Memory-Gym-trained Researcher behavior transfers to LongCodeQA without code-domain training or adaptation: SFT raises accuracy in every context subset, and subsequent Hint-guided GRPO raises it further.

Memory-Gym-trained Researcher behavior transfers to LongCodeQA without code-domain training or adaptation: SFT raises accuracy in every context subset, and subsequent Hint-guided GRPO raises it further. · CiteArk

测量结果

已评估 0/12 项

获得支持 0有争议或冲突 0尚未评估 12

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
contextSubset: 64K · trainingStage: Untrained	56.58%	—	尚未评估
contextSubset: 128K · trainingStage: Untrained	55.43%	—	尚未评估
contextSubset: 256K · trainingStage: Untrained	49.23%	—	尚未评估
contextSubset: Overall · trainingStage: Untrained	54.08%	—	尚未评估
contextSubset: 64K · trainingStage: SFT only	68.42%	—	尚未评估
contextSubset: 128K · trainingStage: SFT only	64.13%	—	尚未评估
contextSubset: 256K · trainingStage: SFT only	64.62%	—	尚未评估
contextSubset: Overall · trainingStage: SFT only	65.67%	—	尚未评估
contextSubset: 64K · trainingStage: SFT+RL	76.32%	—	尚未评估
contextSubset: 128K · trainingStage: SFT+RL	76.09%	—	尚未评估
contextSubset: 256K · trainingStage: SFT+RL	73.85%	—	尚未评估
contextSubset: Overall · trainingStage: SFT+RL	75.54%	—	尚未评估

Memory-Gym-trained Researcher behavior transfers to LongCodeQA without code-domain training or adaptation: SFT raises accuracy in every context subset, and subsequent Hint-guided GRPO raises it further.

报告指标与观测值

Assessment（0）

关联运行（0）

测量结果