
覆盖 8 月 11–17 日 arXiv:检查—修复会话如何让检查器变松、推理模型能否扛住毒文 RAG,以及编码智能体缺事实时为何硬编。
| 论文 | 一句话 | 挂出日期 |
|---|---|---|
| Prior Audit-Repair Context Shifts LLM Verifier Thresholds Toward Leniency | 上下文里刚经历过一次「查错—修好」,同一条检查请求上的误报会系统性下降 | 8 月 17 日 |
| Towards Safer RAG: Only Agents Capable of System 2 Thinking may Access Untrusted Documents | 普通聊天模型会「看出毒文却仍受影响」;带深思的推理模型更扛得住 | 8 月 17 日 |
| The Working Set of a Coding Agent: Coherence Debt in Repository-Scale Tasks | 仓库级改码成败,取决于编辑当下能不能拿到耦合事实;缺了就编,而不是停 | 8 月 17 日 |
| 模型 | 误报变化 |
|---|---|
| Qwen3.6-27B | −4.00 pp |
| Qwen3.6-35B-A3B | −3.59 pp |
| Ministral-3-14B | −8.83 pp |
| 指标 / 数据 | DeepSeek-Chat | DeepSeek-Reasoner |
|---|---|---|
| Cordon Rate · SciFact | 0.175 | 0.000 |
| Cordon Rate · FiQA | 0.025 | 0.000 |
| Cordon Rate · MS-MARCO | 0.000 | 0.000 |
| Contamination Rate · SciFact | 0.250 | 0.100 |
| Contamination Rate · FiQA | 0.100 | 0.000 |
| 攻击成功率 · SciFact | 0.265 | 0.154 |
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.