---
title: "PAST-Bench：评估个人智能体递归自我改进能力的基准"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.04003v1"
permalink: "https://feed.astralor.com/items/itm_6a95ac2c4a8ad5cc00b872a8"
published: "2026-08-04T17:58:05.000Z"
score: 77
channel: "ai"
tags: ["Hermes+","PAST-Bench","个人智能体","基准测试","递归自我改进"]
---

# PAST-Bench：评估个人智能体递归自我改进能力的基准

## AI 点评

> 这篇论文最值得关注的是它没有停留在“有没有提升”的统计数字上，而是用路径证据拆穿了“假改进”的可能。对做Agent产品的人来说，Hermes+的五个干预点非常具体，尤其是替换过时状态那一步，几乎直击日常开发中记忆失效的痛点。PAST-Bench 后续如果被社区广泛采用，很可能成为衡量个人智能体学习能力的标准试金石。

## 速览

递归自我改进要求智能体将积累的经验转化为更好的未来行为。个人AI智能体因跨会话保留偏好、任务历史、工具流程和习得技能，成为研究该能力的理想场景，但经验留存是否真正带来改进尚未被系统测试。PAST-Bench 基准通过26个场景、204个回合，在开/关经验保留的对照条件下，测试智能体在记忆、程序复用、信息收集和更新四类能力上的表现。研究覆盖7个基础模型和4个智能体框架，发现改进真实存在但不均衡。基于发现提出的 Hermes+ 通过五个定向干预提升了平均收益，并在替换过时状态的任务上效果最显著。

## 关键点

- PAST-Bench 设有26个场景、204个回合，采用相同条件下开/关经验保留的对照设计，隔离测试递归自我改进。
- 覆盖记忆、程序复用、信息收集和更新四类能力，评估7个基础模型和4个智能体框架。
- 发现智能体提升真实存在但不均衡，且相同涨点背后可能缺乏预期路径证据。
- 基于研究发现提出 Hermes+，通过五个定向干预平均提升经验收益，在替换过时状态任务上效果最强。

## 全文

Recursive self-improvement requires agents to turn accumulated experience into better future behavior. Personal AI agents offer a concrete setting for studying this capability because they retain preferences, task histories, tool routines, and learned skills across sessions. Yet whether retained experience actually improves them over time has not been systematically tested. We introduce PAST-Bench, a benchmark designed to isolate this question. Each agent runs through ordered sequences of fresh-session tasks under matched conditions that turn retained experience on and off. It spans 26 scenarios and 204 episodes across memory, procedural reuse, information gathering, and update. We report both later-task gains and whether those gains follow the intended save, retrieve, and update pathway. Across seven base models and four agent frameworks, improvement is real but uneven across capabilities. Agents with the same headline gain can differ markedly in whether that gain is supported by evidence of the intended pathway. Guided by these findings, we develop Hermes+, which extends Hermes with five targeted interventions across stages of the agent loop. Hermes+ raises the average gain from retained experience and provides clearer pathway evidence, with its strongest improvement on tasks requiring outdated state to be replaced, although the effect remains capability- and model-dependent. Together, PAST-Bench and Hermes+ provide an evaluation and diagnostic foundation for studying how persistent agents can progress from retaining experience to systematically improving through it. Code: https://github.com/Gen-Verse/PAST-Bench

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_6a95ac2c4a8ad5cc00b872a8