[论文解读] Persona-Driven Benchmarking for Generalizable and Human-Aware Artificial General Intelligence

Shunyu Yao, Jeffrey Zhao+5 more|arXiv (Cornell University)|Oct 6, 2022

Topic Modeling被引用 518

一句话总结

本论文提出 ㊀模型（(symbol)），一种将推理痕迹与行动在大型语言模型中交错的基于提示的方法，旨在提升问答、事实验证和互动决策任务的性能与可解释性，超越基线并实现更好的人工对齐。

ABSTRACT

This research paper, "Persona-Driven Benchmarking for Generalizable and Human-Aware Artificial General Intelligence," proposes a novel architectural solution to transition current Large Language Models (LLMs) from sophisticated pattern-matchers to genuine Artificial General Intelligence (AGI) agents. The core argument of the paper is that achieving AGI is primarily an architectural challenge (creating a "gearbox" of metacognitive functions) rather than solely a scaling or fundamental modeling problem (the "engine"). The proposed framework is an architectural overlay composed of three synergistic components that address the key limitations of contemporary LLMs: lack of persistent memory, poor accountability, and narrow human-centric reasoning. Key Components of the Proposed Architecture Persona Module (for Dynamic Contextualization and Self-Awareness): Function: Handles the dynamic contextualization of all interactions, performs causal intent inference regarding the user's needs, and manages the AGI's evolving, internal self-identity (or "self-model"). Goal: To enable the AGI to move beyond simple response generation to goal-oriented optimization based on a deep, human-aware understanding of the user and its own capabilities. Benchmarking Loop (for Metacognitive Self-Correction): Function: This component acts as the AGI's metacognitive self-correction engine. It rigorously audits the LLM's output against a goal-specific metric, performs causal inference to determine why a failure occurred, and then dictates a permanent learning path. Goal: To foster an internal, continuous learning cycle that ensures the AGI is accountable and capable of generalized skill transfer from one task to a similar, new one. FedRAG (Federated Reflective Augmented Generation): Function: A novel, persistent external memory layer that functions as a dynamic knowledge graph. It is crucial because it decouples learned skills and state from the static LLM weights. It stores successful outcomes, critical causal inferences from the Benchmarking Loop, user preferences, and Persona state updates. Goal: To provide the AGI with persistent, generalizable memory and ensure that learning is permanent, efficient, and does not lead to catastrophic forgetting or excessive model scaling. Central Thesis The paper asserts that the integrated, recursive loop formed by these three components—where the Benchmarking Loop's reflective output is written to the FedRAG memory for use by the Persona Module in future interactions—is the necessary architectural scaffold to achieve true, generalizable, and human-aware AGI.

研究动机与目标

• 推动将推理与行动紧密耦合，以提升LLM驱动任务的鲁棒性和可解释性。
• 展示推理痕迹与行动交错可以在多样基准上超越仅推理或仅行动的基线。
• 通过可见的推理轨迹提升可解释性、可信度和诊断性。
• 评估外部知识检索与内部推理在知识密集型任务中的互补作用。

提出的方法

• 扩展智能体的行动空间，使其包含用于更新未来步骤上下文的语言推理痕迹。
• 用少样本轨迹提示冻结的 PaLM-540B（以及附录中的 GPT-3），包括想法、行动和观测。
• 使用简单的外部知识API（维基百科）并辅以诸如搜索、查找和完成等行动，将推理落地到事实。
• 在多项任务中比较 ㊀模型 与 Standard、CoT、CoT-SC 与 Act 基线。
• 通过在任务反馈基础上在 ㊀模型 与 CoT-SC 之间切换，提出混合提示以同时利用内部与外部知识。
• 在 HotpotQA 上通过对 ㊀模型轨迹（思想、行动、观测）的几千条微调来改进解码能力。

实验结果

研究问题

• RQ1 跨越推理痕迹与行动的交错是否能在性能与可解释性上超越仅推理或仅行动的基线？
• RQ2 与外部知识库（如维基百科）的交互是否提升扎根推理能力并减少幻觉？
• RQ3 将内部推理与外部知识结合对问答和事实核验任务有何影响？
• RQ4 与仿真学习和强化学习基线相比， ㊀模型在互动决策基准上的表现如何？
• RQ5 微调是否增强模型解码轨迹并对新任务的泛化能力？

主要发现

• ㊀模型在 HotpotQA 和 Fever 上超越了仅行动基线，在某些 setting 中将 ㊀模型 与 CoT-SC 结合可获得最佳提示效果。
• ㊀模型通过访问简单的维基百科API获取相关事实，表现出更少的幻觉并更好地扎根。
• 在 ALFWorld 和 WebShop 上， ㊀模型在较少的上下文示例条件下对比更强的基线取得了显著的绝对提升（分别为34%和10%）。
• ㊀模型提供可解释的轨迹（思想、行动、观测），供人类检查以诊断和信任决策。
• 单独进行提示的提升有限，但使用几千条轨迹微调使 ㊀模型成为在 HotpotQA 上评估选项中的最佳方法。
• ㊀模型在跨越多领域时通常提供一个稳健且具泛化能力的框架，与强化学习及多任务训练结合时有进一步提升的潜力。