A progressive data-construction pipeline with GPT-4o-based reordering, followed by DPO and curriculum training, improves soft constraint following on FollowBench, CFBench, and IFEval.
This task im- poses a higher demand on the model’s ability to generate well-organized and structured outputs, aligning with the re- quired presentation structure
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Step-by-Step Mastery: Enhancing Soft Constraint Following Ability of Large Language Models
A progressive data-construction pipeline with GPT-4o-based reordering, followed by DPO and curriculum training, improves soft constraint following on FollowBench, CFBench, and IFEval.