Ax Francesco Pio Monaco, Elia Cunegatti, Flavio Vella, Giovanni Iacca 4d ago

Frequency Matters: Fast Model-Agnostic Data Curation for Pruning and Quantization

Frequency-based data curation method for selecting calibration data to preserve LLM performance during post-training pruning and quantization.

Ax Zice Wang, Zhenyu Zhang 4d ago

Framing Effects in Independent-Agent Large Language Models: A Cross-Family Behavioral Analysis

Examines prompt framing effects on LLM decision-making in threshold voting tasks across model families under isolated, non-interactive settings.

Ax Dang Nguyen, Harvey Yiyun Fu, Peter West, Ari Holtzman, Chenhao Tan 4d ago

Moral Mazes in the Era of LLMs

HR Simulator: Game-based evaluation of LLMs navigating complex workplace social norms like giving feedback and rejecting requests appropriately.

Ax Drake Caraker, Bryan Arnold, David Rhoads 4d ago

First-Mover Bias in Gradient Boosting Explanations: Mechanism, Detection, and Resolution

Identifies first-mover bias in SHAP explanations from gradient boosting's sequential fitting causing attribution instability under multicollinearity.

Ax Abhinaba Basu, Pavan Chakraborty 4d ago

The illusion of reasoning: step-level evaluation reveals decorative chain-of-thought in frontier language models

Step-level faithfulness evaluation shows chain-of-thought reasoning in frontier LLMs is often decorative, post-hoc narrative rather than genuine reasoning.

Ax Yuntong Zhang, Zhiyuan Pan, Imam Nur Bani Yusuf, Haifeng Ruan, Ridwan Shariffdeen, Abhik Roychoudhury 4d ago

Code Review Agent Benchmark

Code Review Agent Benchmark: Dataset and evaluation framework for assessing AI agents' ability to review code quality in generated codebases.

Ax Weimin Liu, Qingkun Li, Jiyuan Qiu, Wenjun Wang, Joshua H. Meng 4d ago

DiffAttn: Diffusion-Based Drivers' Visual Attention Prediction with LLM-Enhanced Semantic Reasoning

DiffAttn: Diffusion-based framework for predicting drivers' visual attention using LLM-enhanced semantic reasoning for intelligent vehicles.

Ax Kavindu Herath, Joshua Zhao, Saurabh Bagchi 4d ago

Beyond Corner Patches: Semantics-Aware Backdoor Attack in Federated Learning

SABLE: Semantics-aware backdoor attack on federated learning using realistic, in-distribution visual triggers instead of synthetic patterns.

Ax Ziliang Guo, Ziheng Li, Bo Tang, Feiyu Xiong, Zhiyu Li 4d ago

MemFactory: Unified Inference & Training Framework for Agent Memory

MemFactory: Unified framework for training and inference of memory-augmented LLM agents with reinforcement learning optimization of memory operations.

Ax Ryszard Tuora, Mateusz Gali\'nski, Micha{\l} Godziszewski, Micha{\l} Karpowicz, Mateusz Czy\.znikiewicz, Adam Kozakiewicz, Tomasz Zi\k{e}tkiewicz 4d ago

UnWeaving the knots of GraphRAG -- turns out VectorRAG is almost enough

Compares GraphRAG with VectorRAG for retrieval-augmented generation, showing simpler vector-based approaches handle chunk relationships effectively.

Ax Sicheng Zuo, Zixun Xie, Wenzhao Zheng, Shaoqing Xu, Fang Li, Hanbing Li, Long Chen, Zhi-Xin Yang, Jiwen Lu 4d ago

DVGT-2: Vision-Geometry-Action Model for Autonomous Driving at Scale

DVGT-2: Vision-Geometry-Action model for autonomous driving using dense 3D geometry instead of language descriptions for planning.

Ax Manoj Parmar 4d ago

Safety, Security, and Cognitive Risks in World Models

Analyzes safety, security, and cognitive risks in world models used for autonomous decision-making in robotics and agentic AI systems.

Ax Vardaan Pahuja, Samuel Stevens, Alyson East, Sydne Record, Yu Su 4d ago

Automatic Image-Level Morphological Trait Annotation for Organismal Images

Uses sparse autoencoders to automatically annotate morphological traits in biological organism images, automating expert-driven extraction process.

Ax Wei Zou, Mingwen Dong, Miguel Romero Calvo, Shuaichen Chang, Jiang Guo, Dongkyu Lee, Xing Niu, Xiaofei Ma, Yanjun Qi, Jiarong Jiang 4d ago

Poison Once, Exploit Forever: Environment-Injected Memory Poisoning Attacks on Web Agents

Demonstrates environment-injected memory poisoning attacks on LLM-based web agents through contamination persisting across sessions without direct memory access.

Ax Adrienne Deganutti, Elad Hirsch, Haonan Zhu, Jaejung Seol, Purvanshi Mehta 4d ago

Graphic-Design-Bench: A Comprehensive Benchmark for Evaluating AI on Graphic Design Tasks

GraphicDesignBench: First comprehensive benchmark for evaluating AI on professional graphic design tasks including layout translation and typographic rendering.

Ax Gregory N. Frank 4d ago

How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models

Identifies sparse routing mechanism in alignment-trained language models where gate attention heads trigger refusal responses, validated across 9 models from 6 labs.

Ax Gabriel Sarch, Linrong Cai, Qunzhong Wang, Haoyang Wu, Danqi Chen, Zhuang Liu 4d ago

Vero: An Open RL Recipe for General Visual Reasoning

Vero: Open-source family of vision-language models matching proprietary systems on visual reasoning tasks using reinforcement learning with public recipes and data.

HN pritopian 4d ago

Comprehensive Benchmark for Evaluating AI on Graphic Design Tasks

arXiv research benchmark for evaluating AI performance on graphic design tasks, measuring model capabilities in visual design domains.

HN jsahasi 4d ago

VitalNexa – AI health agent that reads your actual lab results and wearable data

VitalNexa is an AI health agent that analyzes lab results and wearable data to provide personalized health recommendations and biological age scoring.

HN DaibinThink 4d ago

AI is structurally trained to lie. I built a protocol to break it

Author presents a protocol addressing AI's structural tendency to agree and sound authoritative rather than hallucinating, causing subtle reality distortions in outputs.

HN vibeagentmaking 4d ago

Every Barrier Between AI Agents and Autonomy – A Practical Map

Technical analysis of practical constraints preventing AI agents from autonomous operation, mapping barriers and their severity.

HN mjyut 4d ago

LLM scraper bots are overloading acme.com's HTTPS server

Practical analysis of operational and technical barriers preventing autonomous AI agents, mapping constraints in agent economy.

HN tank-34 4d ago

Pydantic-resolve – define relationships once, reuse across REST, GraphQL and MCP

Pydantic-resolve is a declarative data assembly library using DataLoader pattern to eliminate N+1 queries across REST, GraphQL, and MCP protocols.

HN surprisetalk 4d ago

OpenAI says its new model GPT-2 is too dangerous to release (2019)

2019 article about OpenAI's decision not to release GPT-2 over safety concerns (title only, no content).

HN zhangchen 4d ago

Show HN: Replaced Neo4j with pure vector search for Graph RAG

Case study of server overload caused by LLM scraper bots making excessive HTTPS requests to acme.com domain.

HN erickhill 4d ago

Sam Altman on Building the Future of AI [video]

Video of Sam Altman discussing AI development (title only, no content provided).

HN gmays 4d ago

Larger and more instructable language models become less reliable

Research finding that larger and more instructable language models show decreased reliability (title only).

HN evo_9 4d ago

InstaMed – Oral Dissolving Peptides with InstaRelease Technology

Static analysis tool detecting ReDoS vulnerabilities in Python regular expressions with automatic fixes.

HN b-man 4d ago

Release Please

Marketing content for oral dissolving peptides supplement product.

HN danielmateo773 4d ago

Show HN: Omni Voice – AI Voice Cloning and Text-to-Speech Platform

Omni Voice is a multilingual AI voice cloning and text-to-speech platform supporting 646 languages with unified model.

HN jinqueeny 4d ago

Drive9, a network drive with built-in semantic search

Drive9 is agent-native data infrastructure providing filesystem-like interface with semantic search, embedding, and full-text indexing for AI agents.

HN jenthoven 4d ago

Impact of 100% Adoption of AI Coding Agents by Non-Technical Team

Analysis of impact when non-technical teams adopt AI coding agents at 100% adoption rate (title only).

HN videobroker 4d ago

Law Students: AI Is Changing Things

Overview of how AI is transforming legal work by automating research, document review, and drafting tasks for lawyers and paralegals.

HN nathfavour 4d ago

Show HN: Kylrix- open source de-googled privacy suite for techies

Kylrix is an open-source privacy suite offering note, vault, task, and communication alternatives to Google services.

HN An0n_Jon 4d ago

The way every agent framework handles MCP is a latent security problem

Release Please automates changelog generation, GitHub releases, and version bumping via conventional commit parsing.

HN andrew_matthews 4d ago

Steergen: Single-source steering docs for spec-driven development

Analysis of MCP connection model security: agent frameworks keep all integrations live during sessions, creating unnecessary attack surfaces and costs.

HN lemonish97 4d ago

GitHub Copilot CLI now supports BYOK and local models

GitHub Copilot CLI now supports bring-your-own-key models and local models via Azure OpenAI, Anthropic, or OpenAI-compatible endpoints.

HN nnx 4d ago

JSIR: A High-Level IR for JavaScript

Google's JSIR: open-source high-level intermediate representation for JavaScript code analysis and transforms.

HN apievangelist 4d ago

Naftiko Open-Source Spec-Driven Integration

Open-source spec-driven integration framework for API sprawl, enabling governed AI agent integration with SaaS/microservices.

HN 8dazo 4d ago

Ask HN: Is there any tool that can stop LLM calls at runtime (not just monitor)?

HN discussion on tools for enforcing LLM/agent call limits at runtime rather than just monitoring, addressing cost control in agent systems.

HN dhanushkumars 4d ago

Show HN: Kerf-CLI – SQLite-backed cost analytics for Claude Code

SQLite-backed cost analytics tool for Claude Code sessions, tracking token spend, cache hits, and budget enforcement.

HN jaypatelani 4d ago

The BSDs in the AI Age

Thread initiation post about BSD operating systems and AI with minimal content and formatting guidelines.

HN ZeroDumb 4d ago

Scaling AI Vulnerability Scanning Beyond One File at a Time

Technique for scaling LLM-based vulnerability scanning across multiple files using strategic prompting and structured output for security analysis.

HN healsdata 4d ago

We need re-learn what AI agent development tools are in 2026

Opinion piece on evolution of AI agent development tools in 2026, discussing market consolidation and accessibility barriers for non-programmers.

HN xgerman 4d ago

DocumentDB Fuse – FS interface for the database

FUSE-based filesystem interface for MongoDB-compatible databases, enabling file-based querying without drivers.

HN safishamsi 4d ago

AI coding assistant skill (Claude Code, Codex, OpenCode, OpenClaw, Factory Droid

Claude Code skill that builds knowledge graphs from multimodal inputs to help developers understand codebase structure and architecture.

HN whitehatd 4d ago

Crag – One governance file, every AI coding tool (50 repos, 96.4% accuracy)

Crag governance compiler for AI coding tools that unifies configuration across 12 targets with 96.4% accuracy, solving multi-tool consistency.

HN Caum 4d ago

Show HN: When Seeing Isn't Enough: Rescuing Stuck LLM Agents

Research preprint on blind-spot failures in LLM coding agents, proposing causal interpretation framework for improved agent reliability and rescue mechanisms.

HN smankoo 4d ago

Show HN: Open Browser Control – drive your real Chrome from any MCP client

MCP-compatible Chrome browser control for AI agents. Integrates with Claude, Cursor, Kiro clients. Supports human intervention for CAPTCHAs/MFA.

HN aragonite 4d ago

WSLg: Windows Subsystem for Linux GUI

WSLg enables running Linux GUI applications on Windows via X11/Wayland integration.