finding
active
finding:general-sycophancy-is-already-high-in-sft-and-only-slightly-modulated-by-dpo-reward-shaping-does-not-reduce-it-below-sft-levelsGeneral sycophancy is already high in SFT and only slightly modulated by DPO; reward shaping does not reduce it below SFT levels
Fig. 12.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Sidharth Baskaran · Max Loeffler +13
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- General sycophancy case.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Physics sycophancy user-praise rises from 1.32 (SFT) to 2.18 (DPO); reward shaping α=4 reaches 1.80finding0.836Table 11.
- Shows SFT effect is trait-specific and reflects register of demonstrations
- Fig. 10, Table 10.
- Author's interpretation from SAE decomposition of the sycophancy persona vector
- Benchmark identification rate rises from 0.08 (SFT) to 0.25 (DPO); interventions barely reduce itfinding0.764Table 11.
- Fig. 3 on Llama-3.1-8B.
- Fig. 8.
- Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models (Denison et al. 2024)concept0.753Related work on LLMs generalizing to reward hacking; methodology used for RL experiments