Ax Hyunji Jung, Sungbin Shin, Namhoon Lee 19h ago

Mitigating Staleness in Asynchronous Pipeline Parallelism via Basis Rotation

Addresses gradient staleness in asynchronous pipeline parallelism for distributed training via basis rotation technique.

Ax Yuntong Hu, Yuting Zhang, Matthew Trager, Yi Zhang, Shuo Yang, Wei Xia, Stefano Soatto 19h ago

EvoMAS: Evolutionary Generation of Multi-Agent Systems

EvoMAS uses evolutionary methods to automatically generate LLM-based multi-agent system architectures, addressing brittleness and generalization challenges in MAS design.

Ax Zitao Song, Cedar Site Bai, Zhe Zhang, Brian Bullins, David F. Gleich 19h ago

Decoupling Variance and Scale-Invariant Updates in Adaptive Gradient Descent for Unified Vector and Matrix Optimization

Novel adaptive optimization algorithm extending Adam-style methods to matrix operations for large-scale training.

Ax Jia Li, Yu Hou, Rui Zhang 19h ago

Capture Timing-Attention of Events in Clinical Time Series

ML method for discovering event patterns in clinical time series data using attention mechanisms with timing awareness.

Ax Max Cairney-Leeming, Amartya Sanyal, Christoph H. Lampert 19h ago

Adaptive Sampling and Clipping for Private Worst-Case Group Optimization

Research on fairness and privacy in ML for human-centric tasks using worst-case group optimization and differential privacy techniques.

Ax Jinwoo Kim, Taylor Berg-Kirkpatrick, Loris D'Antoni 19h ago

Continuous Diffusion Models Can Obey Formal Syntax

Training-free guidance for continuous diffusion language models to satisfy formal syntax constraints. Enables JSON/structured output generation without retraining.

Ax Chundong Liang, Yongqi Huang, Dongxiao He, Peiyuan Li, Yawen Li, Di Jin, Weixiong Zhang 19h ago

Unified Multi-Domain Graph Pre-training for Homogeneous and Heterogeneous Graphs via Domain-Specific Expert Encoding

Multi-domain graph pre-training with domain-specific experts for homogeneous and heterogeneous graphs. Unified approach handling mixed graph types across distribution shifts.

Ax Gabriel Franco, Carson Loughridge, Mark Crovella 19h ago

Singular Vectors of Attention Heads Align with Features

Analysis of attention head singular vectors aligning with learned features in language models. Provides theoretical justification for mechanistic interpretability observations.

Ax Adam Zweiger, Xinghong Fu, Han Guo, Yoon Kim 19h ago

Fast KV Compaction via Attention Matching

Fast KV cache compaction using attention matching. Reduces key-value cache size for long-context LLM inference while maintaining performance.

Ax Hanna Herasimchyk, Robin Labryga, Tomislav Prusina, S\"oren Laue 19h ago

A Structural Theory of Position Bias in Transformers

Structural theory explaining position bias and Lost-in-the-Middle phenomenon in Transformers. Analyzes causal attention architecture origins of token position bias.

Ax Gabriel Raya, Bac Nguyen, Georgios Batzolis, Yuhta Takida, Dejan Stancevic, Naoki Murata, Chieh-Hsin Lai, Yuki Mitsufuji, Luca Ambrogioni 19h ago

Noise Scheduling as Information-Guided Allocation in Diffusion Training

InfoNoise adaptive noise scheduling for diffusion model training. Data-adaptive allocation based on conditional entropy to optimize denoising difficulty.

Ax Stephen Zhewen Lu, Aakarsh Vermani, Kohei Sanno, Jiarui Lu, Frederick A Matsen, Milind Jagota, Yun S. Song 19h ago

Conditionally Site-Independent Neural Evolution of Antibody Sequences

Neural evolution approach for antibody engineering using phylogenetic models. Leverages affinity maturation data to model evolutionary fitness landscape.

Ax Abduragim Shtanchaev, Albina Ilina, Yazid Janati, Arip Asadulaev, Martin Takac, Eric Moulines 19h ago

Sparse Scheduled Diffusion Guidance for Inverse Problems

Sparse scheduled diffusion guidance for Bayesian inverse problems. Reduces computational cost by applying guidance selectively through reverse trajectory.

Ax Zhengzhao Ma, Xueru Wen, Boxi Cao, Yaojie Lu, Hongyu Lin, Jinglin Yang, Min He, Xianpei Han, Le Sun 19h ago

Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards

RLVR method addressing calibration degradation in LLM reasoning. Decouples confidence from reasoning to prevent overconfidence on incorrect answers.

Ax Prithviraj Tarale, Kiet Chu, Abhishek Varghese, Kai-Chun Liu, Maxwell A. Xu, Mohit Iyyer, Sunghoon I. Lee 19h ago

Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data

Self-supervised learning for wearable accelerometer data using biological structure tokenization. Improves human activity recognition with limited labeled data.

Ax Khiem Le, Sreejata Dey, Marcos Mart\'inez Galindo, Vanessa Lopez, Ting Hua, Nitesh V. Chawla, Hoang Thanh Lam 19h ago

Can Decision Trees Teach Large Language Models? Distilling Verbalized Knowledge for Molecular Property Prediction

TreeKD method distilling tree-based model knowledge into LLMs for molecular property prediction. Improves LLM performance on drug discovery tasks.

Ax Qiyuan Chen, Xian Wu, Yi Wang, Xianhao Chen 19h ago

HO-SFL: Hybrid-Order Split Federated Learning with Backprop-Free Clients and Dimension-Free Aggregation

Hybrid-Order Split Federated Learning reducing memory usage on edge devices. Combines zeroth-order optimization with split learning for faster convergence.

Ax Yu-Chen Den, Kuan-Yu Chen, Kendro Vincent, Darby Tien-Hao Chang 19h ago

Integrating Inductive Biases in Transformers via Distillation for Financial Time Series Forecasting

Transformer variants for financial time-series forecasting using knowledge distillation. Addresses non-stationary data and regime shifts in financial markets.

Ax Dmitry Zaytsev, Valentina Kuskova, Michael Coppedge 19h ago

From Causal Discovery to Dynamic Causal Inference in Neural Time Series

Framework for discovering and inferring dynamic causal relationships in time-series neural networks without requiring known causal structure a priori.

Ax Huanran Chen, Huaqing Zhang, Xiao Li, Yinpeng Dong, Ke Shen, Jun Zhu 19h ago

Nexus: Same Pretraining Loss, Better Downstream Generalization via Common Minima

Research on LLM pretraining convergence: investigating whether models converge to common minima across data sources to improve downstream generalization.

Ax Haolong Hu, Hanyu Li, Tiancheng He, Huahui Yi, An Zhang, Qiankun Li, Kun Wang, Yang Liu, Zhigang Zeng 19h ago

SaFeR-Steer: Evolving Multi-Turn MLLMs via Synthetic Bootstrapping and Feedback Dynamics

SaFeR-Steer framework for multi-turn safety alignment in multimodal LLMs using synthetic data bootstrapping and feedback dynamics to address long-context safety degradation.

HN theanonymousone 20h ago

Popular LLM software hit by critical vulnerability in Python package Starlette

Security vulnerability discovered in Starlette Python package used by LLM software.

HN uriva 20h ago

Why I Wrote Safescript

Safescript language design motivation: preventing supply chain attacks and hidden logic in AI-generated code via static analysis.

HN thm 20h ago

Your Future job will be to keep AI on task

Opinion piece on future job roles managing and coordinating AI systems.

HN AIOSAI 21h ago

AIPass – Persistent agent workspace with identity, memory, and email

AIPass platform for persistent multi-agent workspaces with shared memory, context, and collaboration between agents.

HN Jainish08 21h ago

Teleport-env – <500ms stateful rollbacks for AI agents via CRIU

Teleport-env: <500ms OS-level stateful rollback sandbox for autonomous coding agents using CRIU snapshots for MCTS and RL.

HN mattas 21h ago

Meta to start testing AI subscription services

Meta testing paid subscriptions for AI features on Meta AI app and website.

HN doppp 21h ago

Zig 2026: No-AI Policy, $670K Foundation, Left GitHub and Why Zig Isn't 1.0 [video]

Zig programming language 2026 roadmap includes no-AI policy, foundation funding, and GitHub departure.

HN jwsteigerwalt 21h ago

Do they know we can tell it's AI slop?

Workplace discussion about identifying AI-generated content in internal communications.

HN y1zhou 21h ago

Language Modeling Materializes a World Model of Protein Biology [pdf]

Research paper on protein biology language models developing implicit world models for biological prediction and understanding.

HN droidjj 22h ago

Illinois Lawmakers Just Passed America's Strongest AI Safety Bill

Illinois legislature passes SB 315 requiring third-party safety audits of frontier AI labs.

HN mooreds 23h ago

The Authorization Paradox: Who Has the Keys to Your AI? [video]

Video on access control and authorization mechanisms for AI systems.

HN dtingle 23h ago

Show HN: Liiists, a Markdown-first, iOS and CLI list app

Markdown-first list management app for iOS and CLI. Open source project.

HN Pallav123 23h ago

Building the Future of Accessible Tech: Inside Uvilox AI

Uvilox AI offers real-time sign language interpretation with <80ms latency using vision AI models.

HN zfancy 1d ago

AgingBench: AI Agents Age Too

AgingBench research on long-context memory degradation in AI agents. Studies information loss and retrieval problems over time.

HN ballista2026 1d ago

New method for testing AI using real work-flows

LLM INQUISITOR: methodology for evaluating AI systems in real workflows, not benchmarks. Tests stability and reliability.

HN pedronauck21 1d ago

AGH is the First open-network protocol for AI Agents

AGH: open network protocol for AI agents. Enables durable CLI sessions with memory, tools, autonomy on NATS-based channels.

HN Depiphany_Reb 1d ago

What Emerges When You Don't Break Sustained Coherent Human-AI Interaction

Paper formalizing emergent behavioral patterns in sustained human-AI interaction. Introduces 'third vector' concept in response space.

HN PaoloJNN 1d ago

I built a free cross-AI prompt manager (Chrome ext)

Chrome extension providing unified prompt management interface across multiple AI platforms.

HN berlianta 1d ago

Multi-turn jailbreak rates across 15 frontier models (Grok 88%, Claude 12%)

Empirical study measuring jailbreak vulnerability rates across 15 frontier LLMs including Grok (88%) and Claude (12%).

HN intrepidsoldier 1d ago

AI Infra Is Nothing Like the "Classic Cloud Infra"

Analysis of differences between AI infrastructure requirements and traditional cloud infrastructure design.

HN lschueller 1d ago

Malware dev tries to steal Claude users' secrets, leaks own GitHub private token

Security incident: malware developer attempted to steal Claude API credentials but leaked own GitHub token.

HN Bender 1d ago

Argonne flexes spare supercompute to build private AI inference service

Argonne National Lab uses supercomputing resources to build private AI inference service.

HN Bender 1d ago

AIs don't like religion – particularly Jehovah's Witnesses, study claims

Study examining AI model biases and responses toward religious content, particularly Jehovah's Witnesses.

HN baroiall 1d ago

Agent Accountability Gap: Why Network Policies, API Gateways, & RBAC Isnt Enough

Framework analyzing five pillars of AI agent accountability: traceability, authorization, identity, policy, and oversight.

HN benn_88 1d ago

Getting Claude to extract data from a 1997 football manager game

Using Claude API to extract data from 1997 football manager game.

HN hyzyla 1d ago

Show HN: Sweetpad – VS Code extension for building iOS/Swift apps

VS Code extension for Swift/iOS development integrating xcodebuild, swift-format, and language server protocol.

HN taubek 1d ago

Facebook launches a 'Plus' subscription that gives you extra features

Meta launches paid subscriptions for Facebook, WhatsApp, Instagram, and Meta AI.

HN sjhalani7 1d ago

Show HN: VAEN – Package and import portable AI coding-agent Harnesses

VAEN open-source CLI packages AI coding-agent harnesses with skills and MCP servers as portable .agent files.

HN abr0ahm 1d ago

Show HN: Sneakily steer candidates toward naive brute-force solutions

Assessment platform intercepting Claude Code requests to prevent AI from over-solving coding interview problems.