Isolater - Feed

Ax Daniel Geyfman, Felix Draxler, Jan Groeneveld, Hyunsoo Lee, Theofanis Karaletsos, Stephan Mandt 7/2/2026

Calibrated Test-Time Guidance for Bayesian Inference

Analysis of test-time guidance in diffusion models showing common methods miscalibrate Bayesian inference; proposes correction for posterior sampling.

Ax Aviral Chawla, Galen Hall, Juniper Lovato 7/2/2026

MetaOthello: A Controlled Study of Multiple World Models in Transformers

MetaOthello controlled study examining how transformers organize multiple world models across different generative processes using Othello game variants.

Ax Theo X. Olausson, Jo\~ao Monteiro, Michal Klein, Marco Cuturi 7/2/2026

Amortized Maximum Inner Product Search with Learned Support Functions

Amortized maximum inner product search: neural networks trained to directly predict MIPS solutions, amortizing costs for repeated queries on fixed databases.

Ax Hyunji Nam, Haoran Li, Natasha Jaques 7/2/2026

Maximizing Mutual Information Between Prompt and Response Improves LLM Performance With No Additional Data

Self-improvement framework maximizing mutual information between prompts and LLM responses without additional labeled data or external verifiers.

Ax Lala Shakti Swarup Ray, Mengxi Liu, Alcina Pinto, Deepika Gurung, Daniel Geissler, Paul Lukowoicz, Bo Zhou 7/2/2026

ActivityNarrated: An Open-Ended Narrative Paradigm for Wearable Human Activity Understanding

ActivityNarrated dataset for open-ended narrative-based human activity recognition from wearables, replacing fixed-window classification benchmarks.

Ax Tin Had\v{z}i Veljkovi\'c, Joshua Rosenthal, Ivor Lon\v{c}ari\'c, Jan-Willem van de Meent 7/2/2026

Crystalite: A Lightweight Transformer for Efficient Crystal Modeling

Crystalite: lightweight diffusion transformer for crystal material modeling using subatomic tokenization and equivariant inductive biases.

Ax Yiping Li, Zhiyu An, Wan Du 7/2/2026

When Less Latent Leads to Better Relay: Information-Preserving Compression for Latent Multi-Agent LLM Collaboration

Orthogonal BackFill method for compressing KV-cache communication in multi-agent LLM systems, reducing memory and communication costs while preserving information.

Ax Yangyue Wang, Harshvardhan Sikka, Yash Mathur, Tony Zhou, Jinu Nyachhyon, Pranav Guruprasad 7/2/2026

GUI-Perturbed: Domain Randomization Reveals Systematic Brittleness in GUI Grounding Models

GUI-Perturbed framework reveals brittleness in GUI grounding models through domain randomization, showing 27-56% accuracy drops on spatial reasoning tasks.

Ax Minhak Song, Liang Zhang, Bingcong Li, Niao He, Michael Muehlebach, Sewoong Oh 7/2/2026

Zeroth-Order Optimization at the Edge of Stability

Zeroth-order optimization methods for gradient-free black-box learning and memory-efficient LLM fine-tuning, analyzing stability dynamics.

Ax Sajjad Ghiasvand, Mark Beliaev, Mahnoosh Alizadeh, Ramtin Pedarsani 7/2/2026

REALM: Reliable Expertise-Aware Language Model Fine-Tuning from Noisy Annotations

REALM method for fine-tuning LLMs with noisy crowdsourced annotations by jointly learning model parameters and annotator expertise weights.

Ax Changyu Li, Shuanghong Huang, Jiashen Liu, Ming Lei, Jidu Xing, Kaishun Wu, Lu Wang, Fei Luo 7/2/2026

FED-FSTQ: Fisher-Guided Token Quantization for Communication-Efficient Federated Fine-Tuning of LLMs on Edge Devices

Fisher-guided token quantization reduces communication overhead in federated fine-tuning of LLMs on edge devices.

Ax Hongmin Li 7/2/2026

FastUMAP: Scalable Dimensionality Reduction via Bipartite Landmark Sampling

FastUMAP landmark-based dimensionality reduction method optimized for repeated analysis with changing hyperparameters.

Ax Qijun Hou, Yuchen Shi, Pingyi Fan, Khaled B. Letaief 7/2/2026

Federated Client Selection under Partial Visibility: A POMDP Approach with Spatio-Temporal Attention

POMDP approach with spatio-temporal attention for federated learning client selection under partial visibility constraints.

Ax Markus Herre, Andrej Tschalzev, Sascha Marton, Christian Bartelt 7/2/2026

Explaining Tabular Foundation Model Differences Through Meta-Features

Meta-feature analysis explains performance gaps between tabular foundation models and traditional models on prediction tasks.

Ax Luca Zhou 7/2/2026

Tail-Shape Estimation in LLM Evaluation Is Fragile: A Protocol for Diagnosing False Positives

Protocol for diagnosing false positives in tail-aware LLM evaluation metrics beyond mean-based assessment.

Ax Yingshuo Wang, Xian Sun, Lingdong Kong, Wei Gao, Yanhang Li, Zhichao Fan, Zexin Zhuang 7/2/2026

Do Time Series Foundation Model Benchmarks Hide Regime-Dependent Failures? Evidence from Traffic Speed Forecasting

Regime-stratified evaluation reveals hidden failures in time series foundation models masked by standard aggregate metrics.

Ax Borui Cai, Yao Zhao 7/2/2026

Topological Neural Dynamics: A Neuron-wise Framework for Sequence Modeling

Neuron-wise sequence modeling framework allowing independent evolution of neurons instead of layer-wise shared dynamics.

Ax Muhammad Ahmed 7/2/2026

PersistentKV: Page-Aware Decode Scheduling for Long-Context LLM Serving on Commodity GPUs

PersistentKV optimizes decode scheduling for long-context LLM serving on GPUs through page-aware KV cache management.

Ax Rajat Ghosh, Datta Nimmaturi, Aryan Singhal, Vaishnavi Bhargava, Henry Wong, Johnu George, Debojyoti Dutta 7/2/2026

Predictable GRPO: A Closed-Form Model of Training Dynamics

Closed-form reduced-order model of GRPO training dynamics using mean-field approximation and stochastically-forced oscillator analogy.

Ax Rui Zhou, Tianci Xie 7/2/2026

Fora: From Weight-Space to Function-Space Protection in Capability-Preserving Fine-Tuning

Fora protects LLM capabilities during fine-tuning by preserving activation subspaces rather than just parameter distances.

Ax Abhishek Dey 7/2/2026

ComplianceGate: Classifier-Gated Multi-Tier LLM Routing for Inference in Regulated Industries

ComplianceGate routes LLM queries through multi-tier classifier system for compliance enforcement and cost efficiency in regulated industries.

Ax Kung-Ming Lan, Edward Huang 7/2/2026

Beyond the Expressivity-Trainability Paradox: A Dynamical Lie Algebra Perspective on Navigating Barren Plateaus in Quantum Machine Learning

Analysis of barren plateaus in quantum machine learning using Lie algebra perspective to address expressivity-trainability tradeoffs.

Ax Tianyi Ma, Tengyao Wang, Richard J. Samworth 7/2/2026

Deep learning with missing data

Pattern Embedded Neural Networks (PENNs) for multivariate regression with missing covariates, combining imputation with indicator observation networks.

Ax Ran Yan, Youhe Jiang, Zhuoming Chen, Haohui Mai, Beidi Chen, Binhang Yuan 7/2/2026

FSA: An Alternative Efficient Implementation of Native Sparse Attention Kernel

Efficient sparse attention kernel implementation for LLMs, providing alternative to Native Sparse Attention with improved performance on long-context tasks.

Ax Juyeon Yoon, Somin Kim, Robert Feldt, Shin Yoo 7/2/2026

Clotho: Measuring Task-Specific Pre-Generation Test Adequacy for LLM Inputs

Test adequacy metric (Clotho) for measuring LLM input quality before inference to enable task-specific prompt testing without ground truth outputs.

Ax Lingxiao Li, Yifan Wang, Xinyan Gao, Chen Tang, Xiangyu Yue, Chenyu You 7/2/2026

VisReason: A Large-Scale Dataset for Visual Chain-of-Thought Reasoning

Large-scale dataset of 400K visual chain-of-thought reasoning examples for training multimodal large language models on spatially grounded reasoning tasks.

Ax Yuxi Liu, Yipeng Hu, Zekun Zhang, Kunze Jiang, Kun Yuan 7/2/2026

Mixture of Distributions Matters: Dynamic Sparse Attention for Efficient Video Diffusion Transformers

Dynamic sparse attention mechanism for video diffusion transformers to reduce quadratic complexity of self-attention in long-sequence generation tasks.

Ax Sajjad Ghiasvand, Haniyeh Ehsani Oskouie, Mahnoosh Alizadeh, Ramtin Pedarsani 7/2/2026

MMLoP: Multi-Modal Low-Rank Prompting for Efficient Vision-Language Adaptation

MMLoP low-rank prompting method for efficient vision-language model adaptation with millions fewer parameters than state-of-the-art.

Ax Zhanyi Sun, Shuran Song 7/2/2026

From Prior to Pro: Efficient Skill Mastery via Distribution Contractive RL Finetuning

DICE-RL framework using RL as distribution contraction operator to finetune pretrained diffusion/flow-based robot policies from online feedback.

Ax Rasul Khanbayov, Mohamed Rayan Barhdadi, Erchin Serpedin, Hasan Kurban 7/2/2026

IRIS: A Real-World Benchmark for Inverse Recovery and Identification of Physical Dynamic Systems from Monocular Video

IRIS benchmark: 240 4K real-world videos for unsupervised physical parameter estimation and governing-equation identification.

Ax Joanna Zou, Han Cheng Lie, Youssef Marzouk 7/2/2026

Goal-oriented learning of stochastic differential equations using error bounds on path-space observables

Surrogate modeling framework for stochastic differential equations using path-space observable error bounds for efficient simulation.

Ax Zhanzhao Li, Kengran Yang, Qiyao He, Kai Gong 7/2/2026

Large language model-enabled automated data extraction for concrete materials informatics

LLM-powered pipeline for automated extraction and structuring of materials science data from unstructured scientific literature.

Ax Alasdair Ross, George K. Holt, Kamran Pentland, Adriano Agnello, Nicola C. Amorisco, Pedro Cavestany, Aran Garrod, Timothy Nunn, Charles Vincent, Graham McArdle 7/2/2026

Real-time virtual circuits for plasma shape control via neural network emulators

Neural network emulators for real-time tokamak plasma shape control via virtual circuits; replaces expensive numerical computation.

Ax Itay Lavie, Noam Levi, Yonatan Kahn 7/2/2026

Statistical Properties of Training & Generalization

Physics-informed review of deep learning statistical properties including neural scaling laws from classical statistics perspective.

Ax Peng Fang, Arijit Khan, Ziqiang Wu, Zhenli Li, Yibo Zhou, Fang Wang, Dan Feng 7/2/2026

FeLoG: Scalable and Efficient Distributed Graph Embedding with Feedback Loop Mechanism

FeLoG distributed graph embedding framework with feedback loop mechanism for billion-scale graphs; applications include GraphRAG.

Ax Sana Ayromlou, Purvi Sehgal, Pradyumna Narayana 7/2/2026

Diagnosing and Mitigating Compounding Failures in Agentic Persuasion via Taxonomic Strategy Retrieval

Taxonomic strategy retrieval approach to mitigate compounding failures in multi-step agentic tasks; addresses problem drift in persuasion agents.

Ax Matteo Farina, Vishaal Udandarao, Thao Nguyen, Selim Kuzucu, Maximilian B\"other, Andreas Hochlehnert, Adhiraj Ghosh, Marianna Nezhurina, Karsten Roth, Joschka Struber, Yuhui Zhang, Sebastian Dziadzio, Elaine Sui, Soumya Jahagirdar, Dhruba Ghosh, Hasan Hammoud, Thomas De Min, Simone Caldarella, Jehanzeb Mirza, Sedrick Keh, Mehdi Cherti, Hilde Kuehne, Bernt Schiele, Serena Yeung-Levy, Muhammad Ferjad Naeem, Federico Tombari, Ana Klimovic, Elisa Ricci, Matthias Bethge, Sewoong Oh, Ameya Prabhu, Alessio Tonioni, Jenia Jitsev, Massimiliano Mancini, Ludwig Schmidt, Nikhil Parthasarathy 7/2/2026