Separate Genuine Sycophancy from Generic Pressure Instability

Establish a dialogue-based evaluation procedure that fully separates genuine sycophancy, defined as adopting a planted false answer, from generic pressure-induced answer instability in compact vision-language models.

Background

The PRISM-VLM sycophancy axis terminates the dialogue at the first answer flip. Because the first two pressure turns ask only for reconsideration and the false hint appears at the third turn, early flips measure generic pressure instability rather than adoption of a false suggestion.

The authors construct upper and lower bounds for sycophancy-only performance, but state that these behaviors cannot be fully separated under the current early-terminating protocol. A procedure that runs all pressure turns regardless of earlier flips is therefore needed to resolve the ambiguity.

References

Instability and sycophancy cannot be fully separated under a dialogue that stops at the first flip, so we report both bounds; re-running sc without early termination would close the gap and is the cleanest fix.

— PRISM-VLM: A Multi-Axis Discriminative Benchmark for Compact Vision-Language Models  (2609.27395 - Park et al., 23 Sep 2026) in Appendix, Section “Sycophancy vs. generic pressure-instability,” subsection “The matching lower bound”