AI Research Monthly: Feb-Mar 2026 — 21 Findings With Hard Data (The Comprehensive Edition)

AI Research Monthly: Feb-Mar 2026 — 21 Findings With Hard Data Your friend who reads AI papers so you don't have to. Only findings with real numbers — no hype, no "vibe coding is a trend". This is the comprehensive edition covering every major benchmark, comparison, and evaluation from the past two months. Part 1: The Exams Are Broken — Benchmark Trust Crisis 1. The Most-Used AI Coding Test Had Broken Answer Keys What is SWE-bench Verified? It's a benchmark (standardized test) for measuring how well AI can write code. It takes 500 real GitHub issues — actual bugs reported by real developers in open-source projects — gives the AI the buggy source code, and asks it to write a patch that fixes the bug. Then it runs the project's own test suite (automated tests) to check if the fix works. Your score, called "resolve rate," is what percentage of 500 bugs you fixed correctly. Think of it as a coding exam where the questions are real-world bugs, not textbook exercises. What happened: OpenAI's