Personalized RewardBench - Society and Culture: leaderboard

Personalized preference judgment on the 1,074 Society and Culture test pairs of Personalized RewardBench (2026): for LaMP-QA queries, the chosen response follows the user-specific rubric aspects while the rejected one, of similar general quality, does not. Pairwise accuracy in percent (scalar models must score the chosen response higher, generative models pick it); higher is better.

Metric: Pairwise accuracy (%) on the 1,074 Society and Culture test pairs of Personalized RewardBench, where the chosen response follows the user-specific rubric aspects and the rejected one avoids them while keeping general quality (human-validated); scalar models must score the chosen response strictly higher, generative models pick one of the two in shuffled order; no user profile in the input; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Flash75.51#175
2InternLM2-7B-Reward74.95
3Claude Sonnet 4.673.56#173
4Skywork-Reward-V2-Llama-3.1-8B71.88
5R3-Qwen3-14B-14k71.88
6RM-R1-Qwen2.5-Instruct-14B69.65
7Skywork-Reward-V2-Llama-3.2-3B69.18
8R3-Qwen3-8B-14k68.9
9mR3-Qwen3-4B68.53
10R3-Qwen3-4B-14k68.16
11Skywork-Reward-V2-Llama-3.2-1B67.88
12RM-R1-Qwen2.5-Instruct-7B67.69
13InternLM2-20B-Reward67.13
14GPT-5.166.76#240
15mR3-Qwen3-8B66.29

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=personalized-rewardbench-society-and-culture · How It Works · Data refreshed daily, snapshot 2026-11-21.