RECAST: Scaling Complex Instruction Following Beyond 10 Constraints
September 4, 2026
RECAST is a framework for synthesizing datasets that force LLMs to follow more than 10 simultaneous constraints per instance. It extracts constraints from real-world prompt-response pairs to push the boundaries of complex instruction following.
HOW THIS AFFECTS YOU
●
builderYou can use this to stress-test how well your agents handle highly complex, multi-requirement user instructions.
●
researcherThis provides a way to evaluate the true upper limits of model reasoning in instruction-following tasks.