IBA-Bench - Work: leaderboard
IBA-Bench (Song et al., 2026) tests whether an LLM agent completes a task while honoring user preferences that are only implicit in a long, noisy interaction history, here in the work domain. Models answer from bge-m3 retrieved history chunks, and the score is the percentage of hidden preference checkpoints the final text or API call satisfies, averaged over three LLM judges; higher is better.
Metric: Checkpoint satisfaction (%; share of hidden preference checkpoints satisfied by the final action, work 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 September 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 63.9 | #349 |
| 2 | GPT-5.1 | 62.4 | #225 |
| 3 | QwQ-32B | 61.6 | #909 |
| 4 | Qwen 3 4B Instruct | 60.4 | #1660 |
| 5 | DeepSeek V3.2 | 60.1 | #403 |
| 6 | DeepSeek V3 | 56.6 | #705 |
| 7 | GPT-4o Mini | 54.9 | #1627 |
| 8 | GPT-4o | 54.1 | #741 |
| 9 | Qwen 2.5 7B Instruct | 51 | #2583 |
| 10 | GLM-4 9B Chat | 47.3 | #2782 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=iba-bench-work · How It Works · Data refreshed daily, snapshot 2026-11-12.