Evolve on Graph Ablation - GrailQA F1: leaderboard
Evolve on Graph v1 evaluates knowledge-graph reasoning using direct answers, chain of thought, self-consistency, ReAct, retrieval, or separately trained Solver/Proposer systems. It uses Qwen2.5-3B/7B-Instruct and Llama3.1-8B-Instruct. Training lasts 140 steps, with 5 rollouts per task and at most 15 assistant turns. Hit@1 requires an exact normalized match to any gold answer; F1 measures overlap between predicted and gold answer lists. Freebase, Wikidata, and task-specific KG-R1 training scopes remain distinct. Trained weights are not aliased to the instruction backbone. Full EoG and ReAct duplicate Table 2 and are not re-emitted. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | EoG study Qwen2.5-7B-Instruct Triangular reward | 62.1 | |
| 2 | EoG study Qwen2.5-7B-Instruct Adversarial reward | 61.01 | |
| 3 | EoG study Qwen2.5-7B-Instruct w/o query exec. (Random Walk) | 60.41 | |
| 4 | EoG study Qwen2.5-7B-Instruct w/o Solver update | 55.69 | |
| 5 | EoG study Qwen2.5-7B-Instruct w/o Proposer update | 55.18 |
Interactive version: theaggregate.ai/benchmark?slug=evolve-on-graph-ablation-grailqa-f1 · How It Works · Data refreshed daily, snapshot 2026-10-28.