arXiv:2606.29823 · 中英对照阅读
经验图:自我改进智能体的数据基础设施
Experience Graphs: The Data Foundation for Self-Improving Agents
中文速览
长周期智能体做科研、写代码或优化硬件时,需要反复试错上百步,但这些过程常被当作临时日志,崩溃后难恢复,也难让不同任务共享经验。论文提出 Trellis11,把每次尝试、产物、工具结果、得分和前后因果关系组成的“经验图”放进可查询、可管理的数据库,让搜索、找旧方案、恢复现场和整理训练数据都变成数据库操作。基于 Meta 的 KernelEvolve 实测,跨会话复用经验后,达到目标加速约快 10 倍,每个有效方案的 token 成本低 52%,同时也发现复用过多可能让搜索变得保守。它的重要性在于把智能体的试错从用完即弃的计算变成可累积、可共享、能持续改进的长期资产。
摘要
数据库领域一再通过认识到新型工作负载需要新的系统架构,推动技术发展至新的水平。我们认为,长时程智能体任务——代码生成、科学发现、硬件设计、安全研究——正是这样一种工作负载。这些智能体并非只生成单一答案,而是会进行探索:在数百个步骤中生成产物、执行工具、观察失败、分支探索、修复问题并比较备选方案,同时逐步超越基线。这样的搜索会生成一种丰富的结构化对象,我们称之为经验图:其中包含可执行产物、工具输出、目标奖励、同级比较、可变搜索统计数据和因果谱系。然而,现有智能体框架将这类经验视为可丢弃的状态——将其序列化为 JSON 检查点和会话日志,而这些数据在故障后无法恢复,无法跨用户查询,无法按相似性搜索,也无法在不依赖脆弱的事后抓取的情况下物化为训练数据。即使是生产环境中的智能体已趋于采用的基于文件的记忆——声明性事实、程序性技能、情景日志——所记录的也是智能体知道什么,而非其搜索尝试过什么所构成的携带奖励信息的经验图。
The database community has repeatedly advanced the state of the art by recognizing that new workloads demand new system architectures. We argue that long-horizon agentic tasks—code generation, scientific discovery, hardware design, security research—are such a workload. Rather than producing a single answer, these agents explore: they generate artifacts, execute tools, observe failures, branch, repair, and compare alternatives over hundreds of steps, improving over a baseline as they go. This search produces a rich, structured object we call an experience graph: executable artifacts, tool outputs, objective rewards, sibling comparisons, mutable search statistics, and causal lineage. Yet existing agent frameworks treat this experience as disposable state—serialized into JSON checkpoints and session logs that cannot be recovered after a crash, queried across users, searched by similarity, or materialized into training data without brittle, after-the-fact scraping. Even the file-based memory that production agents have converged on—declarative facts, procedural skills, episodic logs—captures what an agent knows, not the reward-bearing experience graph of what its search tried.
术语表
- long-horizon agentic tasks
- 长时程智能体任务
- agentic system
- 智能体系统
- experience graph
- 经验图
- Trellis
- Trellis
- data foundation
- 数据基础设施
- database access pattern
- 数据库访问模式
- recursive self-improvement (RSI)
- 递归自我改进(RSI)
- inference-time search
- 推理时搜索
- Monte Carlo Tree Search (MCTS)
- 蒙特卡洛树搜索(MCTS)
- evolutionary search
- 进化搜索
- frontier selection
- 前沿选择
- cross-session reuse
- 跨会话复用
- vector-seeded graph retrieval
- 向量引导的图检索
- materialized view
- 物化视图
- time-travel query
- 时间旅行查询
- as-of temporal query
- 时点查询
- stateless compute
- 无状态计算
- serverless compute
- 无服务器计算
- crash recovery
- 故障恢复
- horizontal scaling
- 水平扩展
- closed-loop training flywheel
- 闭环训练飞轮
- declarative memory
- 声明式记忆
- procedural memory
- 程序性记忆
- episodic memory
- 情景记忆
- multi-modal query planning
- 多模态查询规划
- concurrent tree search
- 并发树搜索
- governed view maintenance
- 受治理的视图维护
- bi-temporal memory
- 双时态记忆
- exploration–anchoring tradeoff
- 探索—锚定权衡