AdaJudge: Adaptive Multi-Perspective Judging for Reward Modeling
Researchers have introduced AdaJudge, a novel framework designed to enhance the accuracy of reward modeling in large language models. This approach tackles limitations in current static pooling strategies by adapting both the model's representations and its aggregation methods. AdaJudge employs gated refinement blocks to create discrimination-oriented representations and an adaptive multi-view pooling module for dynamic evidence combination. Experiments on RM-Bench and JudgeBench demonstrate AdaJudge's superior performance compared to existing reward models and pooling baselines. AI
IMPACT Enhances LLM alignment by improving reward modeling, potentially leading to more nuanced and human-aligned AI behavior.