SUSTech-NLP/UniRRM-8B

UniRRM-8B: Unified Reasoning Reward Model

UniRRM-8B, developed by SUSTech-NLP, is an 8 billion parameter reward model based on Qwen3-8B. Its core innovation lies in its ability to provide unified, multi-faceted evaluation across 103 languages and multiple evaluation paradigms (pairwise, listwise, and pointwise) from a single model. This makes it highly versatile for assessing AI responses in diverse linguistic and comparative contexts.

Key Capabilities

• Multilingual Evaluation: Supports 103 languages, trained on the comprehensive MixReward dataset.
• Unified Paradigms: Handles pairwise, listwise, and pointwise evaluations without needing separate models.
• Adaptive Rubric Generation: Dynamically creates task-generic and instruction-specific evaluation criteria through a staged reasoning process.
• Structured Reasoning: Employs a three-stage pipeline: Deep Analysis, Adaptive Rubric Generation, and Detailed Evaluation, ensuring thorough and objective assessments.
• Efficient Performance: Delivers strong evaluation capabilities within a compact 8B parameter footprint.

Training and Architecture

The model was trained using a two-stage pipeline: Supervised Fine-Tuning (SFT) on the UniRRM-SFT dataset (35,749 samples distilled from GPT-OSS-120B) to establish structured reasoning, followed by Reinforcement Learning with Group Relative Policy Optimization (GRPO) on the UniRRM-RL dataset (32,832 samples). This process optimized for structured output compliance, outcome consistency, and rubric quality. UniRRM-8B uses a Qwen3ForCausalLM architecture with 40960 max position embeddings and bfloat16 precision.

Good for

• Automated evaluation of LLM responses in a multilingual setting.
• Comparing multiple AI-generated responses (pairwise or listwise).
• Detailed, criterion-based scoring of single AI responses.
• Developing robust and adaptable evaluation systems for diverse NLP tasks.