PingAn-NLP at SemEval-2026 Task 9: Multi-Stage Alignment via GRPO and Tiered Ensemble Voting for Multilingual Polarization Detection

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

Summary

The PingAn-NLP system competed in SemEval-2026 Task 9 Subtask 3, focusing on identifying polarization across 18 languages. This system employs a tiered optimization framework that integrates Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO). Key innovations include synthetic reasoning distillation from a 235B teacher model, a Smart-Tradeoff reward function designed to address extreme label imbalance, and a tiered ensemble voting strategy that adaptively adjusts decision thresholds based on language resources. The 8B-GRPO-Vote system demonstrated strong internal performance in English and Hindi, and officially secured second place in the Bengali, English, Odia, and Turkish competitions.

Key takeaway

For NLP engineers developing multilingual classification systems, the PingAn-NLP approach offers a robust framework for handling diverse languages and label imbalances. You should consider integrating Group Relative Policy Optimization (GRPO) with Supervised Fine-Tuning (SFT) and explore adaptive ensemble voting strategies. This can significantly improve performance in low-resource languages and mitigate issues like extreme label skew, as demonstrated by its second-place finishes.

Key insights

A multi-stage alignment system effectively detects multilingual polarization using GRPO and tiered ensemble voting.

Principles

• Combine SFT with GRPO for optimization.
• Distill reasoning from larger teacher models.
• Adapt decision thresholds by language resources.

Method

The system uses a tiered optimization framework, integrating SFT with GRPO, synthetic reasoning distillation, a Smart-Tradeoff reward function, and tiered ensemble voting for multilingual polarization detection.

In practice

• Apply GRPO for fine-tuning language models.
• Use teacher models for synthetic data generation.
• Implement adaptive ensemble voting for diverse languages.

Topics

• Multilingual NLP
• Polarization Detection
• SemEval-2026
• Group Relative Policy Optimization
• Ensemble Voting
• Supervised Fine-Tuning
• Reasoning Distillation

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• mdok-style at SemEval-2026 Task 9: Finetuning LLMs for Multilingual Polarization Detection — Finetuning mid-sized LLMs with QLoRA and augmented data improves multilingual online polarization detection.
• Better Literary Translation: A Multi-Aspect Data Generation and LLM Training Approach — Multi-aspect data generation and iterative refinement significantly enhance LLM literary translation.
• Multi-Stage Training for Abusive Comment Detection in Indic Languages — Multi-stage training with language preprocessing and model ensembles improves abusive comment detection while minimizing false positives.
• IHUBERT: Vector-Based Semantic Deduplication and Domain-Balanced Pretraining for Persian Resources — Semantic deduplication and domain-balanced pretraining significantly enhance Persian PLM performance across diverse NLU tasks.
• AFRILANGTUTOR: Advancing Language Tutoring and Culture Education in Low-Resource Languages with Large Language Models — Dictionary-based synthetic data generation and combined SFT+DPO significantly enhance LLM tutoring for low-resource African languages.
• Macro: Enhancing Multilingual Counterfactual Explanations through Alignment-as-Preference Optimization — Preference alignment, specifically DPO, effectively balances validity and minimality in multilingual counterfactual explanation generation for LLMs.

Open in AIssential →

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