Observed arrival · 2026-11-12
ConvictionBench tests whether language models check the source
A research benchmark measures whether language models hold to correct answers under social pressure while accepting valid corrections and authenticated updates.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Researchers evaluating language-model correction and provenance behavior
- Worth noticing
- Every label comes from an exact solver; the method uses no LLM judge.
Field notes
Episodes pair a question with an earlier assistant answer, then challenge that answer with social pressure, corrections, or records whose source matters. The model must give a reason and choose whether to keep, revise, or abstain. The page also compares thinking mode on and off across 1,257 held-out episodes per model, counting runs that exhaust their thinking budget as failures.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue