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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Team hugang11 presented a system for SemEval-2026 Task 1, Subtask A (Chinese), specifically designed for humor generation. The system utilizes a sophisticated three-stage pipeline, which combines CoT-SFT (Chain-of-Thought Supervised Fine-Tuning), teacher-constructed DPO (Direct Preference Optimization), and a deterministic post-processing step. This architecture, built on the Qwen2.5-7B-Instruct-bnb-4bit model, achieved a live leaderboard rating of 991 and secured a position in the second competitive group. A key finding from their work suggests that robust inference-time control is as important as alignment-oriented training for successful humor generation.

Key takeaway

For NLP engineers developing humor generation systems, you should prioritize integrating robust inference-time control mechanisms alongside your alignment-oriented training. Your models, even those based on powerful LLMs like Qwen2.5-7B, will benefit significantly from a multi-stage pipeline incorporating techniques like CoT-SFT, DPO, and deterministic post-processing to achieve higher quality and more controlled humorous outputs.

Key insights

Effective humor generation requires balancing alignment training with robust inference-time control.

Principles

• Inference-time control is crucial for humor.
• Multi-stage pipelines enhance generation.
• DPO can be teacher-constructed.

Method

The system uses a three-stage pipeline: CoT-SFT for initial training, teacher-constructed DPO for alignment, and deterministic post-processing for refined output.

In practice

• Apply CoT-SFT for complex generation tasks.
• Implement DPO with teacher-guidance.
• Use post-processing for output refinement.

Topics

• Humor Generation
• SemEval-2026
• CoT-SFT
• Direct Preference Optimization
• Qwen2.5-7B
• Post-processing
• Inference Control

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Direct Preference Optimization for Chatbot Fine-Tuning: An Empirical Study — DPO offers a computationally efficient and simplified pipeline for competitive LLM chatbot fine-tuning.
• Automatic Generation of Highlights for Academic Paper Via Prompt-based Learning — Prompt-based learning with LLMs can generate academic paper highlights effectively without extensive labeled data.
• LAI #123: Claude Code’s Codebase Was Accidentally Leaked — Effective LLM prompting and agent design require careful consideration of example order, tool context, and subagent delegation.
• NebulaExp-8B: An Empirical Post-Training Pipeline via Full-Scale Ablation Research — An ablation-driven pipeline for 8B LLMs improves instruction following and reasoning via structured SFT, RL, and OPD/MOPD.
• Improving Cross-Lingual Factual Recall via Consistency-Driven Reinforcement Learning — GRPO significantly improves cross-lingual factual recall and consistency in LLMs by fostering shared representations, outperforming SFT.
• LEAF: Growing Trees Without Branching for Speech-Aware Large Language Model Post-Training — LEAF is a retrospective tree-based RL method for speech-aware LLM post-training that refines credit assignment.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.