AIRA_2: Breaking Bottlenecks in AI Research Agents

Mornings With Markman - April 1st, 2026

Markman Capital Insight

Apr 01, 2026

AI agents crossed a major milestone last week, according to r...

## 正文

AIRA_2: Breaking Bottlenecks in AI Research Agents

Mornings With Markman - April 1st, 2026

Markman Capital Insight

Apr 01, 2026

AI agents crossed a major milestone last week, according to researchers at Facebook. Their AIRA_2 model bests humans at the toughest machine learning problems. It often wins by more than 2 points after a single day of work. This big advance lets these agents run experiments without constant human tweaks. It is a massive shift for the industry.

Previously, AI agents worked like junior scientists. They stalled when they encountered a bad test. They also could not work through compute glitches. AIRA_2 fixes these issues with multiple GPUs running in parallel. It also scales work linearly as hardware grows. No more waiting in line for one machine. Speed is now a built-in feature.

Tests used MLE-bench, a set of 30 real-world data science contests. These mimic challenges that human experts face when building machine learning models from scratch. AIRA_2 hit a 71.8% rank after 24 hours. That tops the old leader at 69.9%. By day three, it reached 76%. Humans rarely crack 70% on these difficult contests.

Smart evaluation drives these gains. Past agents failed because they relied on noisy data. AIRA_2 hides test data until the end. Clean feedback lets agents iterate like pros. They debug co