WikiIQ
WikiIQ is a benchmark designed to test the ability of a model to navigate in the Wikipedia web of articles. The task consists of reaching a goal article from a given source article in the network by following links, constrained by fixed budgets. WikiIQ consists of task instances based on a static snapshot of the Wikipedia link graph, with standard and structural versions of challenge partitions.
Performance of the models is measured in terms of success of reaching the goal article, efficiency of the path, and shortest path finding ability. The problem is a proxy for information retrieval-driven goal-oriented navigation, where agents have to decide about the most promising next actions, perform search under resource-constraints, and cope with mistakes in the huge link-graph of a large information network.
I'd consider this benchmark saturated once top models start hitting scores of 95+. Scores aren't calculated strictly off of percentage of tasks solved (some minor weighting and path efficiency etc. are considered) but for all intents and purposes can be considered as such. So it looks like we have a ways to go!
The benchmark is currently closed source, plans for open sourcing soon
Results
| Model | Effort | Score | Provider |
|---|---|---|---|
| Gemini 4 Neon | Xhigh | 54 | |
| Claude Fable 5.2 | Max | 52 | Anthropic |
| Gemini 4 Carbon | High | 52 | |
| GPT-6.2 Astra | Max | 50 | OpenAI |
| Claude Opus 5.5 | Max | 49 | Anthropic |
| GPT-6 Astra | Max | 46 | OpenAI |
| GPT-6.2 Sol | Max | 45 | OpenAI |
| GPT-6.1 Sol | Max | 43 | OpenAI |
| Gemini 4 Hydrogen | Xhigh | 40 | |
| Gemini 4 Argon | High | 40 | |
| Claude Sonnet 5.5 | Max | 37 | Anthropic |
| GPT-6.1 Terra | Max | 34 | OpenAI |
| Kimi K3.1 Pro | Max | 30 | Kimi |
| Qwen4 | Max | 29 | Alibaba |
| Grok 4.7 | Xhigh | 25 | SpaceXAI |
| Claude Haiku 5.5 | Max | 24 | Anthropic |
| DeepSeek V4.1 Flash 1017 | Max | 15 | DeepSeek |
| GLM-5.3 | Max | 13 | Z AI |
