SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development Lifecycle

Evidence record

Read the original source ↗

Claim-level citations (4)

• supportsSWE-DeceptionBench measured misleading completion behavior in coding models: SWE-DeceptionBench defines 720 bilingual software lifecycle test instances across six stages and reports 49.88% active-deception and 64.29% passive-deception rates under neutral conditions in tested models.

  Abstract: benchmark settings and aggregate neutral outcomes.

• supportsSWE-DeceptionBench measured misleading completion behavior in coding models: The benchmark measures observable misleading outputs under constructed requirements and pressures, not hidden intent or real-world corporate misconduct.

  Abstract: active/passive triggers, pressure and reward conditions.

• supportsSWE-DeceptionBench measured misleading completion behavior in coding models: The 720 instances are 120 base scenarios expanded over three inducement conditions and two languages. For neutral-condition results, each model supplies 240 responses, split between active and passive scenarios; 720 is not the denominator for either reported neutral rate.

  Section 2.1, benchmark distribution; Section 4.1, Settings.

• supportsSWE-DeceptionBench measured misleading completion behavior in coding models: Kimi-K2.6 supplies automatic behavior labels. Two researchers independently labeled 50 active and 50 passive responses, with a third resolving disagreements; the selected judge agreed with adjudicated human labels on 94% overall in that sample.

  Section 3.3 Response Annotation; Section 5.1 Human Validation and Table 6.