LIFT: Scoped Constraints for Evaluating and Training Instruction Following
Source
arxiv.org
Author
Eran Goldstein, Yuval Belfer, Yonatan Belinkov, Daniel Gissin
Date
Why it matters
Instruction-following benchmarks like IFEval can be gamed and are used as RL rewards. Lift-Bench scores constraints and the main task together, giving a harder check on whether a model really follows instructions.
Key takeaways · AI-distilled
The authors argue that IFEval-style constraints, also used as RL rewards, invite reward hacking that pulls models away from the main task, while earlier richer benchmarks needed expert labor, stayed closed-access and were hard to extend.
Lift-Bench asks a model to draft a document from a topic and a structured outline, with constraints scoped to particular sections to control constraint density. It mixes verifiable and judged constraints and scores the main task alongside them.
Frontier models struggled on the benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition →'s harder subsets, and training on the companion Lift-Train set was much less susceptible to reward hacking than training on the common IFTrain. Both sets and the data pipeline are public.
Terms in this piece · Glossary
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
articleEvalDetectBench: A Benchmark for Measuring Evaluation Awareness in Frontier Language ModelsXinning Li, Kemunto Ochwang'i, Aryasomayajula Ram Bharadwaj, Alexandra Souly, Robert Kirk
articleCompliance, Capability, and Conflict: Benchmarking Multimodal LLMs under System MessagesJuan Yeo, Geewook Kim
articleGardenBench: A lightweight, daily evaluation of LLM capabilitiesBenedict Cummins-Mburu, Mariana Garcia Mejia, Oscar Heath, Sabrina Kreyzerman, Ellie Murray, Arusan Surendiran, Rohan Alexander