Large language models struggle to follow instructions in sequence, according to a new benchmark built to test exactly that.
Researchers introduced CDR-Bench, a benchmark of 3,462 tasks drawn from four real-world data refinement domains and covering 29 distinct processing operators. The benchmark tests models at three levels of difficulty: single-step atomic tasks, multi-step tasks where order does not matter, and multi-step tasks where it does. Experiments on more than 10 state-of-the-art LLMs showed a consistent pattern — performance drops sharply as tasks become compositional, and success rates collapse when the correct order of operations is required. The benchmark uses deterministic reference outputs, so scoring is exact rather than subjective.
The finding matters because data refinement — cleaning, transforming, and restructuring text through a defined sequence of steps — is exactly the kind of work enterprises want to hand off to AI pipelines. If a model silently skips a step, reorders operations, or drifts from the recipe, the output is wrong in ways that are hard to catch downstream. CDR-Bench gives researchers a concrete way to measure that failure mode rather than assume it away.
Existing benchmarks either test text editing in isolation or bundle it with code execution, which muddies the results. The consistent failure patterns here suggest that procedural faithfulness — doing the right thing in the right order, reliably — is a gap that scaling alone has not closed.