AdaJudge: Adaptive Multi-Perspective Judging for Reward Modeling

TL;DR

AdaJudge introduces an adaptive framework for reward modeling that jointly refines representations and dynamically aggregates evidence, significantly improving alignment with human preferences in large language models.

Contribution

It proposes a novel adaptive multi-perspective judging framework that enhances reward modeling by jointly optimizing representations and aggregation methods.

Findings

Outperforms existing reward models on RM-Bench and JudgeBench

Improves alignment with human preferences

Demonstrates effectiveness of adaptive multi-view pooling

Abstract

Reward modeling is essential for aligning large language models with human preferences, yet predominant architectures rely on a static pooling strategy to condense sequences into scalar scores. This paradigm, however, suffers from two key limitations: a static inductive bias that misaligns with task-dependent preference signals, and a representational mismatch, as the backbone is optimized for generation rather than fine-grained discrimination. To address this, we propose AdaJudge, a unified framework that jointly adapts representation and aggregation. AdaJudge first refines backbone representations into a discrimination-oriented space via gated refinement blocks. It then replaces the static readout with an adaptive multi-view pooling module that dynamically routes and combines evidence. Extensive experiments on RM-Bench and JudgeBench show that AdaJudge outperforms strong off-the-shelf reward models and traditional pooling baselines.