The authors compared two paradigms for LLM abstention: selective answering, which uses a confidence threshold outside the model, and abstention training, which internalizes the decision.
A modular abstention method separates accuracy training from abstention decisions
The authors show that the advantage of end-to-end abstention training comes from better answering, not better abstention, motivating their two-stage TAPS approach.
Research Lab
Junwei Zhang · Yifei Wang · Wanli Yang · Hexiang Tan · Rui Tang · Xin Chen · +3 more
University of Chinese Academy of Sciences · State Key Laboratory of AI Safety, Institute of Computing Technology, CAS · Beijing Institute of Technology
Research Digest··2 min read
The authors decompose the risk of LLM abstention at matched coverage into answering ability and distinguishing ability, finding that abstention training primarily improves the former.
Why this paper
From University of Chinese Academy of Sciences and 2 others
In one line
Training for accuracy then using a separate confidence estimator for abstention achieves strong and controllable abstention performance.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§