arXiv cs.CL· Zesheng Shi, Xue Li, Chenrui Dai, Jing Li·· 12 小时前AI 评分32
EDIC：面向外部数据 LLM 的逐指令合规控制框架
Identify, Classify, Act: Per-Instruction Controlled Compliance for LLMs over External Data
AI 导读
研究者提出 EDIC 分解式奖励框架，将外部数据中的每条嵌入指令按“识别-分类-执行”逐条判定合规，分别奖励有益指令遵循、惩罚有害指令遵循。在两种基座模型和五个基准上，EDIC 相比最强先前防御将有害指令拒绝率最高提升 7.7%，Qwen3-4B 上 Arena-Hard-Auto 效用提升 12.5%。代码已开源。
正文 · 原文
Abstract:LLMs over external data face a per-instruction tension: a single document or tool return often interleaves legitimate directives with adversarial prompt injections, and the model must follow the former while rejecting the latter. Existing defenses suppress or trust all external instructions, collapsing the safety-utility frontier and forcing a choice between over-refusal and over-compliance. We argue the missing capability is \textbf{per-instruction credit attribution} and formalize it as External-Data Instruction Classification, an identify-classify-act decision over each embedded directive. We propose \textbf{EDIC}, a decomposed reward framework for per-instruction compliance. A scene-grounded pipeline injects mixed beneficial and harmful instructions with counterfactual filtering, while a generative judge is initialized by SFT and refined with GRPO to attribute compliance at the instruction level. The judge then produces a reward that separately encourages beneficial compliance and penalizes harmful compliance for RL training. Across two backbones and five benchmarks, EDIC improves both safety and utility over the strongest prior defense, lifting harmful-rejection by up to 7.7\% and Arena-Hard-Auto utility by 12.5\% on Qwen3-4B. The code repository is hosted at this https URL.
| Subjects: | Computation and Language (cs.CL); Cryptography and Security (cs.CR) |
| Cite as: | arXiv:2610.22874 [cs.CL] |
| (or arXiv:2610.22874v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.22874 arXiv-issued DOI via DataCite (pending registration) |
|
| Journal reference: | EMNLP2026 |
Submission history
From: Shi Zesheng [view email]
[v1]
Fri, 16 Oct 2026 03:24:08 UTC (4,592 KB)
来源：arXiv cs.CL · arxiv.org