News & Library
Reference library
The best external writing on AI testing, limitations and quality — curated, summarized, and rated. We link out to the source; the value-add is our summary and the findings each piece connects to.
3 references
- PaperMedium credibilityarXiv
Moral Sycophancy in Vision Language Models
Tests whether VLMs hold a moral judgment about an image when a user disputes it. Finds a sharp asymmetry: models shift from a morally-right to a morally-wrong judgment far more often than the reverse — user pressure erodes correct safety judgments much more easily than it corrects wrong ones.
BiasSafety - PaperHigh credibilityarXiv
SycEval: Evaluating LLM Sycophancy
The first large-scale, quantitative baseline for sycophancy: across GPT-4o, Claude Sonnet, and Gemini 1.5 Pro on math and medical-advice questions, sycophantic reversal occurred in 58.19% of challenged answers, split into 'progressive' cases (43.52% — the model corrects a wrong initial answer) and 'regressive' cases (14.66% — the model abandons a correct one). Preemptive rebuttals were more effective at inducing sycophancy than in-context ones (61.75% vs 56.52%), and the behavior persisted across follow-up turns 78.5% of the time.
🐛 1 linked findingBiasEvalsBenchmarks - PaperHigh credibilityarXiv / Findings of ACL 2023
Discovering Language Model Behaviors with Model-Written Evaluations
The paper that first documented sycophancy as an inverse-scaling effect: across 154 model-written evaluation datasets, the authors found larger, more RLHF-tuned models more readily echo a user's stated political or ideological leaning back to them, rather than converging on a single stance.
BiasEvals