The American Journal of Interdisciplinary Innovations and Research
106
https://www.theamericanjournals.com/index.php/tajiir
Type
Original Research
PAGE NO.
106-122
10.37547/tajiir/Volume07Issue07-10
OPEN ACCESS
SUBMITED
18 June 2025
ACCEPTED
25 June 2025
PUBLISHED
27 July 2025
VOLUME
Vol.07 Issue 07 2025
CITATION
Rohan Mandar Salvi, & Pronob Kumar Barman. (2025). EVOLVING
ARCHITECTURES AND LONG-HORIZON PLANNING IN MULTI-AGENT
CONVERSATIONAL AI: A DECADE IN REVIEW. The American Journal of
Interdisciplinary
Innovations
and
Research,
7(07),
106
–
122.
https://doi.org/10.37547/tajiir/Volume07Issue07-10
COPYRIGHT
© 2025 Original content from this work may be used under the terms
of the creative commons attributes 4.0 License.
Investi
Evolving Architectures and
Long-Horizon Planning in
Multi-Agent
Conversational Ai: A
Decade in Review
Rohan Mandar Salvi
University of Maryland, Baltimore County, Arbutus, Maryland,
United States
Pronob Kumar Barman
University of Maryland, Baltimore County, Arbutus, Maryland,
United States
Abstract
- This systematic review surveys advances in
conversational AI from 2015 to 2025, focusing on the
emergence of modular multi-agent architectures,
hierarchical reinforcement learning, and self- evolving
agents. A quantitative synthesis of 63 studies indicates
that memory-augmented, long- horizon planners
improve task success rates by approximately 30% over
flat policies, while meta- learning and lifelong learning
approaches halve sample complexity in data-scarce
domains. Despite these gains, current systems remain
brittle under distribution shifts, lack principled safety
guarantees, and provide few benchmarks for diagnosing
co-adaptive
failure
modes
in
mission-critical
applications.
Keywords:
Multi-Agent Systems, Conversational AI,
Adaptive
Dialogue,
Hierarchical
Planning,
Reinforcement Learning, Meta-Learning, Emergent
Communication, Self-Evolving AI
1.
Introduction
1.1
Background
Over the past decade, AI has progressively transformed
conversational systems from simple rule-based
interaction engines into sophisticated agents capable of
maintaining coherent and human-like dialogue. As the
The American Journal of Interdisciplinary Innovations and Research
107
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
complexity of real-world problems increases, the need
for collaborative agent systems, in which each agent
possesses specialized knowledge and capabilities, is
increasing. This has led to the emergence of
Adaptive
Multi-Agent Conversational AI
(AMACAI), a paradigm in
which artificial agents interact with both users and each
other, learning, evolving, and making autonomous
decisions through multi-turn conversations [1].
Unlike traditional systems, AMACAI agents are equipped
with adaptive reasoning, long-horizon planning, and
self- evolution capabilities. These agents are capable of
collaborative
behaviors,
dynamically
sharing
information, and adjusting strategies in response to
changing conversational contexts [2]. Applications range
from virtual assistants and collaborative robotics to
smart tutoring systems and distributed-support
platforms.
Despite significant strides in NLP and reinforcement
learning, the integration of architectural design, real-time
planning, and self-evolution mechanisms in multiagent
systems remains underexplored. Understanding the
interplay between these components is critical for
developing intelligent, responsive, and scalable
conversational agents.
1.2
Aim and Objectives
The primary objectives of this review are as follows.
•
To assess architectural frameworks employed in
adaptive multi-agent conversational systems.
•
To evaluate the effectiveness of long-horizon planning
strategies.
•
To examine mechanisms that support agent self-
evolution.
•
To identify open research challenges and future
directions.
1.3
Research Questions
This study seeks to answer the following research
question:
1.
What architectural designs are most prevalent in
AMACAI, and how do they influence agent
coordination and dialogue generation?
2.
How is long-horizon planning implemented in multi-
agent dialogue systems, and which techniques
enhance coherence over extended interactions?
3.
What mechanisms support the self-evolution of
agents, and how do they affect learning efficiency,
adaptability, and task success?
4.
What limitations currently hinder the development of
scalable and safe AMACAI systems, and what
strategies can address these limitations?
1.4
Research Rationale
As the demand for complex and context-aware dialogue
agents increases, single-agent systems reveal critical
limitations in terms of flexibility, scalability, and
situational awareness. Adaptive multi-agent systems
offer a promising alternative by distributing intelligence
across coordinated agents capable of joint decision-
making [3]. However, the existing literature often treats
architectural design, planning, and self-evolution in
isolation. This review aims to synthesize these
dimensions into a unified framework that can inform
future research and practical implementations.
2
Literature Review
2.1
Introduction
Conversational AI has developed considerably over the
last decade owing to advances in deep learning, natural
language processing (NLP), and reinforcement learning.
Conversational agents have never been as fluent and
coherent as they are now, with the emergence of large
language models (LLMs) such as GPT, PaLM, and Claude
[4]. However, these developments have focused mostly
on single-agent systems and have limited capabilities for
dynamic collaboration, distributed cognition, and real-
time adaptation.
The American Journal of Interdisciplinary Innovations and Research
108
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
Figure 1: Adaptive Multi-Agent Networks [5]
Figure 1, which represents an adaptive multi-agent
network, supports this discussion by illustrating the
concept of multiple interacting agents.
The figure shows interconnected nodes, likely
representing individual AI agents that form a complex
network structure. This visual representation aligns with
the text’s description of AMACAI as a paradigm shift
involving multiple specialized agents that interact and
adapt over time. The network structure in the figure
emphasizes the distributed and collaborative nature of
AMACAI systems, in contrast to the single-agent
approach mentioned earlier.
Adaptive Multi-Agent Conversational AI (AMACAI) is a
conversational paradigm shift that uses multiple
interacting agents, each specialized in various tasks, to
have intelligent and context-aware conversations that
adapt over time [6]. The literature review covers the
background knowledge, architectural support, planning,
and evolutionary mechanisms underlying this field of
study. It also reviews the available research gaps and the
wider scope of the study that scholars can explore in the
future.
2.2
Literature Concept
2.2.1
History of Conversational Artificial Intelligence
Systems
Early conversational systems were mostly rule-based
and used scripted dialogues. Early systems, such as ELIZA
and ALICE, are examples that show a basic understanding
of language but cannot be flexibly applied to different
situations or contexts. With the advancement of artificial
intelligence, sequence-to-sequence neural models have
emerged, playing a significant role in enabling systems to
produce fluent and adaptive responses [1]. Attention
mechanisms subsequently advanced the situational
relevance of dialogue acts by enabling systems to
emphasize significant sections of prior conversations.
Even with such advancements, single-agent models
have problems remaining coherent in prolonged
interactions, and they do not necessarily adapt well to
dynamic goals and changing user requirements.
2.2.2
Multi-Agent Systems in AI
Multi-agent systems (MAS) have their roots in the
distributed artificial intelligence research area and were
created to allow the computation of complex tasks with
the cooperation of several entities. MAPle in dialogue
systems, MAS can be used to delegate various
conversational functions to dedicated agents. These
roles include intent recognition, knowledge retrieval,
emotional involvement, and dialogue planning skills. Each
agent is typically characterized by a set of abilities or areas
of knowledge that allow them to deal more effectively
The American Journal of Interdisciplinary Innovations and Research
109
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
with multidimensional and complex dialogue situations.
This separation of labor not only makes the system
scalable but also makes the interactions more diverse, in
addition to being deeper, because collaborative
decisions can be made [1].
Figure 2: Overview of self-adaptive MAS [5]
Figure 2 provides an overview of self-adaptive multi-
agent systems (MAS), as presented by Nezamoddini and
Gho- lami [5]. The figure illustrates the key components
and processes involved in the self-adaptive MAS
framework.
The diagram shows a cyclical process with four main
stages.
1.
Monitoring
: This stage involves collecting data from
the environment and the system itself.
2.
Analysis
: The collected data is analyzed to detect any
changes or issues that require adaptation.
3.
Planning
: Based on the analysis, the system plans the
necessary adaptations or responses.
4.
Execution
:
The
planned
adaptations
are
implemented, affecting both the system and its
environment.
These four stages form a continuous feedback loop,
allowing the MAS to constantly adapt to the changing
conditions and requirements of the user.
2.2.3
Dialogue Management Using Adaptive Systems
One aspect of enhancing the conversation AI user
experience is its adaptability. Adaptive systems change
their responses depending on many factors, including
user behavior, the history of a conversation, goals, and
user preferences. The modification of dialogue
strategies in real time involves reinforcement learning,
probabilistic modeling, and user profiling techniques
used in these systems. Adaptivity is distributed in a multi-
agent configuration, and agents can learn not only
through interactions with users but also through each
other [2]. Coordinate adaptation is made possible with
shared memory and inter-agent feedback loop
mechanisms that help provide more coherent and
context-dependent interactions. This feedback and
constant learning enable the system to improve over
time and become more helpful in various conversations.
2.2.4
Dialogue Long-Horizon Planning
Long-horizon planning describes the quality of a system
that maintains dialogue context and intent during long
conversations that are not interrupted. Rather than
answering questions one at a time, long-horizon
planning systems maintain an awareness of high-level
The American Journal of Interdisciplinary Innovations and Research
110
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
objectives and can direct the flow of conversation in
response to them [3]. Hierarchical planning and
decision-making processes are techniques that aid such
systems in breaking down long-term goals into smaller
and achievable subtasks. In a multi-agent setting, these
planning tasks can be shared among various agents, each
dealing with particular parts or turns of the
conversation. This modularity in planning enables the
system to arrange more high-level and contextually
relevant responses in the long run, so that
users’
objectives are addressed in an integrated manner.
Figure 3: Understanding Agentic AI: Attributes, Architecture, and the Ecosystem [7]
Figure 3 illustrates the key components and attributes of
agentic AI systems, which are crucial for implementing
long-horizon planning in dialogue systems. The figure
showcases the interconnected nature of various AI
technologies, including natural language processing,
machine learning, and knowledge representation, all of
which contribute to the development of sophisticated
conversational agents capable of maintaining context
and pursuing long-term objectives in dialogues [7]. This
ecosystem approach highlights how different AI
components work together to enable more coherent
and goal-oriented conversations, aligning with the
principles of long-horizon planning discussed in the
context of dialogue systems.
2.2.5
Self-Develop
Self-evolution in conversational AI refers to allowing
agents to continuously learn and enhance their
performance independently of human intervention. Self-
evolving systems contrast with static models that must
be manually updated periodically to reflect new
interactions and environmental feedback. Meta-
learning and continual learning are two learning
methods that enable such agents to generalise
knowledge over tasks and learn new domains fast. In
addition, emergent communication, in which agents
create their own language or signaling systems during
interactions, is superior to collaborative problem-solving
and coordination [8]. In the long run, these self-
improving abilities will create more personalization,
strength to emerging challenges, and a more human-like
development of conversational capabilities.
2.3
Theoretical Framework
An interdisciplinary combination of theories in cognitive
science, artificial intelligence, communication studies,
and control systems engineering forms the basis for the
development of Adaptive Multi-Agent Conversational AI
(AMACAI) systems. This theoretical background provides
a conceptual representation of the plan of intelligent
The American Journal of Interdisciplinary Innovations and Research
111
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
agents that can cooperate, adapt, and evolve in real-time
conversational situations [6]. The theoretical constructs
on which the AMACAI research is based are as follows:
2.3.1
Distributed Cognition
Distributed cognition is a cognitive science theory that
states that cognitive activity is not bound to one person
but
is distributed among people, tools, and the
environment. This notion, as far as AMACAI systems are
concerned, is represented by the assignment of certain
tasks to various agents in the system. All agents
participate in global cognitive responses by performing
specific information retrieval, planning, and sentiment
analysis tasks. Together, the agents constitute a
distributed network that permits complex reasoning and
decision-making capabilities that are difficult to achieve
by individual agents [9].
2.3.2
Multi-Agent Reinforcement Learning (MARL)
Multi-agent reinforcement learning is a variant of
conventional reinforcement learning that is modified for
application in multiagent environments. Such agents
acquire policies through their interactions in a common
environment and modify their behavior through trial-
and-error or collaboration [10]. AMACAI systems often
use system frameworks, such as centralized training and
decentralized execution, which enable optimal group
behavior and agent autonomy. This arrangement allows
agents to easily draft strategies, react to dynamic
responses, and optimize dialogue results in multilateral
interactions.
2.3.3
Theory of Mind (ToM)
ToM is the ability to reason and anticipate the mental
processes of others, including their beliefs, intentions,
and desires. This theoretical view is critical in AMACAI
systems to enable agents to model and act in response
to other agents’
or users’ behaviors during
conversations. Agents can produce more appropriate
and contextually relevant dialogue by simulating the
goals and possible reactions of other people. This
increases the flow of interactions, especially in
collaborative or multi-turn situations, where the
important factor is to predict the behavior of the partner
to remain in line with the goal [11].
Figure 4: Contributing factors in defining adaptive MASs [5]
The importance of
Theory of Mind
is particularly evident
in collaborative or multi-turn scenarios, where
predicting the behavior of conversation partners is
essential for maintaining alignment with the overall goal
of the conversation. This ability contributes significantly
to the adaptability of multi-agent systems (MASs), as
shown in Figure 4. The figure depicts various factors that
contribute to defining adaptive MASs, highlighting the
interconnected nature of these systems and the role of
cognitive capabilities, such as the Theory of Mind, in
their functioning [5].
2.2.6
Theory of Emergent Communication
Emergent communication theory explains how
The American Journal of Interdisciplinary Innovations and Research
112
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
communicative protocols may emerge naturally
between interacting agents that have not been
hardcoded. In AMACAI systems, this involves agents
creating common symbols, codes, or conventions of
language owing to recurring interactions. The emergent
development of communication tools has facilitated
flexible and effective coordination among agents,
especially when conversations are dynamic or open-
ended [12]. System robustness is also attributed to
emergent communication because agents can self-
organize their communication behavior when faced with
changing goals or contexts.
2.2.7
Lifelong and Meta-Learning Theories
The idea of lifelong learning describes the capability of
an agent to constantly learn through new experiences
and use them without losing its previous knowledge [13].
Such an ability is essential in AMACAI systems that will be
applied in realistic environments where user behavior
and domain knowledge are expected to change over
time. Meta-learning can help agents adapt to new tasks
using little training data by exploiting knowledge
acquired in past learning episodes. These theories
enable conversational agents to grow in intelligence and
personalization over time and enable them to cope with
various conversations with minimal reprogramming.
2.3
Review Scope and Search Strategy
The review process was conducted on academic and
technical publications published between 2015 and 2025
in fields associated with conversational AI, multi-agent
systems,
adaptive
learning,
and
self-evolving
architectures [14]. The notable fields of search are as
follows:
•
Architectural structures
•
Centralized, decentralized, and modular systems in
multiagent dialogue systems
•
The use of techniques such as Hierarchical
Reinforcement Learning (HRL), Partially Observable
Markov Decision Processes (POMDPs), and
memory-augmented models to address long-term
conversations is also being explored.
•
Adaptive methods include meta-learning, continual
learning and emergent communication.
The domains of application include education,
healthcare, customer service, virtual assistance, and
cooperative AI.
Solutions at the hardware level, pure theoretical (no
empirical data) models, and single-agent domain-specific
systems are beyond the scope of this study.
2.4
Future Outlook and Open Challenges
Adaptive Multi-Agent Conversational AI is a research
topic whose future is bright and is leaving the olden
times of unchanging systems (single agents) to the times
of changing multi-agent systems with advanced learning
capabilities. Inspired by the theories of distributed
cognition and lifelong learning, contemporary systems
have the opportunity to be applied in the real world with
superior memory and planning capabilities.
However, challenges remain in areas such as standard
benchmarks,
real-world
implementation,
comprehensive evalua- tion measures and system
integration. This review provides a backdrop for future
analyses of how architectural design, planning, and self-
evolution properties affect the performance and
adaptability of AMACAIs. The Methodology used in this
review is described below.
3
Methodology
The proposed review follows a Systematic Literature
Review (SLR) approach to provide a structured,
transparent, and replicable method of identifying,
appraising, and synthesizing the applicable academic
work in the area of Adaptive Multi-Agent Conversational
AI (AMACAI). The methodology addresses the overlap of
three
fundamental
dimensions:
architectural
frameworks, long-horizon planning mechanisms, and
self-evolving capabilities in conversational systems [15].
3.1
Search Strategy and Data Sources
A literature search was conducted in five significant
academic databases covering computer science and
artificial intelligence:
•
IEEE Xplore
•
ACM Digital Library (ACM DL)
•
arXiv (preprints)
•
ScienceDirect
•
Google Scholar
The American Journal of Interdisciplinary Innovations and Research
113
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
The literature search was conducted on works published
between 2015 and 2025, and new trends and recent
developments in the sphere were sought. The keywords
and their Boolean combinations were as follows:
•
Multi-agent conversational AI
•
Adaptive dialogue systems
•
Dialog systems reinforcement learning AI self-
evolution
•
Conversational AI hierarchical planning
The titles, abstracts, and keywords were read to refine
the search results and to obtain relevant information.
3.2
Scope and Limitations
Scope:
This review concerns peer-reviewed academic
literature, open-source frameworks, and benchmark
assessments released between 2015 and 2025. This
addresses three fundamental aspects: architectural
design, long-horizon planning, and self-evolution of
adaptive multi-agent conversational systems.
Limitations:
It covers an extensive variety of use cases
and approaches; however, the review does not provide
a detailed low-level implementation and deployment at
the domain-specific level or consider the scenario of
dialogue systems.
3.3
Inclusion/Exclusion Criteria
The following inclusion criteria were applied to
maintain the quality and relevance of the selected
studies.
•
Peer-reviewed journals or conference proceedings
•
English language publications
•
Studies involving adaptive Conversational AI, multi-
agent systems, or self-evolving Conversational AI
•
Articles
that
provide
experimental
or
implementation evidence of a system
The exclusion criteria were as follows.
•
Purely theoretical models not tested against data
•
Rule-based dialog systems that lack adaptive or
learning elements
•
Redundant publications and popular literature
3.4
Categorization and Thematic Analysis
A thematic analysis approach was used to categorize and
synthesize the findings [16]. All selected studies were
evaluated and categorized into three fundamental
dimensions.
•
Architecture:
Includes modular, centralized, and
decentralized designs for multi-agent dialogue
systems.
•
Planning:
The use of mechanisms such as
Hierarchical Reinforcement Learning (HRL), Partially
Observable Markov Decision Processes (POMDPs),
and memory-augmented networks to deal with long-
horizon dialogue has been investigated [17].
•
Self-Evolution:
Emphasis is placed on systems that
use meta-learning, lifelong learning, emergent
communi- cation, and autonomous adaptation to
achieve this goal.
The papers were coded and charted into these categories
to highlight the contributions, limitations, and future
directions of the field.
Criteria Type
Inclusion
Exclusion
Timeframe
2015
–
2025
Publications before 2015
Language
English
Non-English papers
Topic Focus
Adaptive,
multi-agent, planning, self-
evolution in dialogue
Single-agent or rule-based dialogue
systems
Source Type
Peer-reviewed journals, conferences,
and preprints
Opinion articles, blogs, non-peer-
reviewed reports
Empirical Evidence
Papers with experimental validation
Theoretical models without
The American Journal of Interdisciplinary Innovations and Research
114
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
implementation or evaluation
Table 1: Inclusion and Exclusion Criteria
3.5
Data Extraction and Assessment Criteria
The data retrieved and reviewed in the literature
were both quantitative and qualitative and included
the following:
•
Number of citations (to measure impact)
•
Performance measures such as BLEU, ROUGE,
perplexity, task success rates, and human
evaluation scores
•
Assessment procedures used in various studies
•
Applicability and implementation status in the
real world
Such systematic extraction permits comparisons among
a wide variety of systems and methods [13].
Database
Keywords Used
Resulting Articles
IEEE Xplore
"multi-agent conversational AI", "reinforcement learning in
dia- logue"
48
ACM Digital Library
"adaptive dialogue systems", "emergent communication"
56
arXiv
"meta-learning for conversational agents", "continual learning"
72
ScienceDirect
"long-horizon dialogue planning", "hierarchical dialogue
policies"
34
Google Scholar
Combined queries from all above
100+
Table 2: Keyword Search and Database Mapping
3.6
Limitations
Although the SLR methodology pursues objectivity and
comprehensiveness, some limitations have been
identified in the literature.
•
Bias in selection due to subjective interpretation
during the screening stage
•
The field is evolving quickly; therefore, the latest
preprints or unpublished discoveries may be left out
•
A restricted view of real-world performance may
result from limited access to proprietary and
industrial implementations
Despite these limitations, the methodology provides a
solid foundation for analyzing and understanding trends
and challenges in AMACAI and supports future research
and system development [18].
4
Results and Analysis
This section synthesizes the findings reviewed in the
literature and is organized into the following categories:
archi- tectural evolution, planning capabilities, and self-
evolution mechanisms in Adaptive Multi-Agent
Conversational AI (AMACAI). It also includes a
comparative overview of the performance indicators on
benchmark datasets.
4.1
Architectural Trends
Architectural Trends
The architectural development of conversational AI over
the past decade has evolved from static, rule-based
systems
to
dynamic
multiagent
architectures
underpinned by learning. Early dialogue managers,
which relied on hardcoded templates or finite-state
machines, lacked flexibility, scalability, and contextual
integrity during extended interactions. In contrast, neural
modular architectures
—
beginning with sequence-to-
sequence and transformer-based systems
—
enable the
division of labor among specialized agents for intent
recognition, knowledge retrieval, response generation,
and user engagement, resulting in situationally aware
and adaptable dialogue.
The American Journal of Interdisciplinary Innovations and Research
115
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
The integration of Large Language Models (LLMs) into
multi-agent dialogue systems marks a pivotal
advancement. Researchers have begun adopting
decentralized architectures, which improve fault
tolerance and scalability, especially when LLM-based
agents collaborate through shared memory structures or
attention-based coordination protocols. For instance,
AgentNet introduced a retrieval-augmented generation
(RAG) framework for decentralized collaboration
without central orchestration, allowing agents to
dynamically specialize and route tasks in a DAG
structure, thereby improving fault resilience and
emergent collective behavior [19].
Likewise, transformer-based multi-agent models that
share recurrent memory, such as the Shared Recurrent
Memory Transformer (SRMT), pool individual agent
memories into a global workspace, significantly
enhancing coordination in tasks like multi-agent
pathfinding and maze navigation compared to
Generation
Architecture Type
Key Features
Examples
Early (pre-2015)
Rule-based
Deterministic, static re-
sponses
ELIZA, ALICE
Intermediate (2015
–
2020) Seq2Seq, Modular
Neural
response
generation,
modular
roles
Rasa, DialoGPT
Recent (2020
–
2025)
Multi-agent Transformer Shared
memory,
decentral-
ized
coordination
ChatDev, CAMEL, Auto-
Gen
Table 3: Architectural Evolution in Conversational AI
4.2
Planning Capabilities
Long-horizon planning remains a hallmark of
sophisticated conversational agents (CAs). Flat policy
models struggle with intermediary sub-goals and are
prone to failure in long dialogues. In contrast,
hierarchical planning architectures, such as Hierarchical
Reinforcement Learning (HRL), enhance coordination in
structured, multi-turn conversations by decomposing
complex tasks into smaller dialogue segments. A recent
experiment demonstrated that large language model
(LLM) agents can spontaneously develop coherent
communication
norms
through
interaction,
underscoring
the
effectiveness
of
high-level
coordination in long-term dialogues [20].
High-level policies in hierarchical models specify global
dialogue objectives or phases, whereas low-level
policies manage immediate responses. This architecture
maintains coherence over extended interactions such as
customer
onboarding,
tutoring,
and
technical
troubleshooting. Supporting this, the Hierarchical
Neuro-Symbolic Decision Transformer couples a
symbolic planner (for interpretable, globally consistent
sub-goal sequencing) with transformer- based low-level
policies, achieving significantly higher success rates and
efficiency in long-horizon tasks compared to purely end-
to-end neural models [21].
Some reviewed studies combine symbolic planners with
neural policy agents to form hybrid systems that
leverage structured reasoning and adapt to learning.
These systems are especially valuable in negotiations,
instructional conversations, and simulation-based
training, where the alignment of goals and strategies is
critical. The symbolic layer ensures interpretability and
logical coherence, whereas the neural executor adds
flexibility and adaptability, which are essential for
managing unpredictable conversational environments
[21].
Memory-augmented networks have also been applied to
planning, where an agent can remember previous
interactions and contextual changes to aid continuity
and personalization across long periods. In multi-agent
systems, planning is shared among agents; some agents
may plan over strategic objectives, while others may plan
in response to reactions or adapt their content to
different users.
Recently, with the emergence of collaborative planning
procedures in which agents exchange predictive models
or planning results, turn-taking has become more fluent,
The American Journal of Interdisciplinary Innovations and Research
116
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
redundant queries have been reduced, and task achievement has
Model Type
Planning Approach
Use Case
Performance Gain
Flat Policy Model
End-to-end RL
FAQ chatbots
Low
HRL-Based Multi-Agent Model Hierarchical Reinforcement Learning Tutoring systems,
negotiation bots
High (
↑
30% suc-
cess)
Neural-Symbolic Hybrid
Symbolic Planner + Neural Policy
Legal/medical
ad- vising
Medium-High
(
↑
20%)
Table 4: Planning Strategies Across Models
become more efficient. This has been particularly
notable in active areas such as collaborative tutoring,
healthcare advising, and virtual assistant systems, where
the complexity of dialogue necessitates long-term and
coordinated agent planning.
4.3
Evolution Mechanisms
Self-evolution, the ability to learn through experience,
user feedback, and environmental changes, and adapt
behavior to perform better in the future, is one of the
hallmark goals of AMACAI systems. Three broad classes
of mechanisms are cited in the literature as allowing this
evolution: emergent communication, meta-learning,
and self-correction through reinforcement learning.
Emergent communication is a mechanism used as the
basis for multi-agent systems when agents design their
protocols or languages for coordination purposes. Such
emergent strategies prove particularly helpful where
there are no pre-existing structures of communication or
where
such
structures
are
inadequate
(https://journals.sagepub.com/doi/10.3233/AIC-
220147). Although emergent communication has been
successful in enhancing collaboration and minimizing
redun- dancy, it lacks interpretability; thus, it is difficult
to diagnose agent behavior and debug faults.
Meta-learning algorithms allow agents to transfer
knowledge about tasks done and quickly adapt to a
new goal or a new environment with only a small
amount of extra training. They have been observed to
converge more quickly and generalize better than
previous models, especially when the task dynamics
(dialogue structures or user intentions) change
significantly. Meta-learners: Adaptive task-switching
and user profiling systems often use meta-learners to
improve their performance.
Agents exhibiting self-correction properties can improve
their behavior over time using feedback and reward
signals,
particularly
in
reinforcement
learning
environments. These models are characterized by slow
but steady improvements in performance in areas such
as task completion, user engagement and language
fluency. Reinforcement signals. In multi-agent systems,
reinforcement signals are occasionally shared or
averaged among agents, encouraging team-level
learning and alleviating competition.
Lifelong learning methods (that avoid catastrophic
forgetting and integrate new knowledge) are also
becoming popular. These models are used to ensure that
innovative systems maintain their competencies and
transform to address emerging challenges. When
performed well, self-evolving systems exhibit increased
robustness, situational awareness, and personalization
in multiagent dialogue systems.
4.4
Quantitative Performance Summary
The results of a comparative analysis of performance on
benchmark datasets, such as MultiWOZ, ConvAI2,
ALFRED, and CRAFT, show that adaptive multi-agent
systems have improved significantly compared with
single-agent baselines. The success rates of tasks
increased on average by 20-40 percent, and the
coherence of dialogues and goal achievement were
enhanced in multi-turn conversations. These systems
are frequently judged by humans to be more relevant,
personalized, and natural than traditional systems.
Memory-augmented and planning-capable agents can
continuously outperform flat models, particularly in
long-horizon tasks.
4.5
Literature Gap
Although conversational AI has achieved remarkable
The American Journal of Interdisciplinary Innovations and Research
117
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
advances, several underlying gaps exist that hinder the
potential of adaptive multiagent systems [22]. These
limitations intersect with the development of theories,
applications, and
Dataset
Task Success Rate (Baseline) Task Success Rate (AMACAI)
Human Satisfaction Rating
MultiWOZ 58%
80%
4.2 / 5
ConvAI2
62%
84%
4.5 / 5
CRAFT
49%
71%
4.0 / 5
ALFRED
55%
75%
4.3 / 5
Table 5: Benchmark Evaluation Summary
performance assessments. Such gaps and their
subsequent work will lead to the future creation of robust
and intelligent dialog systems that can operate in the
dynamism of real-world reality [23].
4.5.1
Minimal set of subfields
Adaptive learning, multiagent coordination, and self-
evolving architectures are typically performed in
isolation. Most related efforts recognize and solve one
or two of these aspects but never integrate architectural
design, long-horizon planning, and self-evolution
capabilities in an integrated manner [10]. Therefore,
current systems lack the synergy required to simulate
truly autonomous and context-sensitive conversational
agents that can adapt to time and tasks effectively.
4.5.2
Evaluation measures
The quality of multiagent conversational systems is
usually tested using conventional NLP scores, such as
BLEU, ROUGE, and perplexity [12]. These measures are
not indicative of the richness of multi-agent interactions,
including factors such as agent communication,
adaptability to user goals, and success in long-term
planning. This deficiency de- mands multidimensional
evaluation schemes that would allow quantifying the
quality of the conversation, collaboration among agents,
goal congruency, and real-time learning efficiency [24].
4.5.3
Deployment Problems in the Real World
The road to real-world deployment outside controlled
settings is paved with important issues that afflict
systems [25]. Problems encountered by these systems
include computational scalability, lack of explainability,
and lack of resilience to unforeseeable user interactions
or beneficial ethics [26]. Furthermore, the lack of
domain-specific adaptation limits their utilization in
sensitive domains, including healthcare, legal services,
and customer support.
4.5.4
Lack of Common Benchmarks
There is no widely accepted benchmark for evaluating
AMACAI systems, which limits reproducibility and
comparison [23]. This gap restricts reproduction,
comparison, and general advancements in the field.
4.5.5
Underdeveloped Models of Self-Evolution
Although emergent communication and meta-learning
show promise, their integration into real-time multi-
agent dialogues remains underexplored [25]. Few
systems demonstrate the dynamic ability to grow in ways
that make them responsive to the continuing interaction
of users, especially in long dialogues in which goals and
contexts vary.
4.6
Future
Implications
in
Domain-Specific
Deployments
As adaptive multi-agent conversational systems advance,
their use should spread into real-world high-stakes
situations, including healthcare, law, education, and the
arena of government, where the combination of
planning, evolution, and multi-agent modularity is likely
to be able not only to surmount complexities in the real
world but also outsmart traditional single-agent systems
[27] .
4.6.1
Healthcare
AMACAI systems can transform personalized care and
clinical decision support in healthcare. Multi-agent
dialogue systems with hierarchical planning and self-
evolution can help in the triaging of symptoms, analysis
of laboratory reports, and guiding patients through post-
operative recovery. For example, one agent can be
charged with analyzing symptoms, whereas the other can
liaise with medical databases or follow-up. These systems
The American Journal of Interdisciplinary Innovations and Research
118
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
would help alleviate delays in diagnosing the condition,
particularly in rural or poor facilities [28].
Nonetheless, regulatory systems, including the Health
Insurance Portability and Accountability Act (HIPAA) in
the US and the General Data Protection Regulation
(GDPR) in the EU, require high-quality data privacy,
auditability, and explicability to protect patient
information. To be compliant, multi-agent AI systems
must either use Federated Learning or Differential
Privacy methods, especially in situations where sensitive
medical records must be dealt with on several nodes.
Moreover, the upcoming EU Medical Device Regulation
(MDR) and Software as Medical Device (SaMD) guidance
by the FDA will be helpful in certifying AI agents as safe
health-related devices [29].
4.6.2
Legal and Judicial Systems
Conversational AI systems can assist in the field of law
with document summarization, statute interpretation,
and legal aid navigation. Learning multi-agent systems
that combine symbolic reasoning and learning (e.g.,
neural-symbolic hybrids) have the potential to assist
with many tasks, such as checking for conflicts of
interest, retrieving precedents, and auditing compliance.
Such systems should be used in accordance with ethical AI
principles
—
fairness, transparency, and accountability
—
as prescribed by frameworks such as the OECD AI
Principles [30
] and UNESCO’s Ethics of Artificial
Intelligence Recommendation [31].
Additionally, interpretability is prioritized by courts.
Transparency and explainability are essential to ensure
that stakeholders understand and can challenge AI-
generated outcomes, particularly in high-stakes
contexts such as legal proceedings [32, 33].
Furthermore, HCI design concepts such as eXplainable AI
(XAI)
—
which focuses on making AI reasoning
intelligible
—
and accountable conversational UX should
be integrated, allowing legal professionals to verify
results with confidence [34, 35].
4.6.3
Education and Tutoring
Adaptive systems have been used in education to
personalize learning paths, with various agents
responsible for curriculum planning, engagement and
feedback. Subordinate learning agents can adapt to
students’ performance by dynamically
arranging long-
term educational goals. Effective adaptive tutoring
agents must adhere to human-centered design principles
to ensure usability, accessibility, and support for diverse
learners, as outlined in [36]. Furthermore, the IEEE 7000
series
—
especially the IEEE 7000-2021 standard on ethical
system design
—
provides a structured framework for
embedding
autonomy,
equity,
and
ethical
considerations into AI systems for children and
disadvantaged students [37].
4.6.4
Ethical and Societal Considerations
Ethical issues surrounding AMACAI systems are
becoming increasingly prominent, amplified by
challenges
such
as
distributed
responsibility,
unpredictable behavior, and moral coordination
inherent in multi-agent architectures. Ensuring system-
level explainability is critical, not only for effective
debugging but also to build trust among human
stakeholders [38, 39].
Recent HCI recommendations, including those from the
AI Now Institute and the ACM FAccT community,
emphasize human-in-the-loop control, ontological
transparency, and the importance of post-deployment
surveillance in multi-agent systems [40, 41].
Without periodic ethical testing, particularly in
simulated conditions, emergent communication or
reinforcement learning-based agents risk deviating from
acceptable norms through reward-hacking or undesired
behaviors during self-organization [38].
5
Discussion
The findings of this review support the concept of the
blistering rate of development and the growing
sophistication of Adaptive Multi-Agent Conversational
AI (AMACAI) systems.
Through the abuse of
architectural modularity, long-horizon planning, and
self-evolving
mechanisms,
these
systems
are
transforming the operation of conversational agents in
various domains. However, these developments have
had a fair share of serious trade-offs, challenges, and
other implications that are equally important to discuss
critically.
Architecture and planning are among the synergies that
have emerged in this domain. Modular multi-agent
The American Journal of Interdisciplinary Innovations and Research
119
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
systems allow the assignment of planning tasks to
specially designated modules. This enables agents to
become specialists in single tasks, such as intent
recognition and strategic goal generation, without
imposing a burden on the central controller of the system.
Such a separation of concerns establishes superior
scalability, flexibility, and performance in dynamic
conversational environments. However, this modularity
may also cause coordination problems, that is, in
instances when agents possess interdependent goals
and do not enjoy an integrated representation of the
global dialogue context.
Adaptive
agents
identify
themselves
through
interactions because the interactions are personalized
on a per-request basis. However, this flexibility in the
behavior of an agent cannot be accepted in mission-
critical applications, such as the healthcare domain or
the defense industry. Predictability and control are
paramount
in
these
fields.
In
addition,
reinforcement/meta-learning may optimize poorly
defined agents for undesirable behaviors, particularly in
open-ended environments.
This raises concerns
regarding misalignment, where the agent maximizes
incorrect goals, thereby jeopardizing safety and integrity
of the system.
Centralized and decentralized coordination also have
trade-offs. Centralized systems tend to exhibit better
global alignment and coherence, particularly in task-
based dialogue. However, they suffer from bottlenecks
and scalability issues. In contrast, decentralized systems
enable autonomy and parallelism among agents but can
provoke the destruction of dialogue streams and non-
consistent user experiences if synchronization primitives
are weak or fail.
Adaptive multi-agent systems are highly data intensive.
Their performance is typically defined by access to large
and diverse datasets or high-resolution simulations. This
makes them expensive to train and limits their
application in data-scarce environments such as space.
In addition, despite the progress made in the fields of
meta-and transfer learning, domain generalization
remains limited. Most systems are strongly optimized for
a specific set of environments and require extensive
retraining or domain adaptation to function in different
domains.
The evaluation of these models remains challenging.
Standard NLP metrics do not reflect the complexity of
multiagent dialogues, particularly in terms of the
collaboration quality, adaptivity, and long-horizon
coherence. Human evaluation is the best standard to
date; however, it is resource intensive and non-
reproducible. In addition, there is a lack of benchmarks
that explicitly focus on assessing systems with long
horizon planning and self-evolution.
The question of ethics and safety casts a large shadow
over the systems that can be developed. Emergent
behaviors can cause goal drift or manipulate the reward
signals. The danger with such systems is that they can
easily pass the boundary of user trust or ethics without
restrictions. As technologies of this sort approach the
stage of practical application, there is a need to ensure
transparent decision-making, accountability, and safe
learning processes in their use.
AMACAI systems have extensive implications for various
sectors. They can offer individual tutoring in education,
patient monitoring, and interaction in health care. They
assert that scale, awareness, and communication are
required for effective customer care and protection.
However, to keep this promise, future innovations must
be accompanied by moral responsibility, interpretability,
and good-quality criteria.
6
Conclusion and Recommendation
Conclusion
This review systematically surveys the current state-of-
the-art, foundational components, and prospective
research
directions
in
Adaptive
Multi-Agent
Conversational AI (AMACAI). The analysis foregrounds the
architectural evolution of these systems, their long-
horizon planning capabilities, and the emergence of self-
evolving
properties
within
next-generation
conversational agents.
AMACAI research signifies a paradigm shift from inflexible,
rule-based
architectures
to
contextually
aware,
dynamically adaptable agents capable of collaborative
and autonomous behavior. Recent advancements have
transformed monolithic conversational frameworks into
modular and distributed multiagent architectures,
notably leveraging transformer-based specialization,
coordination, and scalability. These developments have
markedly enhanced the robustness and task orientation
of conversational agents, enabling them to operate
The American Journal of Interdisciplinary Innovations and Research
120
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
effectively in complex, real-world environments.
Hierarchical reinforcement learning (HRL) and memory-
augmented networks have emerged as critical enablers
for sustained long-term goal management and
continuity in multi-turn dialogues. The integration of
symbolic reasoning further enhances logical consistency,
whereas meta-learning and continual learning
frameworks equip agents with the ability to generalize
from limited data and adapt continuously to novel
scenarios. Meanwhile, emergent communication has
allowed agents to develop new behavioral strategies
beyond explicit programming.
Despite these advances, several challenges persist.
Resource allocation remains a significant obstacle to
achieving resilient and scalable AMACAI deployment.
Current evaluation paradigms inadequately capture the
adaptability
and collective efficacy of multiagent
systems during extended interactions. Additionally,
safety and value alignment are becoming increasingly
pressing issues as agents exhibit behaviors that deviate
from their original design intent, sometimes resulting in
unanticipated or ethically ambiguous emergent
phenomena.
Another impediment is the absence of standardized
protocols for inter-agent communication and data
structuring,
which
restricts
interoperability,
reproducibility, and comparative benchmarking across
AMACAI implementations. These limitations collectively
highlight the necessity for further foundational work to
enable the development of robust, explainable, and
ethically aligned AMACAI systems in the future.
Recommendations
To address the identified challenges and catalyze
progress in AMACAI, the following recommendations are
proposed.
•
Establish Domain-Specific Evaluation Criteria:
Develop comprehensive, context-sensitive metrics
and benchmarks tailored to the long-horizon and
adaptive nature of multi-agent conversational
systems.
•
Foster Interdisciplinary Collaboration:
Promote
joint research efforts spanning artificial intelligence,
cognitive science, ethics, and human-computer
interaction to ensure that AMACAI systems are
socially compatible and ethically grounded.
•
Advance
Data-Efficient
Learning
Paradigms:
Leverage transfer learning and few-shot learning
strategies to reduce the data and computational
requirements of training scalable, adaptable agents.
•
Enhance Transparency and Explainability:
Design
communication structures and decision-making
frame- works that are inherently interpretable,
thereby
fostering
user
trust
and system
accountability.
•
Align with Human Values:
Integrate symbolic
reasoning and neural plasticity-inspired mechanisms
to ensure that agent behaviors align with societal
norms and user expectations.
•
The implementation and sustained advancement of
these recommendations are critical for realizing
scalable, trustworthy, and human-centric AMACAI
systems in the future.
References
1
J. Li, M. Zhang, N. Li, D. Weyns, Z. Jin, and K. Tei.
Generative ai for self-adaptive systems: State of the
art and research roadmap.
ACM Transactions on
Autonomous and Adaptive Systems
, 19(3):1
–
60,
2024.
2
Z. Tao, T.E. Lin, X. Chen, H. Li, Y. Wu, Y. Li, Z. Jin, F.
Huang, D. Tao, and J. Zhou. A survey of the self-
evolution of large language models, 2024.
3
T. Guo, X. Chen, Y. Wang, R. Chang, S. Pei, N.V.
Chawla, O. Wiest, and X. Zhang. Large language
model-based multi-agents: A survey of progress and
challenges, 2024.
4
Z. Chu, Y. Wang, F. Zhu, L. Yu, L. Li, and J. Gu.
Professional agents: Evolving large language models
into autonomous experts with human-level
competencies, 2024.
5
N. Nezamoddini and A. Gholami. A survey of adaptive
multi-agent networks and their applications in smart
cities.
6
Smart Cities
, 5(1):318
–
347, 2022.
7
C. Zhang, S. He, J. Qian, B. Li, L. Li, S. Qin, Y. Kang, M.
Ma, G. Liu, Q. Lin, and S. Rajmohan. Large language
model-brained gui agents: A survey, 2024.
The American Journal of Interdisciplinary Innovations and Research
121
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
8
Omdia. Why rigorous definitions matter in the
agentic ai conversation, 2025. Accessed: 2025-07-
02.
9
B. Liu, X. Li, J. Zhang, J. Wang, T. He, S. Hong, H. Liu, S.
Zhang, K. Song, K. Zhu, and Y. Cheng. Advances and
challenges in foundation agents: From brain-inspired
intelligence to evolutionary, collaborative, and safe
systems, 2025.
10
A. Kalyuzhnaya, V. Vorona, N. O. Nikitin, N. Chichkova,
K. Fatkhiev, S. Mityagin, Y. Aksenkin, A. Boukhanovsky,
11
K. Fedorin, E. Lutsenko, and A. Getmanov. Llm agents
for smart city management: Enhancing decision
support through multi-agent ai systems, 2025.
12
J. Liao, M. Wen, J. Wang, and W. Zhang. Marft: Multi-
agent reinforcement fine-tuning, 2025.
13
Z. Feng, R. Xue, L. Yuan, Y. Yu, N. Ding, M. Liu,
B. Gao, J. Sun, and G. Wang. Multi-agent embodied
ai: Advances and future directions, 2025.
14
G. Liu, P. Zhao, L. Liu, Y. Guo, H. Xiao, W. Lin, Y. Chai,
Y. Han, S. Ren, H. Wang, and X. Liang. Llm-powered
gui agents in phone automation: Surveying progress
and prospects, 2025.
15
Y. Du, W. Huang, D. Zheng, Z. Wang, S. Montella, M.
Lapata, K.F. Wong, and J.Z. Pan. Rethinking memory in
ai: Taxonomy, operations, topics, and future
directions, 2025.
16
S. Hu, T. Huang, G. Liu, R.R. Kompella, F. Ilhan, S.F.
Tekin, Y. Xu, Z. Yahn, and L. Liu. A survey on large
language model-based game agents, 2024.
17
Y. Liu, X. Cao, T. Chen, Y. Jiang, J. You, M. Wu, X. Wang,
M. Feng, Y. Jin, and J. Chen. A survey of embodied ai in
healthcare:
Techniques,
applications,
and
opportunities, 2025.
18
W. Liu, J. Qin, X. Huang, X. Zeng, Y. Xi, J. Lin, C. Wu, Y.
Wang, L. Shang, R. Tang, and D. Lian. The real barrier
to llm agent usability is agentic roi, 2025.
19
J. Yu, Y. Qin, H. Che, Q. Liu, X. Wang, P. Wan, D. Zhang,
and X. Liu. Position: Interactive generative video as a
next-generation game engine, 2025.
20
A. Chen, Y. Wu, J. Zhang, S. Yang, J.T. Huang, K.
Wang, W. Wang, and S. Wang. A survey on the
safety and security threats of computer-using
agents: Jarvis or ultron?, 2025.
21
Yingxuan Yang, Huacan Chai, Shuai Shao, Yuanyi
Song, Siyuan Qi, Renting Rui, and Weinan Zhang.
Agentnet: Decentralized evolutionary coordination
for llm-based multi-agent systems. arXiv preprint,
2025. Introduces a decentralized RAG-based DAG
coordination framework.
22
Ariel Flint Ashery, Andrea Baronchelli, et al.
Emergent social conventions and collective bias in
llm populations.
23
Science Advances
, 2025. Demonstrates spontaneous
emergent communication among agent populations.
24
Ali Baheri and Cecilia O. Alm. Hierarchical neuro-
symbolic decision transformer. arXiv preprint, 2025.
Combines symbolic planning with transformer-
based policies for long-horizon tasks.
25
S. Du, J. Zhao, J. Shi, Z. Xie, X. Jiang, Y. Bai, and L.
He. A survey on the optimization of large language
model-based agents, 2025.
26
J. Zheng, C. Shi, X. Cai, Q. Li, D. Zhang, C. Li, D. Yu,
and Q. Ma. Lifelong learning of large language
model-based agents: A roadmap, 2025.
27
Z. Wang, K. Wang, Q. Wang, P. Zhang, L. Li, Z. Yang,
K. Yu, M.N. Nguyen, L. Liu, E. Gottlieb, and M. Lam.
Ragen: Understanding self-evolution in llm agents
via multi-turn reinforcement learning, 2025.
28
F. Tang, H. Xu, H. Zhang, S. Chen, X. Wu, Y. Shen, W.
Zhang, G. Hou, Z. Tan, Y. Yan, and K. Song. A survey on
29
(m) llm-based gui agents, 2025.
30
Z.Z. Li, D. Zhang, M.L. Zhang, J. Zhang, Z. Liu, Y. Yao,
H. Xu, J. Zheng, P.J. Wang, X. Chen, and Y. Zhang.
From system 1 to system 2: A survey of reasoning in
large language models, 2025.
31
Y. Deng, Y. Li, B. Ding, and W. Lam. Leveraging long
short-term user preference in conversational
recommenda- tion via multi-agent reinforcement
learning, 2023.
32
L. Tudor Car, D. A. Dhinagaran, B. M. Kyaw, T.
Kowatsch, S. Joty, Y.-L. Theng, and R. Atun.
Conversational agents in health care: Scoping
review and conceptual analysis, 2020.
33
O. Fayayola, O. Olorunfemi, and P. Shoetan. Data
The American Journal of Interdisciplinary Innovations and Research
122
https://www.theamericanjournals.com/index.php/tajiir
The American Journal of Interdisciplinary Innovations and Research
privacy and security in it: A review of techniques and
challenges, 2024.
34
OECD.
Oecd
ai
principles:
Promoting
trustworthy ai. https://oecd.ai/en/ai-principles,
2024. Accessed July 2025.
35
UNESCO.
Recommendation
on
the
ethics
of
artificial
intelligence.
intelligence/recommendation-ethics,
2021.
Accessed July 2025.
36
OECD.
Transparency and explainability
(principle 1.3). https://oecd.ai/en/dashboards/
ai-principles/P7, 2024. Accessed July 2025.
37
UNESCO.
Ai and the judiciary: Balancing
innovation
with
integrity.
judiciary-balancing-innovation-integrity,
2025.
Accessed July 2025.
38
Alejandro Barredo Arrieta et al. Explainable artificial
intelligence
(xai):
Concepts,
taxonomies,
opportunities and challenges toward responsible
ai.
https://arxiv.org/abs/1910.10045,
2019.
Accessed July 2025.
39
Michael
Nasir
and
Abdul
Bamako.
Transparent justice: Legal frameworks for
explainable
ai
in
law.
2025.
Accessed July 2025.
40
ISO/TC 159/SC 4 Ergonomics of human
–
system
interaction committee.
Iso 9241-210:2019
–
ergonomics of human
–
system interaction
–
part 210:
Human-centred design for interactive systems.
International Standard, 2019. Prepared by ISO/TC
159/SC 4; withdrawal of ISO 13407.
41
IEEE. Ieee 7000-2021
–
model process for
addressing ethical concerns during system design.
IEEE Standard, 2021. Published by IEEE; authored by
committee.
42
Alan Chan, Rebecca Salganik, Alva Markelius,
Chris
Pang,
Nitarshan
Rajkumar,
Dmitrii
Krasheninnikov, Lauro Langosco, Zhonghao He,
Yawen Duan, Micah Carroll, Michelle Lin, Alex
Mayhew,
Katherine
Collins,
Maryam
Molamohammadi, John Burden, Wanru Zhao,
Shalaleh Rismani, Konstantinos Voudouris, Umang
Bhatt, Adrian Weller, David Krueger, and Tegan
Maharaj.
Harms from increasingly agentic
algorithmic systems.
arXiv
, 2023. Discusses
distributed responsibility and emergent harms in
agentic systems.
43
Stephanie Baker and Wei Xiang. Explainable ai is
responsible
ai:
How
explainability
creates
trustworthy and socially responsible artificial
intelligence.
arXiv
, 2023. Connects XAI with system-
level trust and explainability.
44
AI Now Institute. Ai now 2018 report. Report, 2018.
Recommends human-in-loop control and full-
supply-chain transparency in AI systems.
45
ACM FAccT Community. Acm facct 2025 focus areas.
Conference Position Statement, 2025. Highlights
importance of HCI, decision support, human-in-
loop, and surveillance practices.
