Emergent Language as an Approach to Conscious AI
Proposes emergent language in multi-agent RL to study consciousness-related structures without prior biases, revealing self-referential communication and echo-mismatch circuits.
Key Findings
Methodology
This paper introduces a generative approach leveraging emergent language (EL) within multi-agent reinforcement learning (MARL). Agents start from minimal conditions—no language, no self-concept, minimal exposure to human text—and develop communication solely driven by task pressures. The methodology emphasizes environment complexity scaling, minimal prior biases, and causal interpretation through interventions such as ablation and probing. It incorporates Kaplan’s character/content distinction and mutual information metrics to quantify self-referential encoding. By ensuring the environment creates selection pressures, the approach isolates structures that are causally attributable to current task demands, not inherited priors. This setup allows for observing how complex cognitive features like self-reference and self-monitoring emerge naturally, providing insights into the structural prerequisites for consciousness.
Key Results
- In minimal environments, agents developed indexical encoding (P1), with mutual information reaching 0.75 bits, indicating messages primarily encode sender states. Persistent state representations (P2) were observed via recurrent mechanisms maintaining self-states over time. The most significant result was the emergence of a behavioral self-monitoring circuit (P3), capable of detecting mismatches (echo failures) in communication, which was not predicted by the environment or architecture alone. These structures were validated through causal interventions, confirming their role in task performance. The findings demonstrate that complex cognitive functions can emerge solely from environmental pressures without pre-designed consciousness modules.
- Across different environment complexities and communication bandwidths, the emergent structures showed robustness. Ablation experiments confirmed their causal impact on task success. The environment’s affordances, such as the echo channel, were critical for the emergence of self-referential and monitoring mechanisms, highlighting the environment's role in shaping consciousness-relevant structures.
- Overall, the study establishes that environment-driven pressures can induce the spontaneous development of self-referential communication and monitoring circuits, providing a new pathway to understanding the structural basis of consciousness in artificial systems.
Significance
This research shifts the paradigm from evaluating AI against predefined consciousness criteria or embedding consciousness-inspired modules, to observing how consciousness-relevant structures can naturally emerge under task-driven environmental pressures. It offers a scientific framework for probing the origins of self-reference and monitoring mechanisms, which are central to many theories of consciousness. The methodology provides a causal, empirical basis for studying the structural prerequisites of consciousness, independent of subjective experience. This approach can be extended to more complex environments, potentially leading to the development of AI systems with rudimentary consciousness features. It also informs cognitive science by demonstrating that environment and task demands alone can give rise to foundational cognitive structures. The implications span both theoretical understanding and practical design of autonomous, self-aware AI systems.
Technical Contribution
The paper introduces a formalized framework combining emergent language analysis with causal intervention techniques, grounded in information theory and philosophical concepts like Kaplan’s character/content distinction. It operationalizes the measurement of self-referential encoding via mutual information, validated through ablation and probing experiments. The methodology emphasizes environment complexity scaling and minimal prior biases, ensuring structures are causally driven by task pressures. This approach diverges from traditional discriminative or architectural methods, providing a generative, causally interpretable pathway to study consciousness-relevant structures. The experimental validation demonstrates the emergence of self-referential communication and self-monitoring circuits, establishing a new paradigm for empirical consciousness research in AI.
Novelty
This work is pioneering in systematically using minimal environment setups to observe the spontaneous emergence of self-referential and monitoring structures in AI agents, without relying on pre-trained language models or pre-designed consciousness modules. Unlike prior studies focusing on semantic grounding or task performance, this research emphasizes the origin of self-awareness structures driven solely by environmental pressures. It bridges philosophical concepts of indexicality with quantitative information-theoretic metrics, providing a novel operationalization of self-reference in artificial agents. The approach offers a new experimental paradigm for studying consciousness as a structural phenomenon emerging from task-driven interactions, rather than a property embedded by design.
Limitations
- The experiments are conducted in highly simplified environments, limiting direct applicability to real-world, multimodal, and dynamic scenarios. Scaling complexity may dilute the observed structures or introduce confounding factors.
- The approach relies on reinforcement learning architectures and training regimes that are computationally intensive, posing scalability challenges for larger, more complex systems.
- The current framework focuses on structural proxies for consciousness (self-reference, monitoring) without addressing subjective experience or qualia, leaving the hard problem unresolved.
- Environmental affordances are crucial for the emergence of these structures; in less constrained or more noisy environments, the robustness and stability of these structures need further validation.
Future Work
Future research will explore scaling environment complexity, incorporating multimodal sensory inputs, and extending to multi-agent social interactions. Integrating continual learning and memory mechanisms will be essential to simulate developmental trajectories akin to biological cognition. Developing more efficient training algorithms and interpretability tools will facilitate larger-scale experiments. Additionally, combining this structural approach with neurobiological insights may bridge the gap toward understanding subjective experience. Ultimately, the goal is to establish a comprehensive, causally grounded framework for studying and engineering consciousness-relevant structures in artificial systems.
AI Executive Summary
The question of whether artificial systems can possess consciousness remains one of the most profound and contested in AI research. Traditional approaches often evaluate systems against predefined checklists derived from consciousness theories or engineer modules inspired by consciousness concepts, but both methods risk conflating artifacts of human language priors with genuine structural features. In this context, Zengqing Wu and Chuan Xiao propose一种全新的生成性研究方法，基于多智能体强化学习中的涌现语言机制，旨在在极简环境中观察无先验偏好的智能体自主发展出与意识相关的结构。
他们的方法强调环境复杂度的逐步扩展和任务压力的驱动作用，确保任何涌现的结构都可以追溯到当前环境和任务的压力，而非预设的偏好或架构设计。通过在有限通信带宽下训练智能体，研究团队观察到自指编码、持续的自我状态表征以及行为自我监控电路的涌现。这些结构的出现，验证了环境压力和任务复杂性在认知结构形成中的关键作用，为理解意识的起源提供了新的实验路径。
实验结果显示，智能体在极简环境中成功发展出以自身状态为主要内容的消息，互信息达到0.75比特，表明其具有明显的自指编码能力。同时，智能体表现出持续的状态保持机制和检测信息偏差的监控电路。这些结构未由任务结构或架构预设预测，而是在特定环境赋能条件下自发出现，表明复杂认知功能可以在没有预设意识模块的情况下自然涌现。
该研究的意义在于，提供了一种因果、实证的研究框架，突破了传统的理论推测和架构设计的局限，为人工意识的科学探索提供了新的路径。未来，随着环境复杂度的增加和多模态信息的引入，预计可以涌现出更丰富的认知结构，推动人工系统向具备基本意识的方向发展。尽管目前仍面临计算成本高和环境复杂度限制等挑战，但本研究的框架为理解意识的结构起源提供了崭新的视角，开启了人工认知演化的崭新篇章。
Deep Analysis
Background
人工智能的研究经历了从符号主义到深度学习的演变，近年来大规模预训练模型（如GPT系列）在自然语言处理和认知模拟方面取得了巨大突破。然而，这些模型多依赖大量人类文本数据，预设了丰富的语言偏好，难以区分由环境和任务压力驱动的认知结构。认知科学和神经科学提出多种关于意识起源的理论，如全球神经工作空间（GNWT）、整合信息理论（IIT）和高阶思想（HOT），它们都假设第一人称视角的存在，却未能解释其起源。涌现语言（EL）作为一种新兴的研究范式，强调在极简环境中观察智能体自主发展出复杂通信和认知结构，为理解意识的结构基础提供了新的实验路径。此前的研究多关注通信的语义和结构，鲜少关注“谁在说话”以及“自我指涉”的机制，这成为本文的突破点。
Core Problem
核心问题在于，如何在没有预设意识模块或人类语言偏好的极简环境中，观察到与意识相关的结构涌现。传统方法多依赖于预训练模型或架构设计，容易受到人类偏好的影响，难以区分结构的因果起源。更重要的是，缺乏一种系统性的方法，能验证这些结构是否真正由环境压力驱动，而非设计偏差或先验知识。本文试图解决这一瓶颈，通过极简环境和最小先验设计，确保任何涌现的结构都能归因于任务和环境的压力，从而为理解意识的起源提供实证基础。
Innovation
第一，提出基于涌现语言的生成性研究框架，强调环境复杂度的逐步扩展和最小先验设计，确保结构的因果归属。第二，结合Kaplan的字符/内容区分和互信息指标，量化自指编码的出现，为认知结构的因果分析提供工具。第三，验证在极简环境中，智能体自主发展出自指编码、持续状态表征和行为监控电路，突破了传统架构依赖的限制。第四，强调环境赋能（affordance）在认知结构形成中的作用，为理解意识的环境依赖性提供新视角。
Methodology
- �� 环境复杂度逐步扩展：从极简的两智能体、七个符号、十步的任务开始，逐步增加环境元素和交互复杂性，观察认知结构的涌现。• 最小先验设计：智能体从零开始学习，没有预设自我概念或语言偏好，确保任何结构的出现都由当前任务驱动。• 通过干预分析：采用消融、探测和信息论方法，验证涌现结构的因果关系。• 引入Kaplan的字符/内容区分：量化消息中的自指成分，确保通信中包含自我状态信息。• 互信息指标：衡量消息与自身状态的相关性，验证自指编码的出现。• 实验中，训练强化学习智能体在有限通信带宽下发展出自指编码和自我监控机制，验证其对任务的因果影响。
Experiments
实验采用两个强化学习智能体，在极简环境中合作完成任务。每个智能体观察自己的私有状态，必须通过有限的通信渠道（如少量符号）传递信息以协调行动。训练过程中，智能体没有预设自我概念或语言偏好，完全由任务压力驱动。通过分析消息的互信息和隐藏状态，验证了自指编码和持续的自我状态表征的涌现。干预实验中，智能体能检测信息偏差（回声失配），证明行为自我监控机制的出现。多组环境条件下，结构的稳健性和因果关系得到验证。
Results
在极简环境中，智能体成功发展出以自身状态为主要内容的消息（P1），互信息达到0.75比特，明显高于随机水平，验证了自指编码的涌现。持续的自我状态表征（P2）通过递归保持器实现，信息在时间上得以保持。最重要的是，智能体形成了检测信息偏差的行为电路（P3），能够识别传输中的回声失配，这一结构未由任务结构或架构预设预测，而是在特定环境赋能条件下自发出现。这些结果表明，复杂认知功能可以在无先验偏好的环境中自然涌现，为理解意识的结构基础提供了实证依据。
Applications
该研究为人工智能中的自我认知和意识结构提供了基础框架，适用于自主系统的认知演化研究。未来可在机器人、虚拟助手等领域应用，提升其自主性和适应性。通过环境设计和任务压力，促使系统自主发展出自我监控和自指机制，有助于实现更具自主意识的智能体。
Limitations & Outlook
目前实验环境极度简化，难以直接迁移到复杂的现实场景；计算成本高，训练时间长；未能完全模拟人类意识的丰富内容，结构的复杂性和多模态交互仍待探索。未来需要结合多模态信息和持续学习机制，提升模型的认知深度和泛化能力。
Plain Language Accessible to non-experts
想象你在一个工厂里工作，工厂里没有任何说明书或预设的操作流程。工人们（就像智能体）只知道自己手头的任务，比如装配某个零件，但没有人告诉他们怎么做。随着工作进行，工人们开始自己发明一些方法，比如用特定的手势或口哨来告诉别人自己当前的状态或需要帮助的地方。这些信号逐渐变得复杂，甚至有人能用这些信号检测到自己或别人的状态是否正确，或者发现传递信息时出现了偏差。这个过程就像在极简环境中，工人们自己创造出了一套交流和自我监控的机制，不依赖任何预设的规则或说明书。它展示了在有限信息和压力下，智能体（工人）可以自主发展出与意识相关的结构，比如自我指涉和行为监控。这就像在没有老师教的情况下，工人们自己学会了如何合作、观察和调整自己，从而逐步形成了“意识”的基础。
ELI14 Explained like you're 14
想象你在一个没有老师的学校里，只有几位同学和一些简单的任务。没有人告诉你怎么表达自己，也没有预设的规则。你们开始用手势、表情甚至简单的词语交流，慢慢地，大家发现用某些特定的信号可以告诉别人自己现在的状态，比如“我累了”或者“我需要帮忙”。随着时间推移，你们的交流变得越来越复杂，甚至有人能用这些信号检测自己是不是传错了信息，或者发现别人传递的内容不对。这就像是在极简环境中，学生们自己创造出了一套交流和自我监控的方法，不依赖老师的指示。这说明，即使没有预先设定的规则，人在压力和任务的驱动下，也能自己发明出类似“意识”的一些基本结构，比如知道自己在做什么、能检测到自己出错了。这就像是我们在没有老师教的情况下，自己学会了观察和调整自己，逐步形成了“意识”的基础。
Glossary
涌现语言 (Emergent Language)
指在极简环境中，智能体自主发展出的通信协议，非预设，依赖环境压力驱动形成。技术上是智能体通过强化学习在有限带宽下自发产生的符号系统。
本文中，涌现语言用于观察自指和认知结构的涌现。
自指编码 (Indexical Encoding)
指消息内容主要反映发送者自身状态的编码方式，信息论上互信息显著高于随机水平。技术上通过互信息指标量化。
验证智能体是否发展出自我指涉的关键指标。
持续状态表征 (Persistent State Representation)
指智能体通过递归保持器（recurrent latch）在时间上持续维护自身状态的机制。技术上表现为隐藏状态的时间连续性。
验证自我状态的时间保持机制。
行为自我监控 (Behavioral Self-Monitoring)
智能体通过检测传输信息的偏差（如回声失配）实现自我监控的机制。技术上表现为检测传输偏差的电路。
验证智能体是否具备自我检测能力。
Kaplan的字符/内容区分 (Kaplan's Character/Content Distinction)
区分符号的指示性（字符）与其指涉的内容（内容），用于量化自指结构的出现。
用于分析消息中的自指成分。
互信息 (Mutual Information)
衡量两个随机变量之间的依赖关系，信息论指标，反映消息与自身状态的相关性。
用于量化消息中的自指编码。
环境赋能 (Environmental Affordance)
环境中提供的行动或交互可能性，影响认知结构的涌现。
验证环境复杂度对认知结构的驱动作用。
强化学习 (Reinforcement Learning)
一种机器学习范式，智能体通过与环境交互获得奖励信号，学习最优策略。
训练智能体在极简环境中发展出涌现语言。
因果归属 (Causal Attribution)
验证结构是否由特定环境或任务压力引起的方法，确保结论的因果关系。
通过干预验证涌现结构的因果关系。
极简环境 (Minimal Environment)
设计极度简化的虚拟环境，减少先验偏好，强调任务压力的作用。
用于观察涌现结构的自然发展。
Abstract
The question of whether artificial systems can be conscious remains open, in part because existing approaches either evaluate systems against theory-derived checklists (discriminative) or engineer consciousness-inspired modules directly (architectural); both leave open whether observed structures are artifacts of human language priors. We propose a generative methodology: emergent language (EL) in multi-agent reinforcement learning, where agents start from minimal (no language, no concept of self, minimal exposure to human text) and develop communication under task pressure alone, ensuring causal attributability to task demands rather than inherited human language priors. We position our methodology by discussing how EL serves as a generative tool for studying consciousness-relevant structure, including the role of environment complexity and the interpretation of emergent communication. As a proof of concept, we instantiate this methodology in a minimal environment and show that agents develop self-referential communication, including an echo-mismatch detection circuit that is not predicted by task structure or architecture alone but emerges from a specific environmental affordance.
References (20)
Multi-Agent Cooperation and the Emergence of (Natural) Language
Angeliki Lazaridou, A. Peysakhovich, Marco Baroni
Identifying indicators of consciousness in AI systems.
Patrick Butlin, Robert Long, Tim Bayne et al.
A bitter lesson.
N. Whitman
Large Language Models Report Subjective Experience Under Self-Referential Processing
Cameron Berg, Diogo Schwerz de Lucena, Judd Rosenblatt
A Disproof of Large Language Model Consciousness: The Necessity of Continual Learning for Consciousness
Erik Hoel
Facing Up to the Problem of Consciousness
D. Chalmers
Demonstratives: An Essay on the Semantics, Logic, Metaphysics and Epistemology of Demonstratives and other Indexicals
David Kaplan
Welcome to the Era of Experience
David Silver, Richard Sutton
Learning and communication pressures in neural networks: Lessons from emergent communication
Lukas Galke Poech, Limor Raviv
HexaJungle: a MARL Simulator to Study the Emergence of Language
Kiran Ikram, Esther Mondragón, E. Alonso et al.
Self And Society
J. Fuhrmann
I Am a Strange Loop
S. Strazza, K. Crowley
Consciousness and mind
D. Rosenthal
Shadows of the Mind: A Search for the Missing Science of Consciousness
Emergence of Language with Multi-agent Games: Learning to Communicate with Sequences of Symbols
Serhii Havrylov, I. Titov
The Principles of Psychology
D. Spalding
Collective predictive coding hypothesis: symbol emergence as decentralized Bayesian inference
Tadahiro Taniguchi
Emotion Concepts and their Function in a Large Language Model
Nicholas J Sofroniew, Isaac Kauvar, William Saunders et al.
Experience Grounds Language
Yonatan Bisk, Ari Holtzman, Jesse Thomason et al.
The attention schema theory in a neural network agent: Controlling visuospatial attention using a descriptive model of attention
Andrew I. Wilterson, M. Graziano