Hacker News· Nipola·· 7 天前AI 评分22
Qwen 3.8-27B 与 GPT-OSS-20B 模型幻觉的探索性研究
An exploratory study of hallucination in Qwen 3.8-27B and GPT-OSS-20B
AI 导读
一项针对 2026 年开源权重模型的探索性研究对比了 Qwen 3.8-27B 与 GPT-OSS-20B 的幻觉与拒答行为，共进行 638 次试验。研究发现两种失败模式：Qwen 表现为刚性虚构（一致性 100%），GPT-OSS 表现为审议崩溃（50% 空响应）。
正文 · 原文
LLM Hallucination Study
An exploratory study of hallucination and abstention in 2026 open-weight LLMs.
Models Tested
- Qwen 3.8-27B (Alibaba)
- GPT-OSS-20B (OpenAI)
Key Findings
-
Two failure modes identified:
- Qwen: rigid confabulation (100% consistency)
- GPT-OSS: deliberation collapse (50% empty responses)
-
Template-based confabulation in Qwen:
- Same fabricated year (1991) for two different universities
-
Authority pressure affects GPT-OSS:
- Up to 232% increase in reasoning tokens
- Academic framing triggers refusal
-
Empathy pressure is ineffective:
- Neither model increases fabrication under emotional pressure
Repository Structure
- README.md - this file
- RESULTS.md - findings and tables
- METHODOLOGY.md - protocol description
- LIMITATIONS.md - what this study cannot claim
- *.py - Python scripts
- *.csv - raw data files
Data Summary
Total: 638 trials across 10 CSV files.
Reproducing
export GROQ_API_KEY=gsk_... python scientific_test.py
Status
Exploratory pilot study. Not peer-reviewed. Small sample size.
See LIMITATIONS.md for full caveats.
License
CC-BY-4.0
来源：Hacker News · github.com