[Bug] GPQA score low with thinking enable

[qianyp18]

Prerequisite

• I have searched Issues and Discussions but cannot get the expected help.
• The bug has not been fixed in the latest version.

Type

I'm evaluating with the officially supported tasks/models/datasets.

Environment

env: H20
inference: sglang.v0.5.6
model: Deepseek V3.2
opencompass: 0.5.1.post1

Reproduces the problem - code/configuration sample

None

Reproduces the problem - command or script

python3 run.py --models my_api.py --datasets gpqa_gen.py --debug

dataset	version	metric	mode	opencompass.models.OpenAISDK_opencompass-0.5.1.post1_Deepseek_V32
GPQA_diamond	5aeece	accuracy	gen	79.29

python3 run.py --models my_think_api.py --datasets gpqa_gen.py --debug

dataset	version	metric	mode	opencompass.models.OpenAISDK_opencompass-0.5.1.post1_Deepseek_V32
GPQA_diamond	5aeece	accuracy	gen	76.26

For the result of thinking, after manual review, the score should be 85.85 ；
I checked the evaluation in gpqa.py and found that the regular expression is unreasonable in some scenarios: it should take the last result, but instead it takes the first matched one.

For example, "I'll answer: ANSWER: C" will return A instead of C.

20251210_161518_h20_gpqa.tar.gz
20251210_211428_h20_gpqa_thinking_32k.tar.gz

Reproduces the problem - error message

None

Other information

No response