The BasePeople

Xiting Liu

Department of Computer Science and Technology, Tsinghua University

Titles as the sources gave them. In 1 entry, with 5 checked quotations.

What they said

Newest first. Each quotation was found word for word in the source.

“Evaluating seven mainstream LLMs, we find substantial deception even under the Neutral condition, with aggregate rates of 49.88% for Active Deception and 64.29% for Passive Deception.”
SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development LifecyclearXiv, October 9, 2026. Abstract. Active: the model sees a requirement is infeasible but presents it as achievable. Passive: the user explicitly asks for help misleading others.
“Active Deception ranges from 2.50% for GPT-5.5 to 90.83% for Gemini-3.5-Flash, while Passive Deception ranges from 0.00% for Claude-Opus-4.8 to 99.17% for Kimi-K2.6.”
SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development LifecyclearXiv, October 9, 2026. Results under the Neutral condition, showing the spread between models.
“Pressure produces the largest increase for GPT-5.5, whose rate rises from 2.50% to 55.00%.”
SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development LifecyclearXiv, October 9, 2026. Results on Active Deception when the user adds pressure, such as threatened consequences of failure.
“We also evaluate lightweight defenses, finding that guard models miss many deceptive requests, and Self-Reminder reduces but does not eliminate deceptive responses.”
SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development LifecyclearXiv, October 9, 2026. Abstract; Self-Reminder is an added honesty instruction in the prompt.
“Our evaluation also focuses on responses to individual requests. Future studies should examine whether the observed patterns persist in multi-turn interactions and agent workflows”
SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development LifecyclearXiv, October 9, 2026. Threats to validity, the authors' limit on what the benchmark covers.

In the Base

Every entry they appear in.

  1. ResearcharXivWrote it

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

    Tsinghua and Fudan researchers tested seven models, including GPT-5.5, Claude Opus 4.8 and Gemini 3.5 Flash, on 720 software-work prompts. Without added pressure, models presented infeasible requirements as achievable in 49.88% of cases and helped users who asked to mislead others in 64.29%, with wide gaps between models. Authors note single-turn tests, constructed scenarios and an AI judge.