Cost of Mismatch - GSM8K (Pretrained): leaderboard
Cost of Mismatch v1 Table3, original pretrained models on GSM8K, greedy decoding with chain-of-thought and boxed-answer scoring. Qwen2.5 labels denote base checkpoints, as explicitly described in AppendixD and the released base-model reproduction. BOLT and ZO-BOLT trained adapters and best-checkpoint selection are not assigned to their base model. Reported metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 7B | 75.93 | #895 |
| 2 | DeepSeek R1 Distill Qwen 1.5B | 60.97 | #1521 |
| 3 | Qwen 2.5 1.5B | 52.24 | #1412 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cost-of-mismatch-gsm8k-pretrained · How It Works · Data refreshed daily, snapshot 2026-11-05.