Blog | AI21

Skip to Main Menu
Skip to Main Content
Skip to Footer

Oct 25, 2026

Why we built a benchmark before training for instruction following

Standard instruction-following rewards check the rules, not the task. Here’s why we built LIFT-Bench first, and what changed when we…

Oct 14, 2026

The hidden mismatch: Evaluating coding agents on the work developers actually do

Oct 6, 2026

The verifiability litmus test for agent design

Oct 4, 2026

From manual negotiation to automated scheduling: How AI21 manages its GPU fleet with Kueue

Aug 19, 2026

You don’t need a frontier model. You need a verifier.

Jul 15, 2026

Better and cheaper together: Open models explore, frontier models patch

Jul 7, 2026

Improving Best-of-N with Budget-Aware Execution for SWE Agents

Jun 24, 2026

Tipping the scales: Merging weak agents into a state-of-the-art deep researcher

Jun 4, 2026

First scale, then enrich: How the right execution strategy helped us reach state-of-the-art on SWE-rebench

May 13, 2026

Reproducing Variance: Caching in Agentic LLM Pipelines

Apr 28, 2026

Reaching SOTA Performance Without Breaking the Bank

All That Glitters: When "Gold-Like" Answers Mask Functional Failures on Coding Agent Benchmarks

Apr 14, 2026

All that glitters: When “gold-like” answers mask functional failures on coding agent benchmarks

Stride and Prejudice: How a 32-bit overflow corrupted a CUDA kernel (and stayed hidden for weeks)

Mar 25, 2026

Stride and prejudice: How a 32-bit overflow corrupted a CUDA kernel (and stayed hidden for weeks)

1 2