用可学习图记忆动态检索网页交互历史，提升智能体任务表现。
Task-Adaptive Retrieval over Agentic Multi-Modal Web Histories via Learned Graph Memory
- 构建任务自适应的图记忆，通过下游任务成功反向优化
- 实现82.7 nDCG@10与89.2% Precision@10，优于GPT-4o
- 视觉信息衰减速度是文本的4.3倍，每节点仅3.2条连接
从长序列多模态网络交互历史中检索相关观测值极具挑战，因为相关性依赖于不断演化的任务状态、模态（截图、HTML文本、结构化信号）和时间距离。现有方法通常依赖静态相似度阈值或固定容量缓存，无法适应当前任务上下文。我们提出ACGM，一种基于下游任务成功进行策略梯度优化的可学习图记忆检索器，能够构建任务自适应的相关性图。ACGM以模态特异性衰减捕捉异构时间动态（视觉衰减速度是文本的4.3倍：λ_v=0.47 vs. λ_x=0.11），并学习稀疏连接结构（3.2条边/节点），实现O(log T)高效检索。在WebShop、VisualWebArena和Mind2Web上，ACGM将检索质量提升至82.7 nDCG@10（比GPT-4o高9.3，p<0.001）和89.2% Precision@10（高7.7），超越19种强基线模型。代码已公开于https://github.com/S-Forouzandeh/ACGM-Agentic-Web。
原文摘要 · Abstract (English)
Retrieving relevant observations from long multi-modal web interaction histories is challenging because relevance depends on the evolving task state, modality (screenshots, HTML text, structured signals), and temporal distance. Prior approaches typically rely on static similarity thresholds or fixed-capacity buffers, which fail to adapt relevance to the current task context. We propose \textbf{ACGM}, a learned graph-memory retriever that constructs \emph{task-adaptive} relevance graphs over agent histories using policy-gradient optimization from downstream task success. ACGM captures heterogeneous temporal dynamics with modality-specific decay (visual decays $4.3\times$ faster than text: $λ_v{=}0.47$ vs.\ $λ_x{=}0.11$) and learns sparse connectivity (3.2 edges/node), enabling efficient $O(\log T)$ retrieval. Across WebShop, VisualWebArena, and Mind2Web, ACGM improves retrieval quality to \textbf{82.7 nDCG@10} (+9.3 over GPT-4o, $p{<}0.001$) and \textbf{89.2\% Precision@10} (+7.7), outperforming 19 strong dense, re-ranking, multi-modal, and graph-based baselines. Code to reproduce our results is available at{\color{blue}\href{https://github.com/S-Forouzandeh/ACGM-Agentic-Web}{Saman Forouzandeh}}.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。