AI21 Labs opens LIFT-Bench for section-scoped instruction following
AI21 Labs’ LIFT-Bench measures compliance with section-scoped and global instructions in long Markdown documents, with an open 784-example benchmark and a 34,258-example training set.
AT A GLANCE
Key points
- LIFT-Bench contains 784 examples across 10 splits and supports up to 80 constraints per example.
- LIFT-Train provides 34,258 training examples plus an open synthesis pipeline.
- The authors report lower susceptibility to reward hacking than with the compared IFTrain data.
What is changing?
AI21 Labs researchers released LIFT-Bench to evaluate instruction following in long documents. The paper, submitted to arXiv on October 20, asks a model to draft a Markdown document from a topic and outline while applying some constraints to specific sections and others across the full document.
The benchmark contains 784 examples across 10 splits and can use up to 80 verifiable or language-model-judged constraints per example. The team also released LIFT-Train, a 34,258-example training set, together with the synthetic-data pipeline.
The paper reports that open and closed frontier models struggle as constraint density increases. The authors say training on LIFT-Train produced results that were less susceptible to reward hacking than the compared IFTrain data. This does not prove that reward hacking has been eliminated; it is a comparison under the study’s evaluation conditions.
Explore the original sources
Source date: 25 October 2026.
Published by AI Radar: 25 Oct 2026, 14:08 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…