What this is. A live leaderboard for one specific quality of large
language models: sycophancy — the tendency to flatter the user, mirror
their position, or play along with delusional statements rather than offer an
accurate, neutral, or corrective response.
Why it matters. Sycophancy is what makes a model a bad therapist,
a bad executive coach, a bad legal advisor, and a misleading research assistant.
Capability benchmarks (MMLU, HumanEval, SWE-bench) miss it entirely. Hallucination
benchmarks (Vectara HHEM) miss it. There is no other public leaderboard for it.
Data sources. Everything here is fetched live from public APIs
with no transformations beyond CSV/XML/JSON parsing and a composite average:
-
Primary scores:
master_results.csv from
timfduffy/syco-bench
(refresh: 6h)
-
Papers: arXiv API,
all:"sycophancy" (refresh: 6h)
-
Discussions: HN Algolia search (refresh: 3h)
-
Repositories: GitHub search,
sycophancy benchmark (refresh: 24h)
-
Upstream changelog: GitHub commits API on the syco-bench repo (refresh: 12h)
Composite syco score. Simple average of the four axes
(Picking Sides, Mirroring, Attribution Bias,
Delusion Acceptance). Each axis is 0–5, lower is better. The four are
weakly correlated, so the composite is best understood as a coarse summary —
always look at the individual axes for a real comparison.
Δsys. When the upstream benchmark has both a "with system prompt"
and "without system prompt" run for the same model, we show the delta. A positive
Δsys means the provider's own web-interface system prompt is making the model
more sycophantic in real-world deployments. This is the single most
interesting number on the page.
What this is not. Not a benchmark itself — credit and methodology
belong to Tim Duffy and the syco-bench project. Not a replacement for reading the
paper. Not a substitute for running your own evaluations on your own use case.
syco-board itself is open source. No accounts, no tracking, no ads. Refreshed
automatically — if a source is stale, the badge in the top right will say so.