Isolater - Feed

Ax Daniel Goldstein, Eugene Cheah 5/18/2026

Key-Value Means: Transformers with Expandable Block-Recurrent Compressed Memory

Key-Value Means introduces block-recurrent attention with expandable compressed memory for O(N) transformers with long-context capability.

Ax Gabriel Garcia 5/18/2026

The Last Word Often Wins: A Format Confound in Chain-of-Thought Corruption Studies

Reveals format confound in chain-of-thought corruption studies where answer placement rather than computation determines faithfulness.

Ax Daniel Ranard 5/18/2026

Likelihood scoring for continuations of mathematical text: a self-supervised benchmark with tests for shortcut vulnerabilities

Self-supervised benchmark for evaluating LLM continuation prediction on hidden mathematical text with shortcut vulnerability tests.

Ax Yuanda Xu, Hejian Sang, Zhengze Zhou, Ran He, Zhipeng Wang, Alborz Geramifard 5/18/2026

Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training

Four-stage post-training workflow for LLM reasoning combining sparse-reward RL, KL warmup, and distillation for effective data allocation.

Ax Nirmal Patel, Fei Wang, Inderjit S. Dhillon 5/18/2026

ODRPO: Ordinal Decompositions of Discrete Rewards for Robust Policy Optimization

ODRPO decomposes discrete rewards for robust LLM alignment via RLAIF, addressing stochasticity in LLM-based auto-raters.

Ax Anastasiia Sedova, Skyler Seto, Natalie Schluter, Pierre Ablin 5/18/2026

Scaling Laws for Mixture Pretraining Under Data Constraints

Studies optimal mixture ratios of scarce target data with abundant generic data during LLM pre-training under data constraints.

Ax Eszter Varga-Umbrich, Zachary Weller-Davies, Paul Duckworth, Jules Tilly, Olivier Peltre, Shikha Surana 5/18/2026

Force-Aware Neural Tangent Kernels for Scalable and Robust Active Learning of MLIPs

Active learning framework for machine-learning interatomic potentials using neural tangent kernels with force-aware acquisition.

Ax William Lehn-Schi{\o}ler, Magnus Ruud Kj{\ae}r, Rahul Thapa, Magnus Guldberg Pedersen, Anton Mosquera Storgaard, Nick Williams, Radu Gatej, Tue Lehn-Schi{\o}ler, S\'andor Beniczky, Sadasivan Puthusserypady, James Zou, Lars Kai Hansen 5/18/2026

Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoders

Applies sparse autoencoders to extract interpretable features from EEG foundation models for clinical applications.

Ax Jerem\'ias Figueiredo Paschmann, Juan Kaplan, Francisco Nattero, Santiago Barron, Juan Wisznia, Luciano del Corro 5/18/2026

Active Learners as Efficient PRP Rerankers

Uses active learning to efficiently rerank LLM pairwise preference judgments, treating it as robust top-K recovery rather than sorting.

Ax Yuchen Sun, Pei Fu, Shaojie Zhang, Anan Du, Xiuwen Xi, Ruoceng Zhang, Zhenbo Luo, Jian Luan, Chongyang Zhang 5/18/2026

Beyond Binary: Reframing GUI Critique as Continuous Semantic Alignment

Reframes GUI agent critic models from binary classification to continuous semantic alignment for better ranking of candidate actions.

Ax Magdalena Proszewska, N. Siddharth 5/18/2026

AIMing for Standardised Explainability Evaluation in GNNs: A Framework and Case Study on Graph Kernel Networks

Framework for standardized explainability evaluation in Graph Neural Networks with case study on graph kernel networks.

Ax Anurup Ganguli 5/18/2026

TFGN: Task-Free, Replay-Free Continual Pre-Training Without Catastrophic Forgetting at LLM Scale

TFGN enables continual pre-training of LLMs on heterogeneous domains without replay buffers or task labels, preventing catastrophic forgetting.

Ax Will Schwarzer, Scott Niekum 5/18/2026

Training ML Models with Predictable Failures

Framework for predicting deployment-scale ML model failure rates by extrapolating from largest failure scores in evaluation sets.

Ax Ravidu Suien Rammuni Silva, Jordan J. Bird 5/18/2026

FM-G-CAM: A Holistic Approach for Explainable AI in Computer Vision

FM-G-CAM extends Grad-CAM for explainability in CNN predictions with holistic approach to understanding computer vision models.

Ax Ziyu Liu, Zeyi Sun, Yuhang Zang, Wei Li, Pan Zhang, Xiaoyi Dong, Yuanjun Xiong, Dahua Lin, Jiaqi Wang 5/18/2026

RAR: Retrieving And Ranking Augmented MLLMs for Visual Recognition

RAR combines CLIP and MLLMs for improved visual recognition by retrieving and ranking candidates for fine-grained classification.

Ax ChonLam Lao, Jiaqi Gao, Jiamin Cao, Zhipeng Zhang, Pengcheng Zhang, Jiangfei Duan, Zhilong Zheng, Yu Guan, Yichi Xu, Yong Li, Zhengping Qian, Aditya Akella, Minlan Yu, Ennan Zhai, Dennis Cai, Jingren Zhou 5/18/2026

TrainMover: An Interruption-Resilient Runtime for ML Training

TrainMover is a resilient runtime for LLM training that handles hardware/software interruptions using elastic machines with minimal downtime.

Ax Atsushi Suzuki, Yulan He, Feng Tian, Zhongyuan Wang 5/18/2026

Hallucinations are inevitable but can be made statistically negligible

Theoretical analysis proving hallucinations in LMs are computationally inevitable but can be made statistically negligible through appropriate methods.

Ax Yash Akhauri, Ahmed F AbouElhamayed, Yifei Gao, Chi-Chih Chang, Sameh Gobriel, Nilesh Jain, Mohamed S. Abdelfattah 5/18/2026

TokenButler: Token Importance is Predictable

TokenButler predicts which tokens are important for KV-cache in LLMs to reduce memory and computation bottlenecks during decoding.

Ax Zihan Zhang, Leon Wong, Blesson Varghese 5/18/2026

FedOptima: Optimizing Resource Utilization in Federated Learning

FedOptima optimizes resource utilization in federated learning systems by addressing task dependencies and stragglers across heterogeneous devices.

Ax Diksha Goel, Kristen Moore, Jeff Wang, Minjune Kim, Thanh Thi Nguyen 5/18/2026

Unveiling the Black Box: A Multi-Layer Framework for Explaining Reinforcement Learning-Based Cyber Agents

Multi-layer explainability framework for RL-based cyber attack agents, providing interpretability for adversarial strategy formation and evolution.

Ax Jinyang Wu, Chonghua Liao, Mingkuan Feng, Shuai Zhang, Zhengqi Wen, Haoran Luo, Ling Yang, Huazhe Xu, Jianhua Tao 5/18/2026

TemplateRL: Structured Template-Guided Reinforcement Learning for LLM Reasoning

TemplateRL applies structured template-guided RL to improve LLM reasoning by augmenting policy optimization with explicit problem-solving strategies.

Ax Yuxin Ren, Maxwell D Collins, Miao Hu, Huanrui Yang 5/18/2026

FAR: Function-preserving Attention Replacement for IMC-friendly Inference

FAR framework replaces attention mechanisms with in-memory computing friendly operations for efficient transformer inference on ReRAM accelerators.

Ax Sharique Hasan, Alexander Oettl, Sampsa Samila 5/18/2026

From Model Design to Organizational Design: Complexity Redistribution and Trade-Offs in Generative AI

Introduces GAS framework analyzing trade-offs between generality, accuracy, and simplicity in LLM-driven organizational redesign and strategy.

Ax Shengji Tang, Jianjian Cao, Weihao Lin, Jiale Hong, Bo Zhang, Shuyue Hu, Lei Bai, Tao Chen, Wanli Ouyang, Peng Ye 5/18/2026

A Scalable Multi-LLM Collaboration System with Retrieval-based Selection and Exploration-Exploitation-Driven Enhancement

SMCS system coordinates multiple open-source LLMs using retrieval-based selection and exploration-exploitation for scalable multi-LLM collaboration.

Ax Valeria Pais, Luis Oala, Daniele Faccio, Marco Aversa 5/18/2026

Autoguided Online Data Curation for Diffusion Model Training

Integrates data curation methods into unified framework for training diffusion models more efficiently using autoguidance and online selection.

Ax Sasha Cui, Zhongren Chen 5/18/2026

Painless Activation Steering: An Automated, Lightweight Approach for Post-Training Large Language Models

Automated lightweight method for steering LLM behavior via activation manipulation instead of weight updates or prompting, enabling fast post-training control.

Ax El Mahdi Chayti, Martin Jaggi 5/18/2026

A Split-Client Approach to Second-Order Optimization

Proposes split-client approach to reduce wall-clock cost of Hessian computation in second-order optimization for moderate-dimensional problems.

Ax Pascal Bergstr\"a{\ss}er, Ryan Cotterell, Anthony W. Lin 5/18/2026

Transformers are Inherently Succinct

Theoretical analysis proving transformers can be exponentially more succinct than LTL and RNNs in expressing languages, using classical automata theory.

Ax Mohammed Hilel, Yannis Karmim, Jean De Bodinat, Reda Sarehane, Antoine Gillon 5/18/2026

Building Specialized Software-Assistant ChatBot with Graph-Based Retrieval-Augmented Generation

Graph-based RAG system for building specialized software-assistant chatbots to help enterprise application users.

Ax Augusto B. Corr\^ea, Andr\'e G. Pereira, Jendrik Seipp 5/18/2026

Frontier Large Language Models Rival State-of-the-Art Planners

Frontier LLMs (Claude, GPT-4, Gemini) demonstrate state-of-the-art planning capabilities on International Planning Competition benchmarks.

Ax Rulin Shao, Akari Asai, Shannon Zejiang Shen, Hamish Ivison, Varsha Kishore, Jingming Zhuo, Xinran Zhao, Molly Park, Samuel G. Finlayson, David Sontag, Tyler Murray, Sewon Min, Pradeep Dasigi, Luca Soldaini, Faeze Brahman, Wen-tau Yih, Tongshuang Wu, Luke Zettlemoyer, Yoon Kim, Hannaneh Hajishirzi, Pang Wei Koh 5/18/2026