IBA-Bench - Writing and Design: leaderboard

Metric: Checkpoint satisfaction (%; share of hidden preference checkpoints satisfied by the final action, writing and design domain; mean of three LLM judges: DeepSeek-V3.2, Qwen3-4B-Instruct and Qwen3-30B-A3B-Instruct; retrieval-augmented baseline with bge-m3 top-3 chunks). Source: arxiv.org. Saturation forecast: Around December 2026. 10 models tracked.

Top models

#ModelScore
1GPT-5.172.9
2GPT-5 Mini70.5
3QwQ-32B68.7
4DeepSeek V3.268.1
5Qwen 3 4B Instruct65.2
6DeepSeek V365.1
7GPT-4o63.5
8GPT-4o Mini57.7
9Qwen 2.5 7B Instruct52.7
10GLM-4 9B Chat51.4

Interactive version: theaggregate.ai/benchmark?slug=iba-bench-writing-and-design · How It Works · Data refreshed daily, snapshot 2026-09-26.