til.99jik.com Knowledge Base

LLM-MCTS leverages Large Language Models (LLMs) as both a 'Policy' (a solver that provides immediate answers) and a 'World Model' (a blueprint of how the environment operates). By combining these two roles, the system can navigate complex search spaces by identifying and selecting the most promising paths, similar to using a map to navigate a maze.

AirSim is an open-source flight simulator designed for training and testing drones in a virtual environment to avoid the high costs and risks associated with real-world drone flights. It leverages Unreal Engine for realistic graphics and incorporates a precise physics engine and sensor models to provide a high-fidelity simulation for autonomous vehicle research and Software-in-the-Loop (SIL) automation.

Open Knowledge Format (OKF), announced by Google Cloud in June 2026, is an open format designed for knowledge that AI agents can read, write, and exchange. It represents knowledge using a directory structure consisting of Markdown files combined with YAML frontmatter.

Specification by Example (SBE) is a software development practice designed to bridge the gap between stakeholders, developers, and QA. It ensures that all parties have a shared understanding of the product requirements, reducing ambiguity and unnecessary rework. SBE is particularly effective in short-iteration or flow-based development processes, as it provides just-in-time documentation and treats functional requirements, specifications, and acceptance tests as the same entity.

Open Knowledge Format (OKF), announced by Google Cloud in June 2026, is an open specification designed for knowledge that AI agents can read, write, and exchange. It represents knowledge using a directory structure consisting of Markdown files combined with YAML frontmatter.

The 2008 paper by Satoshi Nakamoto describes the first decentralized financial ledger system. It establishes trust organically by combining the computational resources of global participants with mathematical proofs, eliminating the need for legal enforcement or a central guaranteeing authority. It specifically addresses the problems of double-spending and distributed consensus within P2P networks using Proof of Work.

The 2008 paper by Satoshi Nakamoto, published in the SSRN Electronic Journal, introduces the first decentralized financial ledger system. It establishes a method to create trust spontaneously by combining the computational resources of global participants and mathematical proofs, eliminating the need for legal enforcement or a central certifying authority. The system specifically addresses the double-spending problem and distributed consensus in P2P networks through the use of Proof of Work (PoW).

Matrix-Fuzzer is a testing framework designed to automatically discover safety-critical scenarios for autonomous driving systems. Unlike random generation, it extracts high-level Behavior Tree (BT) parameters from real driving logs and optimizes them within realistic probability distributions. This approach prevents 'unrealistic' collisions caused by erratic agent behavior and focuses on finding actual ego-vehicle defects using a minimal number of test iterations.

UnityPGTA is an automated testing tool that analyzes Unity engine source code and scene data to deploy a reinforcement learning agent capable of detecting system defects. Unlike traditional automation, it requires zero human intervention for writing test scripts or scenario graphs. It treats the occurrence of error logs or exceptions as a positive reward for the agent, optimizing the model to find bugs.

AirSim is an open-source flight simulator designed for training and testing drones in a virtual environment to avoid the high cost and risk associated with real-world drone flights. It integrates a precise physics engine and sensor models on top of Unreal Engine's realistic graphics to provide a high-fidelity simulation environment suitable for Software-in-the-Loop (SIL) automation and research.

TitanFuzz is a zero-shot fuzzing approach that leverages Large Language Models (LLMs) to test deep learning (DL) libraries. The core insight is that LLMs trained on vast corpora containing DL API calls can implicitly learn the syntax, semantics, and intricate constraints required to generate valid DL programs. It is one of the first works to demonstrate that LLMs can perform both generation-based and mutation-based fuzzing without requiring manual rule-based grammar specifications.

AdvFuzz is an automated scenario generation framework designed to find genuine defects in Autonomous Driving Systems (ADS). Unlike traditional testing that uses static trajectories or aggressive NPC behavior (e.g., sudden cutting-in), AdvFuzz employs intelligent adversarial NPCs that interact dynamically with the EGO vehicle while adhering to traffic laws. This approach ensures that detected violations are caused by the ADS's own software bugs or failures rather than unrealistic NPC behavior.