R0040/2026-03-28/Q002/SRC06
Wei et al. paper on using synthetic data to reduce sycophancy.
Source
| Field |
Value |
| Title |
Simple Synthetic Data Reduces Sycophancy in Large Language Models |
| Publisher |
arXiv |
| Author(s) |
Jerry Wei et al. |
| Date |
2024-02-16 |
| URL |
https://arxiv.org/abs/2308.03958 |
| Type |
Research paper |
Summary
| Dimension |
Rating |
| Reliability |
Medium-High |
| Relevance |
High |
| Bias: Missing data |
Low risk |
| Bias: Measurement |
Low risk |
| Bias: Selective reporting |
Low risk |
| Bias: Randomization |
N/A |
| Bias: Protocol deviation |
N/A |
| Bias: COI/Funding |
Low risk |
Rationale
| Dimension |
Rationale |
| Reliability |
From a credible research team. Pre-print with solid experimental methodology. |
| Relevance |
Directly demonstrates that sycophancy can be reduced through data-level intervention without changing the training algorithm. |
| Bias flags |
No significant concerns. |
| Evidence ID |
Summary |
| SRC06-E01 |
Synthetic non-sycophantic training data significantly reduces sycophancy |