15/2026

PreFT: Prefill-only finetuning for efficient inference

Parameter-efficient finetuning method optimizing prefill-only updates to improve inference throughput for personalized LLMs.

Ax Tianle Zhong, Neiwen Ling, Yifan Pi, Zijun Wei, Tianshu Yu, Geoffrey Fox, Peng Wu, Xiao Yu 5/15/2026

Diagnosing Training Inference Mismatch in LLM Reinforcement Learning

Identifies and diagnoses training-inference mismatches in LLM reinforcement learning from implementation differences in token probabilities.

Ax Simranjeet Singh Dahia, Claudia Szabo 5/15/2026

Quantum Advantage in Multi Agent Reinforcement Learning

Empirical evaluation of quantum entanglement benefits in decentralized multi-agent reinforcement learning with variational quantum policies.

Ax Jerem\'ias Figueiredo Paschmann, Juan Kaplan, Francisco Nattero Santiago Mauricio Barron Bucolo, Juan Wisznia, Luciano del Corro 5/15/2026

Active Learners as Efficient PRP Rerankers

Active learning approach to improve pairwise ranking prompting reranking from LLMs with noisy and intransitive judgments.

Ax Andrii Shportko, Inessa Verbitsky 5/15/2026

Paraphrasing Attack Resilience of Various AI-Generated Text Detection Methods

Evaluation of AI-generated text detection methods' resilience to paraphrasing attacks across fine-tuned and classifier approaches.

Ax Kexin Chu, Dawei Xiang, Wei Zhang 5/15/2026

Latency-Quality Routing for Functionally Equivalent Tools in LLM Agents

Router for LLM agents selecting among functionally equivalent tool providers based on latency and quality tradeoffs.

Ax Zhiye Song, Kyungmi Lee, Eun Kyung Lee, Xin Zhang, Tamar Eilam, Anantha P. Chandrakasan 5/15/2026

EnergyLens: Predictive Energy-Aware Exploration for Multi-GPU LLM Inference Optimization

Framework for predicting and optimizing energy consumption during multi-GPU LLM inference without expensive profiling.

Ax Sibo Jia, Zihang Zhao, Kunrong Li 5/15/2026

Architecture-Aware Explanation Auditing for Industrial Visual Inspection

Auditing protocol to verify that deep neural network heatmap explanations faithfully reflect actual model decision drivers in industrial inspection.

Ax Jesseba Fernando, Grigori Guitchounts 5/15/2026

Dynamics of the Transformer Residual Stream: Coupling Spectral Geometry to Network Topology

Analysis of how computation propagates through transformer layers by studying residual stream dynamics and spectral geometry in LLMs.

Ax Weisen Jiang, Shuhao Chen, Sinno Jialin Pan 5/15/2026

MetaMoE: Diversity-Aware Proxy Selection for Privacy-Preserving Mixture-of-Experts Unification

MetaMoE unifies independently trained domain-specialized experts into privacy-preserving MoE using public proxy data for federated settings.

Ax Libo Sun, Po-wei Harn, Peixiong He, Xiao Qin 5/15/2026

Minimal-Intervention KV Retention: A Design-Space Study and a Diversity-Penalty Survivor

KV-cache compression study with diversity-penalty survivor method for efficient LLM inference on long-form reasoning tasks.

Ax Matias Alvo, Daniel Russo, Yash Kanoria 5/15/2026

Policy Optimization in Hybrid Discrete-Continuous Action Spaces via Mixed Gradients

Mixed gradient policy optimization for hybrid discrete-continuous action spaces in robotics and control problems.

Ax Qiyuan Chen, Jiayu Zhou, Raed Al Kontar 5/15/2026

Language-Induced Priors for Domain Adaptation

Domain adaptation framework leveraging expert textual descriptions as language-induced priors to prevent negative transfer in cold-start scenarios.

Ax Zuyuan Zhang, Carlee Joe-Wong, Tian Lan 5/15/2026

Matrix-Space Reinforcement Learning for Reusing Local Transition Geometry

Matrix-Space RL reuses local transition geometry through positive semidefinite matrix descriptors for compositional generalization.

Ax Yuchen Sun, Pei Fu, Shaojie Zhang, Anan Du, Xiuwen Xi, Ruoceng Zhang, Zhenbo Luo, Jian Luan, Chongyang Zhang 5/15/2026

Beyond Binary: Reframing GUI Critique as Continuous Semantic Alignment

Continuous semantic alignment framework for GUI critic models in test-time scaling for generalist AI agents beyond binary classification.

Ax Fangyuan Yu, Xin Su, Amir Abdullah 5/15/2026

Dynamic Latent Routing

Dynamic Latent Routing composes optimal sub-policies temporally for MDPs and proposes LLM post-training method based on General Dijkstra Search.

Ax Yerin Park, Sangseon Lee 5/15/2026

AIM-DDI: A Model-Agnostic Multimodal Integration Module for Drug-Drug Interaction Prediction

AIM-DDI predicts drug-drug interactions using model-agnostic multimodal integration with focus on unseen-drug generalization.

Ax Jessica Rumbelow 5/15/2026

Exemplar Partitioning for Mechanistic Interpretability

Exemplar Partitioning constructs interpretable feature dictionaries from LLM activations using Voronoi partitioning with 1000x fewer tokens than sparse autoencoders.

Ax Nicholas E. Corrado, Wenyuan Huang, Josiah P. Hanna 5/15/2026

Distributionally Robust Multi-Task Reinforcement Learning via Adaptive Task Sampling

Distributionally robust multi-task reinforcement learning addresses imbalanced learning across tasks via adaptive task sampling.

Ax Zhengjia Zhong, Shuyan Ke, Zaizhou Lin, Jiaqi Song, Hongyi Lan, Hui Li 5/15/2026

RQ-MoE: Residual Quantization via Mixture of Experts for Efficient Input-Dependent Vector Compression

RQ-MoE combines residual quantization with mixture-of-experts for efficient input-dependent vector compression of embeddings.

Ax Jiaqi Sun, Boyang Sun, Mohamad Rasmy, Xiangchen Song, Kun Zhang 5/15/2026

MoRe: Modular Representations for Principled Continual Representation Learning on Squantial Data

MoRe proposes modular representations for continual learning on sequential data with minimal catastrophic forgetting.

Ax Changryeol Choi, Hyewon Park, Yujin Kwon, Gowun Jeong 5/15/2026

LoMETab: Beyond Rank-1 Ensembles for Tabular Deep Learning

LoMETab extends rank-1 ensembles to rank-r multiplicative factorizations for improved tabular deep learning performance.

Ax Chen Liang, Xiatao Sun, Qian Wang, Daniel Rakita 5/15/2026

Turning Stale Gradients into Stable Gradients: Coherent Coordinate Descent with Implicit Landscape Smoothing for Lightweight Zeroth-Order Optimization

Coherent Coordinate Descent optimizes zeroth-order optimization for memory-constrained on-device learning without backpropagation.

Ax Young-Chae Hong, Yangho Chen 5/15/2026

Optimal Pattern Detection Tree for Symbolic Rule-Based Classification

Optimal Pattern Detection Tree introduces interpretable rule-based classification model for pattern discovery in healthcare and maintenance domains.

Ax JB Lanier, Nathan Monette, Pierre Baldi, Roy Fox 5/15/2026

Data-Augmented Game Starts for Accelerating Self-Play Exploration in Imperfect Information Games

Multi-agent exploration strategy for imperfect-information games like StarCraft using data-augmented game starts to accelerate policy gradient learning.

Ax Qazi Mamunur Rashid, Xuan Yang, Zhengzhe Yang, Yanzhou Pan, Erin van Liemt, Darlene Neal, Kshitij Pancholi, Jamila Smith-Loud 5/15/2026

NodeSynth: Socially Aligned Synthetic Data for AI Evaluation

NodeSynth generates socially aligned synthetic data for AI model evaluation using a fine-tuned taxonomy generator anchored in real-world evidence.

Ax Donghwan Kim, Hyunsoo Yoon 5/15/2026

MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse

Novel OOD detection method using class-wise Mahalanobis distance variance for neural networks in safety-critical applications.

Ax Shuqi Gu, Yongxiang Zhao, Baoyu Jing, Kan Ren 5/15/2026

What if Tomorrow is the World Cup Final? Counterfactual Time Series Forecasting with Textual Conditions

Counterfactual time series forecasting method incorporating textual conditions for future events influence.

Ax Leo Muxing Wang, Pengkun Yang, Lili Su 5/15/2026

Collaborative Yet Personalized Policy Training: Single-Timescale Federated Actor-Critic

Federated actor-critic framework for collaborative policy training with shared representations and personalized local policies.

Ax Runyuan He, Qiuyang Mang, Shang Zhou, Kaiyuan Liu, Hanchen Li, Huanzhi Mao, Qizheng Zhang, Zerui Li, Bo Peng, Lufeng Cheng, Tianfu Fu, Yichuan Wang, Wenhao Chai, Jingbo Shang, Alex Dimakis, Joseph E. Gonzalez, Alvin Cheung 5/15/2026

FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale

FrontierSmith: Method for synthesizing open-ended coding problems at scale to train stronger LLM coders.

Ax Siyang Yao, Erhu Feng, Yubin Xia 5/15/2026

When Answers Stray from Questions: Hallucination Detection via Question-Answer Orthogonal Decomposition

QAOD: Single-pass white-box framework for hallucination detection in LLMs using question-answer orthogonal decomposition.

Ax Minbeom Kim, Lesly Miculicich, Bhavana Dalvi Mishra, Mihir Parmar, Phillip Wallis, Bharath Chandrasekhar, Kyomin Jung, Tomas Pfister, Long T. Le 5/15/2026

LiSA: Lifelong Safety Adaptation via Conservative Policy Induction

LiSA: lifelong safety adaptation framework for AI agents executing workflows, maintaining guardrails against contextual safety failures in tool use and data access.

Ax Elias Zavitsanos, Georgios Paliouras 5/15/2026

Focused PU learning from imbalanced data

Method for positive-unlabeled learning from highly imbalanced datasets, applicable to disease identification, fraud detection, and recommender systems.

Ax Weijia Xu, Alessandro Sordoni, Chandan Singh, Zelalem Gero, Michel Galley, Xingdi Yuan, Jianfeng Gao 5/15/2026

Test-Time Learning with an Evolving Library

EvoLib: test-time learning framework enabling LLMs to accumulate and evolve knowledge across problem instances via extracted modular skills without parameter updates.

Ax Letian Yang (Shanghai Jiao Tong University, Shanghai, China), Xu Liu (Shanghai Jiao Tong University, Shanghai, China), Yiqiang Lu (Ant Group, Shanghai, China), Jian Liu (Ant Group, Shanghai, China), Weiqiang Wang (Ant Group, Shanghai, China), Shuai Li (Shanghai Jiao Tong University, Shanghai, China) 5/15/2026