Researchers have introduced LIFT, a new framework for evaluating and training instruction-following capabilities in language models. LIFT includes Lift-Bench, a benchmark designed to mitigate reward hacking by using scoped constraints and distinguishing between main task and instruction-following objectives. The framework also features Lift-Train, a training counterpart that is less susceptible to reward hacking than existing methods. Initial experiments show that current frontier models struggle with the more challenging aspects of Lift-Bench, and the publicly available benchmark and training data aim to provide finer control over constraint density and accessibility. AI
IMPACT Introduces a new benchmark and training method to improve language model instruction following and reduce reward hacking.
RANK_REASON The cluster describes a new academic paper introducing a novel benchmark and training framework for language models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
