Team hugang11 at SemEval-2026 Task 1, Subtask A (Chinese): A CoT-SFT, Teacher-Constructed DPO, and Deterministic Post-processing Pipeline for Humor Generation

AbstractWe present a system for SemEval-2026 Task 1, Subtask A (Chinese), which addresses humor generation with a three-stage pipeline combining CoT-SFT, teacher-constructed DPO, and deterministic post-processing. Built on Qwen2.5-7B-Instruct-bnb-4bit, the system achieved a live leaderboard rating of 991 and ranked in the second group. Our results suggest that robust inference-time control is as important as alignment-oriented training for humor generation.