88Z AI, 4 August — Tencent Hunyuan, together with Tsinghua AIR and Southeast University, has released E-Bench, an agent evaluation benchmark. Built on fully synthetic virtual environments modeled after Honor of Kings, QQ Music, and Tencent Meeting, it includes 323 state-change tasks and over 76,000 data entries. By incorporating information and tool gaps, it prevents models from relying on memory shortcuts. Evaluations show that 11 leading models achieved an average success rate of only 54.56%, with the top model
At a glance
- E-Bench builds fully synthetic environments based on three Tencent apps, using dual-gap designs to prevent memory cheating
- Eleven leading models averaged a 54.56% success rate, with the top model's Pass³ metric staying below 60%
The update
Tencent Hunyuan, in collaboration with Tsinghua AIR and Southeast University, has launched E-Bench, a new benchmark for evaluating AI agents. Modeled after Honor of Kings, QQ Music, and Tencent Meeting, this fully synthetic virtual environment comprises 323 state-changing tasks and more than 76,000 data rows. To prevent models from taking shortcuts via memorization, E-Bench incorporates specific designs for information and tool gaps. In evaluations covering 11 cutting-edge models, the average success rate was just 54.56%. The best-performing model achieved an Avg@3 score of 73.79%, while its Pass³ rate remained below 60%, indicating room for improvement in handling complex tasks.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 腾讯混元Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.