AI21 Labs releases LIFT-Bench, scoping each constraint to one document section

AI21 Labs released LIFT-Bench on October 25, 2026, an instruction-following benchmark in which every constraint is scoped to one section of a document and the task content itself is scored.

What was released

• LIFT-Bench: the evaluation benchmark. Each constraint applies to a single section of a document, and scoring covers the task content itself.
• LIFT-Train: a training counterpart containing 34,258 tasks.
• Open pipeline: the pipeline that generated both LIFT-Bench and LIFT-Train is open.

The release is covered by Unite.AI under the headline "AI21 Labs Releases LIFT-Bench to Curb Reward Hacking in RL Training."

The paper

A companion paper, "LIFT: Scoped Constraints for Evaluating and Training Instruction Following," was submitted to arXiv on October 20. Its authors are Eran Goldstein, Yuval Belfer, Yonatan Belinkov, and Daniel Gissin.

Benchmark scores for specific models, licensing terms, and download locations are not yet available.

What this means

LIFT pairs an evaluation set with a training set built from the same open pipeline, so the same constraint design can be used to both measure and train instruction following. Scoping each constraint to one document section and scoring the task content itself defines how the benchmark checks a response. The Unite.AI headline links the release to reward hacking in reinforcement learning training.