3/9/2026
How Reliable is Language Model Micro-Benchmarking?
Meta-evaluation study analyzing reliability of micro-benchmarks for ranking LLMs, comparing against full benchmarks and random sampling approaches.
3/9/2026
CanvasMAR: Improving Masked Autoregressive Video Prediction With Canvas
CanvasMAR: masked autoregressive video generation model improved with canvas mechanism to reduce frame distortion and enhance sampling quality with few steps.
3/9/2026
Escaping Model Collapse via Synthetic Data Verification: Near-term Improvements and Long-term Convergence
Study on model collapse in generative models trained iteratively on synthetic data, proposing verification methods to prevent performance degradation and reverse collapse trends.
3/9/2026
Mixed Monotonicity Reachability Analysis of Neural ODE: A Trade-Off Between Tightness and Efficiency
Interval-based reachability analysis method for Neural ODEs using mixed monotonicity techniques to verify continuous-time machine learning models of dynamical systems.
3/9/2026
Real-Time Learning of Predictive Dynamic Obstacle Models for Robotic Motion Planning
Real-time learning framework for predicting dynamic obstacle motions in robotic systems using Hankel-DMD for online nonlinear model identification from noisy measurements.
3/9/2026
FireScope: Wildfire Risk Prediction with a Chain-of-Thought Oracle
FireScope-Bench: dataset and benchmark for wildfire risk prediction using satellite imagery, climate data, and chain-of-thought reasoning with multimodal integration.
3/9/2026
SPINE: Token-Selective Test-Time Reinforcement Learning with Entropy-Band Regularization
Test-time reinforcement learning method for LLMs using token-selective entropy-band regularization to prevent response collapse.
3/9/2026
Uncertainty-Aware Subset Selection for Robust Visual Explainability under Distribution Shifts
Analyzes robustness of subset selection-based visual explanation methods under distribution shifts and out-of-distribution conditions.
3/9/2026
A Novel Patch-Based TDA Approach for Computed Tomography Imaging
Applies topological data analysis to CT imaging for improved feature extraction in medical imaging ML models.
3/9/2026
Restoring Exploration after Post-Training: Latent Exploration Decoding for Large Reasoning Models
Addresses exploration collapse in large reasoning models post-training by proposing latent exploration decoding to restore sampling diversity.
3/9/2026
Validating Interpretability in siRNA Efficacy Prediction: A Perturbation-Based, Dataset-Aware Protocol
Validates interpretability of saliency maps in siRNA efficacy prediction using perturbation-based counterfactual sensitivity testing.
3/9/2026
SWE-MiniSandbox: Container-Free Reinforcement Learning for Building Software Engineering Agents
Container-free reinforcement learning framework for training software engineering agents at scale without per-task container overhead.
3/9/2026
GaiaFlow: Semantic-Guided Diffusion Tuning for Carbon-Frugal Search
Proposes carbon-efficient neural ranking architecture for information retrieval using semantic-guided diffusion tuning.
3/9/2026
The Limits of Long-Context Reasoning in Automated Bug Fixing
Evaluates whether LLMs can reliably perform long-context code debugging and patch generation, testing limits of agentic workflows on software engineering tasks.
3/9/2026
What Topological and Geometric Structure Do Biological Foundation Models Learn? Evidence from 141 Hypotheses
Investigates geometric and topological structures learned by biological foundation models like scGPT using autonomous hypothesis screening with AI-driven workflows.
3/9/2026
Modality Collapse as Mismatched Decoding: Information-Theoretic Limits of Multimodal LLMs
Information-theoretic analysis of multimodal LLM failure modes. Frames modality collapse as mismatched decoding problem, explains 98% information loss.
3/9/2026
Fine-Tuning and Evaluating Conversational AI for Agricultural Advisory
Fine-tuning conversational LLMs for agricultural advisory with domain-specific improvements. Addresses recommendation accuracy and farmer communication alignment.
3/9/2026
Fragile Thoughts: How Large Language Models Handle Chain-of-Thought Perturbations
Empirical study evaluating LLM robustness to chain-of-thought reasoning perturbations across five error types. Assesses reasoning reliability under corruption.
3/9/2026
Improving the accuracy of physics-informed neural networks via last-layer retraining
Research on improving physics-informed neural networks accuracy through post-processing retraining. Domain-specific ML application for solving PDEs.
3/9/2026
Good-Enough LLM Obfuscation (GELO)
Arxiv paper proposing obfuscation method to protect LLM prompt privacy on shared accelerators. Addresses KV cache security against adversarial memory access.
3/9/2026
OpenAI's Symphony: Agent Management Layer
OpenAI's Symphony orchestrates autonomous coding agents for project work, monitoring task boards and managing PR delivery with proof-of-work artifacts.
3/9/2026
Zero Lines Written by a Human but 750 Pull Requests Later
Engineer completed production app with 750+ PRs across 4 languages in 45 days using only AI code generation, no human-written code.
3/9/2026
Show HN: Tilnote – An AI note workspace from rough ideas to publishable content
Tilnote AI note workspace uses agent to structure ideas into publishable content from keywords, with web clipper and writing assistance.
3/9/2026
CLI tool for deterministic linting of LLM output
Vale is open-source CLI linting tool for editorial style guides, runs offline, integrates with VS Code and GitHub. Tangentially useful for LLM output processing.
3/9/2026
How to manipulate running LLM outputs via GGUF page cache poisoning
Proof-of-concept exploit demonstrating persistent manipulation of LLM outputs via GGUF page cache poisoning in running inference servers.
3/9/2026
Ask HN: Are we going to see more job postings asking for only agentic coding?
Discussion of job market shift toward agentic coding workflows. Zapier job posting requires experience directing AI agents, handling failures, and multi-agent patterns.
3/9/2026
Show HN: Cantrip – Agent-native GTM engine I built for solo technical founders
AI-powered GTM engine for solo founders. Describes product and generates customer acquisition strategy to reach first 100 users.
3/9/2026
Show HN: Ajen – Open-source platform where AI employees build your startup
Open-source platform where AI agents (CEO, CTO, CMO) collaborate to plan and build startups based on descriptions. Early-stage project seeking feedback.
3/9/2026
Show HN: Running multiple Claude Code agents in parallel Git worktrees (ChatML)
Tool for running multiple Claude Code agents in parallel using Git worktrees to avoid filesystem conflicts, enabling concurrent AI-assisted development workflows.
3/9/2026
Show HN: Nervous System v1.9 – Governance for Multi-Agent AI (MCP)
Nervous System governance framework enforces 7 rules preventing multi-agent AI failures, battle-tested on 13-agent system with zero bypasses of 58+ violations.
3/9/2026
Show HN: Own your AI's context and memories across every model and device
Personal memory system using knowledge graph, pgvector, and MCP server to share context across multiple LLM providers and devices.
3/9/2026
Your binary is no longer safe: LLM-assisted Decompilation
LLM-assisted decompilation technique for reverse-engineering binary programs, automating binary-to-source code conversion.
3/9/2026
I ran the same AI security test 4 times – 75% found critical bypasses
Mutation testing engine reveals GPT-4 prompt injection vulnerabilities, finding different critical bypasses in 75% of runs despite identical inputs.
3/9/2026
What if you never had to get an API key ever again?
Val Town platform founder discusses eliminating API key friction in developer workflows, relevant for agent/LLM app development experience.
3/9/2026
Show HN: Spotr – Client-side fuzzy search for collections in TypeScript
TypeScript fuzzy search library for client-side collection searching with configurable scoring. Developer tool but not AI-specific.
3/9/2026
Show HN: Mguard – First defense against MINJA memory poisoning attacks
Security library defending against memory poisoning attacks (MINJA, AgentPoison, MemoryGraft) on AI agents. Drop-in protection for Mem0, LangChain, custom systems.
3/9/2026
Show HN: Volt HQ – MCP server comparing AI inference pricing across providers
MCP server for comparing AI inference pricing across providers with budget alerts and optimization recommendations.
3/9/2026
Show HN: Security toolkit for OpenClaw – scanner, hardened configs, guides
Security toolkit for OpenClaw personal AI assistant including scanner, hardened configs, and vulnerability guides. Addresses exposed instances.
3/9/2026
Agency: Specialized Expert Agents with Personality
Collection of specialized AI agent personalities with distinct expertise, processes, and deliverables for various tasks.
3/8/2026
AI allows hackers to identify anonymous social media accounts, study finds
Research on using LLMs to de-anonymize social media accounts and link identities across platforms.
3/8/2026
Show HN: Self-hosted financial analyst – Plaid and Claude and Next.js, –$5/month
Self-hosted personal finance app integrating Plaid, Claude API, and Next.js for AI-powered investment analysis.
3/8/2026
State-of-the-Art Prompting for AI Agents (2025)
Summary of prompt engineering techniques from YC founders for building AI agents.
3/8/2026
AI Assistants Are Moving the Security Goalposts
Analysis of security implications and risks introduced by autonomous AI agents with computer access.
3/8/2026
Show HN: Wa-agent – Framework for building AI agents on WhatsApp
Node.js framework for autonomous AI agents on WhatsApp using YAML config, multi-step tool use, and multiple model providers.
3/8/2026
Claude Custom Chat – customize your Claude Code extension
VS Code/Cursor extension providing custom chat interface for Claude Code CLI. Self-modifying extension with rollback capability.
3/8/2026
Oly – Run AI agents, close your terminal, intervene when it needed from anywhere
Session-persistent PTY daemon for long-running CLI AI agents with intervention capabilities from anywhere.
3/8/2026
Show HN: Engram — a brain-inspired context database for AI agents
Engram: persistent context database for AI agents and LLMs that manages memory like human cognition to prevent context collapse and agent coordination issues.
3/8/2026
Threat-Modeling the OWASP Top for LLM Applications
Security threat modeling and case studies of LLM application vulnerabilities including data exfiltration and prompt injection.
3/8/2026
Luma AI releases new Uni-1 image model beating Nano-banana 2
Luma AI releases Uni-1, a unified multimodal understanding and generation model for images and video. Technical progress in visual AI.
3/8/2026
A simple L7 proxy for vLLM that manages LoRA adapter storage via NVMes
L7 proxy for vLLM managing LoRA adapter storage, routing, and workload affinity across distributed clusters.