nvidia/Qwen3-Nemotron-14B-BRRM

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:14BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 22, 2025License:nvidia-internal-scientific-research-and-development-model-licenseArchitecture:Transformer0.0K Featherless Exclusive Warm

The nvidia/Qwen3-Nemotron-14B-BRRM is a 14 billion parameter Branch-and-Rethink Reasoning Reward Model developed by NVIDIA, designed to evaluate LLM-generated responses. This model employs a novel two-turn reasoning framework, utilizing adaptive branching to focus on critical evaluation dimensions and branch-conditioned rethinking for targeted deep analysis. It achieves state-of-the-art performance on major reward modeling benchmarks by addressing the "judgment diffusion" problem, making it suitable for integrating into RLHF pipelines for enhanced response quality assessment.

Loading preview...

nvidia/Qwen3-Nemotron-14B-BRRM: A Novel Reward Model

The nvidia/Qwen3-Nemotron-14B-BRRM is a 14 billion parameter Branch-and-Rethink Reasoning Reward Model developed by NVIDIA. Unlike traditional reward models that provide a single scalar evaluation, BR-RM introduces a two-turn reasoning framework to offer a more nuanced and accurate assessment of LLM-generated responses. This model is specifically designed to overcome the "judgment diffusion" problem by adaptively focusing on instance-critical evaluation dimensions.

Key Capabilities

  • Adaptive Branching: Dynamically selects 1-3 critical evaluation dimensions (e.g., Logical Reasoning, Computational Precision) per instance in the first turn.
  • Branch-Conditioned Rethinking: Performs targeted, deep analysis based on the selected dimensions in the second turn, leading to a final comparative judgment.
  • State-of-the-Art Performance: Achieves top results on major reward modeling benchmarks, including 92.1% on RewardBench, 85.9% on RM-Bench, and 74.7% on RMB.
  • RLHF Compatibility: Engineered for seamless integration into standard Reinforcement Learning from Human Feedback (RLHF) pipelines.

Good for

  • Evaluating LLM Response Quality: Provides a robust and detailed mechanism for assessing the quality of responses from large language models.
  • Improving LLM Alignment: Can be used as a reward signal in RLHF to fine-tune LLMs for better alignment with human preferences and specific quality criteria.
  • Complex Evaluation Scenarios: Particularly effective in scenarios where multiple quality dimensions need to be considered and prioritized dynamically.