Modelpublic

SeerAttention/SeerAttention-Decode-Qwen3-4B-AttnGates

sourceHugging Facemitupdated 1y agoView on Hugging Face

2likes105downloads

Model Card

SeerAttention-R

This repo contains the decode stage AttnGate weights from paper SeerAttention-R. The current support models are:

Results of Reasoning Tasks

Results of reasoning task with different token budgets. All the results are the averaged pass@1 results with 64 sample per query for AIME, 16 samples for GPQA, and 8 samples for MATH-500.

AIME24

Model	2k	4k	6k	8k	Full Attention
Qwen3-4B	55.42	68.75	70.94	72.50	71.25
Qwen3-8B	56.56	72.29	74.22	75.05	74.48
Qwen3-14B	62.24	75.78	78.02	78.65	78.91
DeepSeek-R1-Distill-Qwen-14B	55.78	66.35	67.50	66.82	67.50

AIME25

Model	2k	4k	6k	8k	Full Attention
Qwen3-4B	45.73	57.60	60.20	62.90	66.41
Qwen3-8B	42.60	56.77	60.31	64.17	67.86
Qwen3-14B	46.67	62.66	67.19	69.01	70.21
DeepSeek-R1-Distill-Qwen-14B	38.44	47.19	52.25	50.05	50.00

MATH500

Model	1k	2k	4k	6k	Full Attention
Qwen3-4B	84.80	92.20	93.60	93.60	93.93
Qwen3-8B	82.82	91.53	94.17	94.53	94.43
Qwen3-14B	85.13	93.20	94.77	94.80	95.22
DeepSeek-R1-Distill-Qwen-14B	87.65	92.10	93.05	93.12	93.30

GPQA Diamond

Model	1k	2k	4k	6k	Full Attention
Qwen3-4B	39.61	51.20	55.20	55.90	56.19
Qwen3-8B	37.59	54.32	59.60	60.48	60.54
Qwen3-14B	44.54	59.72	63.76	64.20	65.25
DeepSeek-R1-Distill-Qwen-14B	51.26	56.79	56.41	57.48	57.80