FinDendro - Controlled Fact Accuracy: leaderboard
Dated financial fact recall and prediction under explicit time prompts. FinDendro v1 contains 2,226 dated US facts from 1999–2026. Task 1 accuracy uses 1,340 controlled facts that suppress persistence shortcuts. Table 4 also reports Task 2 Predict and Task 3 dated-Predict accuracy on their native prompt cohort. Each prompt is evaluated once without external tools, using temperature zero where supported and at most 1,024 output tokens, increased to 4,096 for GPT-5. Locally hosted Qwen3-8B and Qwen3.5-9B use greedy decoding with thinking disabled. API snapshot revisions are unreported. Claim frequencies and subsets selected by each model’s claims or knowledge window are diagnostics, not a shared capability cohort. Question access: Primary result reviewed; exact scored release availability not confirmed.
Metric: Accuracy (fraction). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (Benchmark Sweep Run: source=FinDendro; API; temperature 0 where supported; native token limit) | 0.76 | |
| 2 | DeepSeek-V3.2 (Benchmark Sweep Run: source=FinDendro; API; temperature 0 where supported; native token limit) | 0.54 | |
| 3 | GPT-4o (Benchmark Sweep Run: source=FinDendro; API; temperature 0 where supported; native token limit) | 0.5 | |
| 4 | Llama-3.3-70B-Instruct (Benchmark Sweep Run: source=FinDendro; API; temperature 0 where supported; native token limit) | 0.43 | |
| 5 | Qwen3.5-9B (Benchmark Sweep Run: source=FinDendro; greedy; thinking disabled) | 0.37 | |
| 6 | Qwen3-8B (Benchmark Sweep Run: source=FinDendro; greedy; thinking disabled) | 0.36 |
Interactive version: theaggregate.ai/benchmark?slug=findendro-controlled-fact-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-06.