Where Do Large Language Models Fail on Competitive Programming? A Taxonomy of Failures by Algorithm Type and Difficulty Rating

This is a collection of recent AI/SE research (arXiv 05 Jun 2026) presenting benchmarks, taxonomies, and datasets that reveal reliability, safety, and deployment gaps in LLM-driven software and agent ecosystems. Key items: an empirical failure taxonomy for LLMs on competitive programming showing CoT can worsen algorithmic correctness; DeployBench and SWE-InfraBench exposing fragile artifact deployment and IaC editing failures; a first empirical taxonomy of runtime faults in MCP (Model Context Protocol) servers that explicitly calls out security-validation and integration fault classes; REStack