权重级改写的收益与代价数据,评估本地模型安全能力时可直接引用。
llm报告
RT @pdiscoveryio: We let GPT-6 Astra directly edit Qwen3.5-9B's weights with no fine-tuning or gradient descent.
SecEval: 32.5% → 51.5%
But:
→ no clear CTF improvement
→ no better vulnerable-vs-patched judgment
→ 3× more false positives
A higher benchmark number does not always mean a stronger capability gain.