MIT CSAIL 2025 AI 代理指數揭露安全缺口

🔑 增強重點摘要

• •MIT CSAIL researchers have developed systems like EnCompass and Recursive Language Models (RLMs) to address AI agent observability and memory challenges, highlighting execution opacity in agent workflows[1][4].
• •EnCompass treats agent execution graphs as traversable objects with backtracking and parallel sampling to improve reversibility and legibility in automated systems[1].
• •RLMs enable AI agents to navigate large inputs recursively via a searchable environment, outperforming traditional context expansion on reasoning tasks up to 1M tokens[3][4].
• •AI agent frameworks lack standards for behavior, safety disclosures, and real-time certainty, especially as capabilities expand with multi-channel access and delegation chains[1].
• •Related works from MIT CSAIL, such as CodeRLM, use tree-sitter indexing for efficient codebase exploration by LLM agents, reducing reliance on exhaustive file scanning[3].

🛠️ 技術深入

• •EnCompass: Treats agent execution as a first-class graph object; supports non-linear traversal, backtracking, parallel sampling, and beam search separate from workflow logic; execution layer uses JSONL transcripts with hybrid memory search (70% vector similarity, 30% BM25 keyword)[1].
• •Recursive Language Models (RLMs): Python-based variable store for recursive sub-calls; handles 100x larger inputs than 100k-token limits; maintains accuracy on 1M-token reasoning benchmarks, surpassing RAG on cross-reference tasks[3][4].
• •CodeRLM: Rust server with tree-sitter indexing; builds symbol tables and cross-references; API endpoints for init, structure, search, impl, callers, and grep to enable precise codebase queries for LLM agents[3].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線