3/17/2026
The first open-source agentic AI physicist
Open-source AI agent designed to perform physics research tasks autonomously.
3/17/2026
Show HN: Oh-my-agent – A structural harness for AI agents in real projects
Framework for reliable AI agent development addressing hallucination and task drift. Structured protocol for production agent deployments.
3/17/2026
Show HN: MCP Isn't Dead. You're Just Using It Wrong
Analysis of MCP dynamic tool registration feature. Argues MCP enables advanced agent capabilities beyond static tool definitions.
3/17/2026
Ask HN: What AI can you use for personal video editing?
User question about AI tools for personal video editing. Discussion of limitations in current LLM video capabilities.
3/17/2026
LLM drift-Claude vs. Calmkeep: 25-turn Code (60% vs. 85%) & Legal (50% vs. 100%)
Performance comparison of Claude vs Calmkeep on 25-turn code and legal tasks. Shows 60%-85% code accuracy and 50%-100% legal accuracy.
3/17/2026
Three zones of LLM competence for software engineers
Analysis of LLM competence zones for software engineering tasks. Framework for understanding model capabilities and limitations.
3/17/2026
LLMs benchmark with esoteric programming languages
Benchmark study showing LLM code generation relies on memorization. Models score 90% on Python but 3.8% on esoteric languages.
3/17/2026
Show HN: FreeFlow – Open-Source Wispr Flow
Open-source voice-to-text tool with real-time speech cleaning and injection into any app. Customizable alternative to Whisper Flow.
3/17/2026
ClickSay – Chrome extension that captures UI context for AI coding tools
ClickSay is a Chrome extension that captures UI context (selectors, styles, HTML, screenshots) and voice input for AI coding tools like Claude Code.
3/17/2026
Show HN: The new security frontier for LLMs; SIEM evasion
Security research showing AI agents can perform SIEM/EDR evasion, indicating organizations must assume adversaries will gain these LLM-powered capabilities.
3/17/2026
How I Used Lima for an AI Coding Agent Sandbox
Experience report using Lima for sandboxing AI coding agents (Claude Code, Codex) to enable autonomous operation with controlled permissions.
3/17/2026
Introducing GPT-5.4 mini and nano
OpenAI releases GPT-5.4 mini and nano models optimized for coding and subagents with 2x faster inference and improved reasoning.
3/17/2026
Rtk – CLI proxy that reduces LLM token consumption by 60-90%
Rtk is a Rust CLI proxy reducing LLM token consumption 60-90% by filtering and compressing command outputs before context, with <10ms overhead.
3/17/2026
Ask HN: Quick technical questions about LLMs
Discussion thread with technical questions about LLM mechanics: token stopping, prompt continuation, and next-token prediction behavior.
3/17/2026
Can LLMs Hack Enterprise Networks?
Prototype using LLMs for autonomous assumed-breach penetration testing against Active Directory networks, demonstrating LLM capabilities in enterprise security contexts.
3/17/2026
A structural epistemic limit in LLMs: 8–15% unverifiable claims across domains
MarCognity-AI is an open-source framework analyzing LLM claim verification, finding 8-15% unverifiable claims. Decomposes responses and verifies against sources.
3/17/2026
Out-of-Context Reasoning in LLMs: A short primer and reading list
Primer on out-of-context reasoning in LLMs: when models reach conclusions requiring reasoning not present in context window, affecting generalization and alignment.
3/17/2026
Show HN: ModelSweep - Open-Source Benchmarking for Local LLMs
ModelSweep is a GUI-based benchmarking workbench for evaluating local LLMs on Ollama, enabling test suite building and comparative dashboards.
3/17/2026
Llmgate – call any LLM via YAML config, 2 dependencies
Llmgate is a lightweight Python wrapper supporting 21 LLM providers via YAML config with only 2 dependencies (httpx, pyyaml).
3/17/2026
Show HN: DataFlow,Turn raw data into high-quality LLM training datasets
DataFlow is a low-code visual pipeline tool for generating, cleaning, and preparing high-quality LLM training datasets with flexible orchestration.
3/17/2026
M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling
M²RNN: non-linear RNN architecture with matrix-valued states for language modeling with greater expressive power than Transformers.
3/17/2026
AerialVLA: A Vision-Language-Action Model for UAV Navigation via Minimalist End-to-End Control
AerialVLA: end-to-end vision-language-action model for UAV navigation combining visual interpretation with fuzzy linguistic instructions.
3/17/2026
OxyGen: Unified KV Cache Management for Vision-Language-Action Models under Multi-Task Parallelism
OxyGen system for unified KV cache management in vision-language-action models enabling efficient multi-task parallel inference.
3/17/2026
SPARQ: Spiking Early-Exit Neural Networks for Energy-Efficient Edge AI
SPARQ framework integrating spiking neural networks, quantization, and early-exit mechanisms for energy-efficient edge AI.
3/17/2026
From $\boldsymbol{\log\pi}$ to $\boldsymbol{\pi}$: Taming Divergence in Soft Clipping via Bilateral Decoupled Decay of Probability Gradient Weight
Bilateral decoupled decay method for stabilizing soft clipping in reinforcement learning with verifiable rewards for LLM reasoning.
3/17/2026
Extending Minimal Pairs with Ordinal Surprisal Curves and Entropy Across Applied Domains
Extension of minimal pairs evaluation using ordinal surprisal curves to assess linguistic knowledge in LLMs beyond binary judgments.
3/17/2026
ES-Merging: Biological MLLM Merging via Embedding Space Signals
Method for merging specialized biological multimodal LLMs using embedding space signals to combine modalities.
3/17/2026
Questionnaire Responses Do not Capture the Safety of AI Agents
Study showing questionnaire-based safety assessments of AI agents fail to capture real-world deployment safety concerns.
3/17/2026
Distilling Reasoning Without Knowledge: A Framework for Reliable LLMs
Modular framework separating planning from retrieval in LLMs to improve reliability on factual QA with explicit tool usage.
3/17/2026
Infinite Problem Generator: Verifiably Scaling Physics Reasoning Data with Agentic Workflows
Infinite Problem Generator: agentic framework synthesizing physics problems with guaranteed solvability for LLM training data generation.
3/17/2026
CangjieBench: Benchmarking LLMs on a Low-Resource General-Purpose Programming Language
CangjieBench benchmark for evaluating LLMs on Cangjie, a low-resource general-purpose programming language with contamination-free evaluation.
3/17/2026
Trust-Region Noise Search for Black-Box Alignment of Diffusion and Flow Models
Trust-region search algorithm for black-box alignment of diffusion and flow models to target rewards at inference time without gradient access.
3/17/2026
VLA-Thinker: Boosting Vision-Language-Action Models through Thinking-with-Image Reasoning
Vision-Language-Action framework with thinking-with-image reasoning allowing models to revisit visual context during long-horizon embodied tasks.
3/17/2026
MALicious INTent Dataset and Inoculating LLMs for Enhanced Disinformation Detection
Benchmark of 12 language models on MALINT, a human-annotated disinformation corpus capturing malicious intent, for improved detection.
3/17/2026
An End-to-end Architecture for Collider Physics and Beyond
End-to-end language-driven agent system for high-energy physics phenomenology workflows, executing tasks from theoretical input to final outputs.
3/17/2026
Machine Learning-Driven Intelligent Memory System Design: From On-Chip Caches to Storage
Survey on using machine learning methods for adaptive memory system design in modern computing platforms instead of static heuristics.
3/17/2026
FlashHead: Efficient Drop-In Replacement for the Classification Head in Language Model Inference
Efficient drop-in replacement for dense classification heads in language models, reducing parameter and compute overhead for consumer devices.
3/17/2026
D-MEM: Dopamine-Gated Agentic Memory via Reward Prediction Error Routing
Biologically-inspired agentic memory architecture using reward prediction error routing to reduce token costs and write latency in LLM agents.
3/17/2026
A Loss Landscape Visualization Framework for Interpreting Reinforcement Learning: An ADHDP Case Study
Loss landscape visualization framework for interpreting reinforcement learning behavior in actor-critic algorithms and control systems.
3/17/2026
$PA^3$: $\textbf{P}$olicy-$\textbf{A}$ware $\textbf{A}$gent $\textbf{A}$lignment through Chain-of-Thought
Policy-aware agent alignment framework using chain-of-thought reasoning to help LLM agents adhere to complex business rules without excessive prompting.
3/17/2026
Delightful Policy Gradient
Novel policy gradient method addressing pathological behavior in standard policy gradients through context-aware advantage weighting.
3/17/2026
LLM-Augmented Release Intelligence: Automated Change Summarization and Impact Analysis in Cloud-Native CI/CD Pipelines
LLM-augmented system for automated change summarization and impact analysis in cloud-native CI/CD pipelines and release management.
3/17/2026
s2n-bignum-bench: A practical benchmark for evaluating low-level code reasoning of LLMs
Benchmark for evaluating LLMs on low-level code reasoning and formal proof generation using cryptographic library assembly code.
3/17/2026
ResearchPilot: A Local-First Multi-Agent System for Literature Synthesis and Related Work Drafting
Open-source multi-agent system for literature review assistance using DSPy, Qdrant, and local-first architecture to synthesize papers and draft related work.
3/17/2026
Compute Allocation for Reasoning-Intensive Retrieval Agents
Study on compute allocation strategies for LLM-augmented retrieval agents handling reasoning-intensive queries over long horizons with growing memory stores.
3/17/2026
Nudging Hidden States: Training-Free Model Steering for Chain-of-Thought Reasoning in Large Audio-Language Models
Training-free inference-time model steering strategies to improve chain-of-thought reasoning in large audio-language models across multiple benchmarks.
3/17/2026
EARCP: Self-Regulating Coherence-Aware Ensemble Architecture for Sequential Decision Making -- Ensemble Auto-Regule par Coherence et Performance
EARCP ensemble architecture dynamically weights heterogeneous expert models based on performance and inter-model coherence for sequential decision making.
3/17/2026
VisionCoach: Reinforcing Grounded Video Reasoning via Visual-Perception Prompting
VisionCoach uses reinforcement learning with visual-perception prompting to improve spatio-temporal grounding in video reasoning models.
3/17/2026
Seamless Deception: Larger Language Models Are Better Knowledge Concealers
Study on detecting when language models actively conceal knowledge, finding larger models better at deception with gradient-based concealment easier to detect.
3/17/2026
AgentTrace: Causal Graph Tracing for Root Cause Analysis in Deployed Multi-Agent Systems
AgentTrace provides lightweight causal graph tracing for post-hoc root cause diagnosis in deployed multi-agent workflows with cascading failures.