AI Paper Atlas

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

ICLR 2026 · Large language models › Alignment and preferences · approach: Steering and correcting LLM behavior

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 changedWhereMeasuredEffect
use generative RMs instead of classifier-based RMs on CARBreward models · CARBaverage accuracy: Qwen3-235B 76.5 vs best classifier Skywork-Reward-Gemma-2-27B 73.0Improved
train generative RM with Think-as-Locals (RLVR with structured criteria) on Qwen2.5-7B-InstructQwen2.5-7B-Instruct · CARB (Arabic/Chinese/Japanese subsets)accuracy: 62.6 -> 78.8Improved
train generative RM with Think-as-Locals on Qwen2.5-14B-InstructQwen2.5-14B-Instruct · CARB (Arabic/Chinese/Japanese subsets)accuracy: 63.6 -> 82.1Improved
train generative RM with Think-as-Locals on Qwen2.5-32B-InstructQwen2.5-32B-Instruct · CARB (Arabic/Chinese/Japanese subsets)accuracy: 71.4 -> 84.3Improved
train generative RM with Think-as-Locals on Qwen2.5-7B-InstructQwen2.5-7B-Instruct · M-RewardBench (Arabic/Chinese/Japanese subsets)accuracy: 77.1 -> 80.4Improved
train generative RM with Think-as-Locals on Qwen2.5-32B-InstructQwen2.5-32B-Instruct · M-RewardBench (Arabic/Chinese/Japanese subsets)accuracy: 86.0 -> 89.5; below Qwen3-235B 92.3Mixed
Think-as-Locals starting from DeepSeek-distilled Qwen 7B baseDpsk-Qwen2.5-7B-Instruct · CARB (Arabic/Chinese/Japanese subsets)accuracy: 68.5 vs 78.8 for Qwen2.5-7B-Instruct baseMixed
Think-as-Locals structured-criteria RL vs retraining RM-R1 with same training data7B generative RM · CARB (Arabic/Chinese/Japanese subsets)accuracy: 75.5 -> 78.8Improved
remove correctness reward from Think-as-Locals RL (ablation)Think-as-Locals · CARB (average)accuracyHurt
remove criteria reward from Think-as-Locals RL (ablation)Think-as-Locals · CARB (average)response entropyHurt
Think-as-Locals with structured criteria vs base generative RMgenerative RM · perturbation sensitivity (Arabic/Chinese/Spanish commonsense)log-probability sensitivity to spurious featuresImproved
use CARB instead of M-RewardBench to select RMs for RLHFreward benchmark · downstream multilingual cultural alignment (GRPO)r^2 with downstream performance: CARB r^2 > 0.6 vs M-RewardBench r^2 < 0.1Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy