GABench - Open-Ended QA - Insight (Text-Paired Graphs): leaderboard
GABench (2026) gives an LLM agent, run in the OpenClaw harness, graph-analysis questions over real graphs that it must answer by calling graph tools rather than reading the graph from the prompt; a DeepSeek-V4-Flash judge checks the final answer against a ground truth computed by executing the reference tool trajectory. This board covers insight tasks that answer a scenario-based open question on text-paired graphs (the HIV and PCBA molecule collections). Score is the task success rate in percent; higher is better.
Metric: Task success rate (%; share of tasks whose final answer a DeepSeek-V4-Flash judge marks consistent with the verified ground truth; insight scenario-based open-ended graph questions (insight and prediction) with its 10 analysis tools on text-paired graphs (the HIV and PCBA molecule collections, 200 generated tasks per dataset and category); LLM agent in the OpenClaw harness, one run). Source: arxiv.org. Saturation forecast: Around July 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5-Turbo | 14.29 | #231 |
| 2 | Gemini 3 Flash | 11.63 | #164 |
| 3 | DeepSeek V4 Pro | 10.81 | #188 |
| 4 | GPT-5 Mini | 10.11 | #348 |
| 5 | Qwen 3 235B A22B | 0.6 | #623 |
| 6 | GLM-4.5 Flash | 0.57 | #979 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gabench-open-ended-qa-insight-text-paired-graphs · How It Works · Data refreshed daily, snapshot 2026-11-13.