Isolater - Feed

HN carlual 7/14/2026

Show HN: ZenStack – access control at the ORM layer, built for coding agents

ORM-layer access control system designed for AI coding agents with RBAC/ABAC support.

Ax Maxime Heuillet, Yufei Cui, Boxing Chen, Audrey Durand, Prasanna Parthasarathi 7/14/2026

Nested-ReFT: Efficient Reinforcement Learning for Large Language Model Fine-Tuning via Off-Policy Rollouts

Efficient RL fine-tuning for LLMs using off-policy rollouts to reduce computational cost in verifiable reward-based reinforcement learning for reasoning tasks.

Ax Francesco Emanuele Stradi, Eleonora Fidelia Chiefari, Matteo Castiglioni, Alberto Marchesi, Nicola Gatti 7/14/2026

Beyond Slater's Condition in Online CMDPs with Stochastic and Adversarial Constraints

Algorithm for online constrained MDPs achieving improved regret bounds under stochastic and adversarial constraints beyond Slater's condition.

Ax Yunhao Liang, Pujun Zhang, Yuan Qu, Jingyuan Yang, Shaochong Lin, Zuo-jun Max Shen 7/14/2026

Graph Optimization Foundation Model: Tokenizing Graph via A Language-Model Paradigm

Foundation model for graph optimization using language model paradigm to handle OR problems on graph structures while managing combinatorial constraints.

Ax Pengxiao Lin, Zheng-An Chen, Zhi-Qin John Xu 7/14/2026

Unveiling the Mechanisms of Multi-Hop Reasoning in Transformers via Identity Bridge

Identifies missing bridge entity supervision as cause of multi-hop reasoning failure in transformers; proposes identity bridge supervision improving out-of-distribution composition.

Ax Yuchen Zhu, Wei Guo, Jaemoo Choi, Petr Molodyk, Bo Yuan, Molei Tao, Yongxin Chen 7/14/2026

Enhancing Reasoning for Diffusion LLMs via Distribution Matching Policy Optimization

Distribution Matching Policy Optimization: RL algorithm designed for diffusion LLMs enabling reasoning tasks with higher inference throughput than autoregressive models.

Ax Patrick Pynadath, Jiaxin Shi, Ruqi Zhang 7/14/2026

CANDI: Hybrid Discrete-Continuous Diffusion Models

CANDI: hybrid diffusion model for discrete and continuous data, analyzing token corruption mechanisms through identity and rank degradation.

Ax Jatin Prakash, Aahlad Puli, Rajesh Ranganath 7/14/2026

Controllably Efficient Language Models

Framework enabling transformers to trade off inference efficiency and quality dynamically, controlling sparse/linear attention and convolutions per layer.

Ax Zhuoyun Du, Runze Wang, Huiyu Bai, Zouying Cao, Xiaoyong Zhu, Yu Cheng, Bo Zheng, Wei Chen, Haochao Ying 7/14/2026

Enabling Agents to Communicate Entirely in Latent Space

Interlat: enables LLM-based agents to communicate directly in latent space, bypassing discrete tokens for richer information exchange in collaborative problem-solving.

Ax Quan Liu, Feng Ye, Chenhao Lu, Shuming Zhen, Guanliang Huang, Lunzhe Chen, Xudong Ke 7/14/2026

Enhancing Adversarial Transferability through Block Stretch and Shrink

Block Stretch and Shrink method improving adversarial attack transferability through implicit ensemble perspective on model frontend responses.

Ax Songqiao Su, Xiaoya Li, Albert Wang, Guoyin Wang, Jiwei Li, Chris Shum 7/14/2026

CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learning

CUDA-L2: combines LLMs and RL to automatically optimize HGEMM CUDA kernels, outperforming cuBLAS and torch.matmul on matrix multiplication.

Ax Ijun Jang, Jewon Yeom, Juan Yeo, Hyunggyu Lim, Taesup Kim 7/14/2026

Stable On-Policy Distillation through Adaptive Target Reformulation

On-policy knowledge distillation method stabilizing training by reformulating targets to address distribution mismatch between student and teacher LLMs.

Ax Pengyang Shao, Naixin Zhai, Lei Chen, Yonghui Yang, Fengbin Zhu, Xun Yang, Meng Wang 7/14/2026

BalDRO: A Distributionally Robust Optimization based Framework for Large Language Model Unlearning

BalDRO: distributionally robust optimization framework for LLM unlearning addressing sample-wise imbalance in forgetting difficulty.

Ax Khalid Oublal, Quentin Bouniot, Qi Gan, Stephan Cl\'emen\c{c}on, Zeynep Akata 7/14/2026

TimeSAE: Causal Sparse Decoding for Faithful Explanations of Black-Box Time Series Models

TimeSAE: sparse decoding method providing causal explanations for black-box time series models with out-of-distribution generalization.

Ax Oliver Wei{\ss}l, Vincenzo Riccio, Severin Kacianka, Andrea Stocco 7/14/2026

HyperNet-Adaptation for Diffusion-Based Test Case Generation

HyperNet-Adaptation framework using diffusion models for generating realistic test cases for deep learning systems beyond simple adversarial perturbations.

Ax Noam Koren, Rafael Moschopoulos, Kira Radinsky, Elad Hazan 7/14/2026

SFO: Learning PDE Operators via Spectral Filtering

Spectral Filtering Operator (SFO): neural operator for solving PDEs using Universal Spectral Basis to capture long-range nonlocal interactions.

Ax Juntao Fang, Shifeng Xie, Shengbin Nie, Yuhui Ling, Yuming Liu, Zijian Li, Keli Zhang, Lujia Pan, Themis Palpanas, Ruichu Cai 7/14/2026

Rethinking Zero-Shot Time Series Classification: From Task-specific Classifiers to In-Context Inference

TIC-FM: in-context learning framework for zero-shot time series classification using foundation models without task-specific training.

Ax Mingyue Cheng, Yaguo Liu, Daoyu Wang, Xiaoyu Tao, Qi Liu 7/14/2026

CoGenCast: A Coupled Autoregressive-Flow Generative Framework for Time Series Forecasting

CoGenCast: generative framework combining autoregressive LLMs and flow models for time series forecasting with semantic and stochastic modeling.

Ax Sudipto Ghosh, Sujoy Nath, Sunny Manchanda, Tanmoy Chakraborty 7/14/2026

Disentangling Intrinsic Importance from Emergent Structure in Multi-Expert Orchestration

INFORM: interpretability framework for analyzing multi-LLM orchestration policies, decoupling expert interaction structure from intrinsic importance in collaborative reasoning systems.

Ax Shizhou Xu 7/14/2026

Separation-Utility Pareto Frontier: An Information-Theoretic Characterization

Information-theoretic analysis of the trade-off between utility and fairness (separation) in predictive models, proving Pareto frontier concavity.

Ax David Chanin, Adri\`a Garriga-Alonso 7/14/2026

SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data

Benchmark and toolkit for evaluating sparse autoencoders using large-scale synthetic data with realistic features.

Ax Bowen Yu, Sheng Zhang, Binhao Wang, Yi Wen, Jingtong Gao, Bowen Liu, Zimo Zhao, Shanshan Ye, Wanyu Wang, Maolin Wang, Xiangyu Zhao 7/14/2026

BRIDGE: Bridging Reasoning In Distillation Gap Elimination via Structure-Aware Masking

Knowledge distillation method for compressing chain-of-thought reasoning from large LLMs to smaller models using structure-aware masking.

Ax Mohan Tang, Sidi Lu 7/14/2026

Turbo Connection: Reasoning as Information Flow from Higher to Lower Layers

Transformer architecture extension enabling multi-step reasoning by allowing information flow across layers, addressing fixed-depth computation limits.

Ax Adhyyan Narang, Sarah Dean, Lillian J Ratliff, Maryam Fazel 7/14/2026

Dynamics of Learning under User Choice: Overspecialization and Peer-Model Probing

Analysis of machine learning platform dynamics when users select between competing learners, studying convergence and model specialization.

Ax Jay Revolinsky, Harry Shomer, Jiliang Tang 7/14/2026

Local Message-Passing for Discrete Graph Generation

GenGNN message-passing backbone for discrete graph generation using coupled node-edge refinement without graph transformers.

Ax Kadir-Kaan \"Ozer, Ren\'e Ebeling, Markus Enzweiler 7/14/2026

ECoLAD: Selecting Anomaly Detectors for Automotive Deployment via Compute-Reduction Evaluation

Deployment-focused evaluation protocol for selecting time-series anomaly detectors suitable for automotive applications.

Ax Jiaheng Hu, Jay Shim, Chen Tang, Yoonchang Sung, Bo Liu, Peter Stone, Roberto Martin-Martin 7/14/2026

Simple Recipe Works: Vision-Language-Action Models are Natural Continual Learners with Reinforcement Learning

Demonstrates vision-language-action models naturally handle continual learning with sequential fine-tuning and RL.

Ax Robert Welch, Emir Konuk, Kevin Smith 7/14/2026

The Cost of Reasoning: Chain-of-Thought Induces Overconfidence in Vision-Language Models

Shows chain-of-thought reasoning in vision-language models degrades uncertainty quantification reliability.

Ax Ivan Pasichnyk 7/14/2026

Critical Damping as a Momentum Schedule: Multi-Seed Validation, a Hybrid Recipe, and an Exhaustive Negative Result on Surgical Layer Selection

Critical damping-based momentum schedule for SGD achieving faster convergence without hyperparameter tuning.

Ax Hui Sun, Yun-Ji Zhang, Zheng Xie, Ren-Biao Liu, Yali Du, Xin-Ye Li, Ming Li 7/14/2026

ACES: Who Tests the Tests? Leave-One-Out AUC Consistency for Code Generation

Framework for selecting LLM-generated code using test voting consistency without determining individual test correctness.

Ax Allassan Tchangmena A Nken, Baimam Boukar Jean Jacques, Miriam Rateike, Celia Cintas, Skyler Speakman 7/14/2026

Retrieval with Multiple Query Vectors through Anomalous Pattern Detection

Retrieval method using multiple query vectors simultaneously via anomalous pattern detection for complex reasoning tasks.

Ax Yazheng Liu, Yuxuan Wan, Rui Xu, Xi Zhang, Sihong Xie, Hui Xiong 7/14/2026

Attribution-Guided Continual Learning for Large Language Models

Attribution-guided approach for continual learning in LLMs that identifies and preserves parameters storing important knowledge.

Ax Ben Rachmut, Luise Ge, William Yeoh, Ning Zhang, Yevgeniy Vorobeychik 7/14/2026

COSMOS: Model-Agnostic Personalized Federated Learning with Clustered Server Models and Pseudo-Label-Only Communication

Federated learning framework handling heterogeneous client architectures and data distributions via server-side personalization.

Ax Dongjun Lee, Ga-eun Bae, Insu Yun 7/14/2026

CTFusion: A CTF-based Benchmark for LLM Agent Evaluation

Benchmark using Capture The Flag challenges to evaluate LLM agents on cybersecurity tasks with contamination safeguards.

Ax Bat-Sheva Einbinder, Hen Davidov, Yee Whye Teh, Yarin Gal, Yaniv Romano 7/14/2026

Selective Safety Steering via Value-Filtered Decoding

Decoding-time steering method for LLMs that selectively applies safety constraints without unnecessary interventions.

Ax Diego Coello de Portugal Mecke, Tom Hanika, Lars Schmidth-Thieme 7/14/2026

Prune, Update and Trim: Robust Structured Pruning for Large Language Models

Post-training pruning method for reducing LLM inference costs by removing model parameters while maintaining performance.

Ax Umut Onur Yasar 7/14/2026

Student Capacity Moderates Knowledge Distillation Effectiveness: A Systematic Study Across ResNet Teacher-Student Pairs on CIFAR-10

Investigates how teacher-student capacity ratios affect knowledge distillation effectiveness in ResNet image classification on CIFAR-10.

Ax Peiman Mohseni, Nick Duffield, Raymond K. W. Wong 7/14/2026

Revisiting Neural Processes via Fourier Transform and Volterra Series

Neural processes model using Fourier transforms and Volterra series for learning latent functions from irregular measurements with translation equivariance.

Ax Jacques Raynal, Pierre Slangen, Elsa Raynal, Jacques Margerit 7/14/2026

From Performance to Representational Adequacy: A Representational Bootstrap Framework for Adaptive Biological Systems

Representational bootstrap framework for analyzing adaptive biological systems beyond aggregate performance metrics.

Ax Farhin Farhad Riya, Olivera Kotevska, Jinyuan Stella Sun 7/14/2026

IntraShuffler: A Privacy Preserving Framework for Heterogeneous DP Federated Learning

Privacy framework for federated learning with heterogeneous differential privacy budgets, addressing gradient pattern leakage in client updates.

Ax Haoyang Hong, Zichen Wang, Quanquan Gu, Huazheng Wang 7/14/2026

Online KL-Regularized Reinforcement Learning with Function Approximation under Misspecification

Analysis of KL-regularized reinforcement learning under function approximation with model misspecification, extending bounds to non-realizable settings.

Ax Jacques Raynal, Pierre Slangen, Elsa Raynal, Jacques Margerit 7/14/2026

Bootstrap Theory of Representational Emergence: Explanatory Insufficiency as a Driver of Representation Learning and World Models

Theoretical framework (TBER) examining how new representational levels emerge in learning systems when existing representations become insufficient for explanation.

Ax Hyunseok Paeng 7/14/2026