SetupAI
EN
來源較早收集於 5h

針對人類與大型語言模型集合的對抗性社會知識論

針對人類與大型語言模型集合的對抗性社會知識論
PostLinkedIn
📄閱讀原文： ArXiv AI
#trust-and-safety#epistemology#llm-governanceadversarial-social-epistemology-(ase)arxiv

💡學習如何檢測並防止 AI 與人類溝通網路中對信任的策略性操縱。

⚡ 30 秒速覽

有什麼變化

引入對抗性社會知識論（ASE）以分析 LLM 輔助溝通中的信任剝削。

為什麼重要

該框架為開發 AI 整合資訊系統的開發者提供了一個關鍵視角，有助於減輕錯誤資訊並維持系統的可靠性。

下一步行動

在您的應用程式中為 LLM 生成的輸出加入自動化審計軌跡，以追蹤聲明的推論鏈。

誰應關注：Researchers & Academics

關鍵要點

  • •引入對抗性社會知識論（ASE）以分析 LLM 輔助溝通中的信任剝削。
  • •識別溝通主體如何扭曲或捏造資訊以破壞制度性認證。
  • •提出審計推論鏈的機制，以確保公開斷言的完整性。
  • •利用推論主義語義學來解釋並驗證 AI 輔助聲明的有效性。

🧠 深度解析

本篇為 AI 生成分析，非原文內容。

🔑 增強重點摘要

  • •The framework draws heavily on Robert Brandom’s inferentialism, treating LLM outputs as 'commitments' within a social game of giving and asking for reasons.
  • •ASE specifically addresses the 'epistemic free-riding' problem, where agents use LLMs to generate high-volume, low-effort content that mimics institutional authority.
  • •The research introduces a formal 'proof-of-provenance' protocol for inferential chains, requiring LLMs to cryptographically link claims to verifiable source datasets.
  • •It identifies 'semantic drift' as a primary vulnerability, where LLMs subtly alter the inferential role of terms during multi-step reasoning to bypass safety filters.
  • •The proposed auditing machinery utilizes 'adversarial verification,' where a secondary, specialized LLM acts as a dialectical opponent to stress-test the primary agent's inferential consistency.

🛠️ 技術深入

  • Implementation utilizes a Directed Acyclic Graph (DAG) structure to map inferential dependencies across multi-agent interactions.
  • Employs 'Inferential Traceability Tokens' (ITTs) to maintain a verifiable log of how specific premises lead to a final assertion.
  • Integrates with existing Knowledge Graph (KG) architectures to cross-reference LLM-generated inferences against ground-truth ontologies.
  • Utilizes a Bayesian belief-updating mechanism to quantify the 'trust score' of an agent based on its historical adherence to inferential validity.

🔮 前景展望基於引用來源的 AI 分析

Mandatory provenance metadata will become a standard for AI-generated public policy documents by 2027.
The increasing risk of institutional trust erosion necessitates verifiable inferential chains for high-stakes decision-making.
Adversarial verification will replace static safety fine-tuning as the primary method for LLM alignment.
Static fine-tuning fails to account for dynamic, context-dependent subversion, whereas adversarial auditing adapts to the agent's reasoning process.

⏳ 時間線

2025-03
Initial conceptualization of inferentialist semantics applied to LLM-human hybrid systems.
2025-11
Development of the first prototype for auditing inferential chains in closed-loop environments.
2026-06
Publication of the Adversarial Social Epistemology framework on ArXiv.
📄閱讀原文： ArXiv AI
📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #trust-and-safety

同產品

更多關於 adversarial-social-epistemology-(ase)

同來源

最新來自 ArXiv AI

AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： ArXiv AI ↗

這是摘要，不是原文。去看原站，或訂閱每週簡報。

每週電子報

每週一封，可隨時退訂。