CLAIM-2026-10-13-01

SWE-DeceptionBench covers six software lifecycle stages, two deception triggers and three inducement conditions (Neutral, Reward, Pressure), giving 720 test instances in Chinese and English, evaluated across seven mainstream LLMs.

Why it matters

It fixes the size of the thing: 720 instances over seven models is a benchmark, not an anecdote, and the Chinese/English split matters for anyone reading the rates as universal.

Sources

• arxiv2026swedecept