Tag
一个 Apache-2.0 开源桌面应用 Open-Science 试图让 AI 辅助科研像带校验和与版本控制的代码一样可追溯：AI 代理阅读论文、调用科学数据库、运行 Python/R 代码，并将输入数据、代码与运行历史一并保存为可分享的 .science 文件，实现本地优先、模型无关的可复现研究。
This paper proposes using controlled synthetic pretraining with the O'PRIOR task generator to make training-data attribution experimentally verifiable for tabular foundation models, showing that top-attributed task removal measurably degrades held-out performance and revealing that provenance association and interventional faithfulness can diverge.
The article explores designing memory systems for AI agents that are auditable by humans, suggesting fields like provenance, scope, and expiration rules to maintain clarity and prevent stale information.
This paper introduces a provenance-guided incremental learning framework to handle rule-induced concept shifts, where target definitions change, and evaluates it on a new benchmark with improved efficiency and accuracy.
OriginBlame is a record- and token-level data provenance system that propagates author identity through AI training data pipelines, enabling precise forget sets for machine unlearning. It eliminates over-deletion from dataset-level systems and improves unlearning effectiveness.
This paper introduces a novel dataset watermarking method for closed LLMs that uses co-occurrence patterns of word pairs to provably detect if proprietary data was used in training, even when it constitutes a small fraction of the dataset.