SRC04 — From Yes-Men to Truth-Tellers: Addressing Sycophancy with Pinpoint Tuning¶
Source¶
| Title | From Yes-Men to Truth-Tellers: Addressing Sycophancy in Large Language Models with Pinpoint Tuning |
| Publisher | ICML 2024 / arXiv |
| Authors | Wei Chen, Zhen Huang, Liang Xie, et al. |
| Date | September 2024 (accepted ICML 2024; revised February 2025) |
| URL | https://arxiv.org/abs/2409.01658 |
| Type | Peer-reviewed conference paper |
Summary Ratings¶
| Dimension | Rating |
|---|---|
| Reliability | High |
| Relevance | High |
| Missing data bias | Low |
| Measurement bias | Low |
| Selective reporting bias | Low |
| Randomization bias | N/A |
| Protocol deviation bias | Low |
| COI / Funding bias | Low |
Rationale¶
| Dimension | Rationale |
|---|---|
| Reliability | Peer-reviewed at ICML 2024; quantified results with multiple model sizes |
| Relevance | Directly proposes a method to reduce sycophancy by targeting specific model components |
Evidence Extracts¶
| Evidence | Summary |
|---|---|
| SRC04-E01 | Supervised Pinpoint Tuning reduces sycophancy by targeting <5% of model modules |