Research map / Large language models / Alignment and preferences
Evaluating and Improving Cultural Awareness of Reward Models for LLM Alignment
Hongbin Zhang, Kehai Chen, Xuefeng Bai, Yang Xiang, Min Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use generative RMs instead of classifier-based RMs on CARB | reward models · CARB | average accuracy: Qwen3-235B 76.5 vs best classifier Skywork-Reward-Gemma-2-27B 73.0 | Improved |
| train generative RM with Think-as-Locals (RLVR with structured criteria) on Qwen2.5-7B-Instruct | Qwen2.5-7B-Instruct · CARB (Arabic/Chinese/Japanese subsets) | accuracy: 62.6 -> 78.8 | Improved |
| train generative RM with Think-as-Locals on Qwen2.5-14B-Instruct | Qwen2.5-14B-Instruct · CARB (Arabic/Chinese/Japanese subsets) | accuracy: 63.6 -> 82.1 | Improved |
| train generative RM with Think-as-Locals on Qwen2.5-32B-Instruct | Qwen2.5-32B-Instruct · CARB (Arabic/Chinese/Japanese subsets) | accuracy: 71.4 -> 84.3 | Improved |
| train generative RM with Think-as-Locals on Qwen2.5-7B-Instruct | Qwen2.5-7B-Instruct · M-RewardBench (Arabic/Chinese/Japanese subsets) | accuracy: 77.1 -> 80.4 | Improved |
| train generative RM with Think-as-Locals on Qwen2.5-32B-Instruct | Qwen2.5-32B-Instruct · M-RewardBench (Arabic/Chinese/Japanese subsets) | accuracy: 86.0 -> 89.5; below Qwen3-235B 92.3 | Mixed |
| Think-as-Locals starting from DeepSeek-distilled Qwen 7B base | Dpsk-Qwen2.5-7B-Instruct · CARB (Arabic/Chinese/Japanese subsets) | accuracy: 68.5 vs 78.8 for Qwen2.5-7B-Instruct base | Mixed |
| Think-as-Locals structured-criteria RL vs retraining RM-R1 with same training data | 7B generative RM · CARB (Arabic/Chinese/Japanese subsets) | accuracy: 75.5 -> 78.8 | Improved |
| remove correctness reward from Think-as-Locals RL (ablation) | Think-as-Locals · CARB (average) | accuracy | Hurt |
| remove criteria reward from Think-as-Locals RL (ablation) | Think-as-Locals · CARB (average) | response entropy | Hurt |
| Think-as-Locals with structured criteria vs base generative RM | generative RM · perturbation sensitivity (Arabic/Chinese/Spanish commonsense) | log-probability sensitivity to spurious features | Improved |
| use CARB instead of M-RewardBench to select RMs for RLHF | reward benchmark · downstream multilingual cultural alignment (GRPO) | r^2 with downstream performance: CARB r^2 > 0.6 vs M-RewardBench r^2 < 0.1 | Improved |
Closest papers by meaning
- RMB: Comprehensively benchmarking reward models in LLM alignment (ICLR 2025 · similarity 0.92)
- CultureLLM: Incorporating Cultural Differences into Large Language Models (NeurIPS 2024 · similarity 0.90)
- Interpreting Language Reward Models via Contrastive Explanations (ICLR 2025 · similarity 0.89)
- MM-RLHF: The Next Step Forward in Multimodal LLM Alignment (ICML 2025 · similarity 0.88)
- Theoretical Limits of Language Model Alignment (NeurIPS 2026 · similarity 0.88)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
