Tencent Hunyuan and Fudan Launch ExplorationBench to Measure Exploration in LLMs

Tencent Hunyuan and researchers at Fudan University have introduced ExplorationBench, an evaluation suite for how well AI systems "explore." The announcement appeared on the official @TencentHunyuan account on X, and @TencentAI_News then reposted it and added its own analysis. The authors are going after one of the hardest agent traits to measure: the ability to act when the right path is not known in advance. The primary source material we have on our desk contains no numerical model results, and we will not invent any.

What we know so far

There are few facts, but they are clear. The benchmark is called ExplorationBench. Its stated goal is to measure how an AI system "explores." The work is done by a team from Fudan, which has released CL-bench before, with participation from Tencent Hunyuan. For Tencent this is the second joint evaluation project with a university partner, which points to an established pattern rather than a one-off.

Key details of the announcement:

• Name: ExplorationBench.
• What it measures: the exploratory behavior of AI systems, not the accuracy of an answer to a fixed question.
• Authors: a Fudan University team and Tencent Hunyuan.
• Predecessor: CL-bench from the same Fudan group.
• Publication channels: the X accounts of Tencent Hunyuan and Tencent AI.
• Not disclosed in the available text: the number of tasks, the format of the environments, the metrics, the list of tested models, the license and a link to the code.

Why exploration became the bottleneck

Most popular tests check the outcome: a correct answer, passing unit tests, a solved task. Agentic scenarios work differently. An agent lands in an environment with incomplete information and has to decide what to check, where to go and when to stop. The classic exploration-exploitation dilemma has long been familiar from reinforcement learning, but it shows up in a new form in LLM agents. A model can confidently exploit the first successful hypothesis and miss a better one. Or it can gather information endlessly and never get around to acting.

A final success metric does a poor job of telling these failures apart. Two agents receive the same score even though one found the solution systematically and the other stumbled on it by guesswork. A separate exploration benchmark is needed to see the process itself: how completely the agent covers the space, how often it repeats actions, and how it reacts to new information.

What the announcement leaves out

The main question for ExplorationBench is how exactly it defines exploration. Everything depends on that. If the environments are deterministic and finite, the benchmark risks turning into a brute-force test. If they are open-ended, reproducibility and scoring become a problem. It also needs to be clear whether the suite separates exploration ability from a model's general intelligence. Strong models tend to win across the board, and without control conditions you cannot prove the contribution of exploratory behavior specifically.

The second question is openness. If the code and data are released publicly, the benchmark can become a shared tool. If not, it will remain a showcase for Hunyuan's in-house models. The CL-bench experience shows that the group is willing to release evaluation work publicly, but we cannot yet confirm the timing or terms for the new suite.

How to use this now

A team building agents can already add process metrics to its own evaluation: the number of repeated actions, the share of unique states, and the point at which the agent first consults the right source. Once ExplorationBench ships with code, it will be easy to compare that data against its results. When choosing a model for an agentic product, it is worth waiting for published tables and looking not at a single average score but at the breakdown by environment type.

When you read the eventual report, check three things: whether it includes independent runs of competitors' models, whether it reports variance between runs, and whether it publishes agent trajectories. Without these, any ranking will remain advertising rather than measurement.