---
title: "Fudan NLP × miHoYo: Reading the ~50k-Character “AI Agents” Survey"
author: "Unique Research"
sourcePublication: "Unique Research (FFCAP English library)"
originalPublishedAt: "2023-09-18T18:00:00+08:00"
canonical: "https://ffcap.cn/en/research/20230918a055l200"
source: "https://ffcap.cn/en/research/20230918a055l200"
language: "en"
cnSourceUrl: "https://view.inews.qq.com/a/20230918A055L200"
---

# Fudan NLP × miHoYo: Reading the ~50k-Character “AI Agents” Survey

_Unique Research · Insight · September 18, 2023 · Full editorial rewrite of the Fudan NLP Lab + miHoYo joint survey on LLM-based Agents — TOC §§1–7 covered end-to-end (CN source ~12.6k chars of survey digest; not a stub)_

Fudan University’s NLP Lab and miHoYo jointly released a long survey on **LLM-based Agents** (智能体). Unique Research walks the full outline so English readers get the map without losing the Chinese structure.

**Glossary:** LLM · Agent · AI Agents (no single standard Chinese term; 智能体 used) · LLM-based Agents · AaaS (Agent-as-a-Service).

## 1 Introduction

LLMs moved from chat interfaces to systems that plan, use tools, remember, and coordinate. The survey treats Agents as the next organizing abstraction: a brain (LLM) plus perception and action loops—alone, in swarms, or with humans—eventually forming simulated societies. The goal is a shared vocabulary for builders and researchers across academia and games/entertainment (miHoYo’s stake).

## 2 Background

### 2.1 Origins

Agency traces to Greek philosophy (Socrates/Plato), then Hume/Kant on subjectivity and agency—academic foundations for “agentic” thought. Turing (1950s) imported related ideas into CS via the intelligence test. Mainstream AI long chased narrower techniques; Agents became a core AI concept mainly from the 1980s. Unlike passive objects that only accept inputs, Agents stress autonomy, goal-directedness, initiative, and sociality—pushing AI toward stronger systems and remaining a premise for more capable intelligence.

### 2.2 Technical trends

Milestones: symbolic rule/logic systems (brittle under uncertainty) → statistical learning and neural connectionism (learn/adapt) → large neural LMs as Agent substrates with early successes. The field is still fast-moving; LLM-driven Agent design remains open research.

### 2.3 Why LLMs fit as the “brain”

Natural language I/O enables social interaction with humans and other Agents; broad world knowledge; in-context learning; emergent reasoning and planning; transfer across tasks without full retraining—making LLMs a practical central controller for Agents.

## 3 Birth of the Agent: building on LLMs

### 3.1 Brain

The brain is the hub; an LLM supplies language understanding, memory, and reasoning. Five capability upgrades:

-   **3.1.1 Natural language interaction** — multilingual I/O (EN/ZH…); optional NLG modules for richer, clearer expression.
-   **3.1.2 Knowledge** — parametric knowledge plus retrieval over external stores for factual grounding.
-   **3.1.3 Memory** — store prior dialogues as condensed frames; retrieve in later turns so long-horizon interaction does not reset.
-   **3.1.4 Reasoning & planning** — hypothetical reasoning over decision paths (prompt-guided); decompose hard goals into sub-tasks; revise plans from mid-execution feedback.
-   **3.1.5 Transfer & generalization** — in-context adaptation to new tasks without training from scratch; few-shot generalization for open, complex environments (historically a hard point; recent LLMs show strong transfer).

### 3.2 Perception

-   **Text** — still primary; challenge is implicit context + common-sense; also extract from semi-structured web pages.
-   **Images** — caption→text pipelines, or direct visual encoders (e.g. CNN) for richer representations.
-   **Audio / other sensors** — widen the world model beyond language.

### 3.3 Action

-   **Text output** — default channel.
-   **Tool use** — call external APIs/services instead of re-training; mastering tool interfaces is key to efficiency.
-   **Embodied action** — e.g. robot-arm control: 3D scene understanding, spatial reasoning, and mapping language instructions to low-level motion remain hard.

## 4 Agents in practice: “always leverage AI”

### 4.1 Single-Agent general capability

-   **Task-oriented** — finish defined workflows.
-   **Innovation-oriented** — explore novel solutions/creativity.
-   **Lifecycle-oriented** — continuous self-improvement and experience accumulation, not one-shot deploy.

### 4.2 Multi-Agent synergy

Figure-class pattern in the survey: cooperative interaction (unordered or ordered toward a shared goal) vs adversarial interaction (turn-taking competition to raise each Agent’s performance). Complementary coordination assigns roles so each Agent plays to its strengths. Adversarial setups should surface new viewpoints—not anxious rivalry—with mechanisms against over-aggression.

### 4.3 Human–Agent engagement

-   **Coach–executor** — humans instruct/feedback; Agents execute. Coaches should not over-constrain: grant autonomy to explore alternate solutions; allow low-cost trial-and-error; balance guidance vs freedom of action space.
-   **Equal partner** — Agents excel at scale compute/data precision; humans at abstraction under uncertainty. Typical split: Agent generates options, human judges; or human sketches prototypes, Agent refines—complementary synergy from clear role division.

## 5 Agent society: from individuals to societies

### 5.1 Behavioral attributes

**Social behavior:** healthy Agent societies need both collaboration and competition—over-emphasizing either harms system evolution. **Personality:** distinct personas add character but risk group bias/polarization—use filters for extreme speech, introduce impartial third-party Agents, and anchor personality in reason and civility.

### 5.2 Environments

Text-based worlds · sandbox simulations · physical environments—each shapes what social dynamics can be studied or deployed.

### 5.3 Social simulation

Study key attributes/mechanisms of Agent societies; harvest insights for human social science; and surface ethical/social risks inside simulated societies before they hit production.

## 6 Discussion

-   **6.1 Mutual benefit** — LLMs give Agents strong brains; Agent scenarios push LLM research into a virtuous cycle. Exploration does not excuse potential harm—researchers must own social responsibility and walk their talk.
-   **6.2 Evaluation** — benchmarks for task success, collaboration quality, safety, and long-horizon consistency remain immature relative to model leaderboards.
-   **6.3 Safety, trustworthiness, risks** — adversarial robustness; trustworthiness; misuse and cascading failures in multi-Agent systems.
-   **6.4 Scaling Agents** — compute cost, coordination overhead, and AaaS delivery models.
-   **6.5 Open problems** — grounding, long-term memory fidelity, multi-Agent norms, human alignment under growing autonomy.

## 7 Conclusion

LLM-based Agents unify language, tools, embodiment, and social simulation into a practical research program spanning labs and industry (including game studios like miHoYo). The Fudan × miHoYo survey’s lasting value is the shared map: brain–sense–act stack, deployment modes, society-level questions, and an explicit safety/ethics agenda. Unique Research’s English rewrite preserves that TOC so practitioners can navigate the Chinese original section by section.

_CN paper figures; attach on ingest._

— END —

---

Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20230918a055l200
On-site reading page: https://ffcap.cn/en/research/20230918a055l200