arXiv · Computation and Language· Nishanth Nayakanti, Prasang Gupta, Ashutosh Bilthare, Kevin Paul·· 4 hr agoAI score38
仅有结论而无标注证据：拒绝采样还是仅标签后训练用于证据恢复？
Verdicts Without Annotated Evidence: Rejection Sampling or Label-Only Post-Training for Evidence Recovery?
AI brief
研究在 ContractNLI 上衡量小语言模型仅凭结论后训练时能恢复多少证据，人类证据片段直到评估才放出。匹配结论与吻合证据片段并不等同：六个系统中两项分数仅弱相关且排序不同，accuracy 不足以指导需可审查引用的场景。
Source: arXiv · Computation and Language · arxiv.org