skill-rm-unifying-reward-modeling

name	skill-rm-unifying-reward-modeling
description	Skill Reward Model (Skill-RM) framework for unified reward modeling in RL pipelines. Treats reward computation as structured agentic task orchestrating heterogeneous evaluation criteria.

Skill-RM: Unified Reward Modeling via Agent Skill

Core Concept

Reformulate reward modeling as execution of reusable Reward-Evaluation Skill. Instead of static evaluation, treat reward computation as structured agentic task that dynamically selects and aggregates evidence.

Key Innovation

• Unified interface: rule-based verifiers, ground-truth references, procedural checklists, complex rubrics
• Dynamic orchestration: Select evidence tailored to each input's requirements
• Transparency & consistency: Beyond static evaluation, structured task execution

Implementation

1. Reward-Evaluation Skill: Reusable skill module for reward computation
2. Evidence aggregation: Dynamic selection based on input requirements
3. Integration: RL pipelines, RFT, best-of-N selection

Applications

• RLHF pipelines
• Reinforced fine-tuning (RFT)
• Best-of-N selection
• Reward benchmarks

Results

Consistently outperforms traditional judge baselines on reward benchmarks and downstream RL applications.

Source

• arXiv: 2606.03980
• Title: Skill-RM: Unifying Heterogeneous Evaluation Criteria via Agent Skill
• Authors: Tao Chen, Gangwei Jiang, et al. (Qwen team)
• Code: https://github.com/Qwen-Applications/Skill-RM

Activation Keywords

skill-rm, reward modeling, heterogeneous evaluation, RLHF reward, reward orchestration, agentic reward