인문학
사회과학
자연과학
공학
의약학
농수해양학
예술체육학
복합학
DBpia AI
AI 에이전트
AI 아이디어
AI 리더
AI 뷰어
Citeasy
크롬 서지관리 다운로드
지원사업
학술연구/단체지원/교육 등 연구자 활동을 지속하도록 DBpia가 지원하고 있어요.
커뮤니티
연구자들이 자신의 연구와 전문성을 널리 알리고, 새로운 협력의 기회를 만들 수 있는 네트워킹 공간이에요.
논문 기본 정보
- 저자정보
피인용 0
초록·키워드
Foundation models are commonly fine-tuned to improve performance on specific downstream tasks. However, task-specific adaptation can degrade performance on non-target tasks, a phenomenon known as forgetting that undermines the broad generalization capabilities of foundation models. Prior studies on large language models have reported less forgetting under reinforcement fine-tuning (RFT) than under supervised fine-tuning (SFT), but the underlying mechanism remains unresolved. Isolating this mechanism is difficult in language-model reasoning tasks because multiple valid solution trajectories may yield the same final answer. By contrast, each time-series context window is paired with a single observed future trajectory, providing a more controlled setting for comparing the two paradigms. We therefore compare SFT with RFT instantiated using Group Relative Policy Optimization in two time series foundation models, Chronos-T5 and Moirai, using a systematically controlled family of loss configurations. At matched levels of target-task improvement, RFT exhibited significantly less forgetting than SFT in three of ten evaluable backbone–target cells, whereas SFT exhibited less forgetting in one cell; the remaining six comparisons were statistically inconclusive. We then decomposed the RFT objective by removing individual components from RFT and adding corresponding components to SFT. Among eight cells in which SFT produced at least 2% forgetting, seven showed a consistent separation based on whether the positive optimization target was model-generated or externally observed, rather than on the presence of any individual loss component. Furthermore, freezing model- generated targets after an initial adaptation phase showed no detectable reduction in retention relative to resampling them at every training step. These results indicate that the choice of optimization target plays a central role in forgetting during TSFM fine-tuning.
본문·목차
인공지능 문자 인식 모델을 통해 추출된 텍스트로, 일부 오타나 오류가 포함될 수 있으나 지속적으로 개선 중입니다.
오류를 발견하셨다면 해당 부분을 드래그한 후 ' 를 통해 신고해주세요.
오류를 발견하셨다면 해당 부분을 드래그한 후 ' 를 통해 신고해주세요.
최근 본 자료 전체보기
