Attention Dropout Clean GSM8K - Raw Accuracy: leaderboard
Attention Dropout Clean GSM8K - Raw Accuracy: Raw exact-answer accuracy over 250 derivations (%). Attention dropout arXiv 2611.14209v1 Table 1: clean no-dropout controls over 250 paired GSM8K derivations. Zero-shot user prompt, no system or exemplars; deterministic greedy, max_new_tokens=512, Qwen enable_thinking=False. Raw accuracy includes answer-format failures. Intervention arms and conditional-on-answer accuracy are not substituted for baseline capability.
Metric: Raw exact-answer accuracy over 250 derivations (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 8B | 93.6 | #1838 |
| 2 | Llama 3.1 8B Instruct | 87.2 | #4277 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=attention-dropout-clean-gsm8k-raw-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.