With Qwen3-8B greedy decoding on eight LongBench-v1 tasks, four tasks agreed within 0.05 points, the largest task deviation was 0.50 points, and the task-macro average was 51.92 for AutoRegressive versus 51.72 for ASPIRE (−0.20 points) across 1,460 examples. The authors attribute the small differences to floating-point nondeterminism in batched-kernel tie-breaking.

With Qwen3-8B greedy decoding on eight LongBench-v1 tasks, four tasks agreed within 0.05 points, the largest task deviation was 0.50 points, and the task-macro average was 51.92 for AutoRegressive versus 51.72 for ASPIRE (−0.20 points) across 1,460 examples. The authors attribute the small differences to floating-point nondeterminism in batched-kernel tie-breaking. · CiteArk

测量结果

已评估 0/36 项

获得支持 0有争议或冲突 0尚未评估 36

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
指标与条件	论文报告	最近可用实测	证据判断
task: HotpotQA · model: Qwen3-8B · decoding: greedy	200 examples	—	尚未评估
task: HotpotQA · model: Qwen3-8B · method: AutoRegressive · decoding: greedy	62.98%	—	尚未评估
task: HotpotQA · model: Qwen3-8B · method: ASPIRE · decoding: greedy	62.95%	—	尚未评估
task: HotpotQA · model: Qwen3-8B · decoding: greedy	-0.03 percentage points	—	尚未评估
task: TriviaQA · model: Qwen3-8B · decoding: greedy	200 examples	—	尚未评估
task: TriviaQA · model: Qwen3-8B · method: AutoRegressive · decoding: greedy	90.01%	—	尚未评估
task: TriviaQA · model: Qwen3-8B · method: ASPIRE · decoding: greedy	90.01%	—	尚未评估
task: TriviaQA · model: Qwen3-8B · decoding: greedy	0 percentage points	—	尚未评估
task: NarrativeQA · model: Qwen3-8B · decoding: greedy	80 examples	—	尚未评估
task: NarrativeQA · model: Qwen3-8B · method: AutoRegressive · decoding: greedy	26.76%	—	尚未评估
task: NarrativeQA · model: Qwen3-8B · method: ASPIRE · decoding: greedy	26.81%	—	尚未评估
task: NarrativeQA · model: Qwen3-8B · decoding: greedy	0.05 percentage points	—	尚未评估

报告指标与观测值

Assessment（0）

关联运行（0）

测量结果