PaperlayerView exact revision on arXiv

Source-linked AI summary

From LLM Reasoning to Autonomous AI Agents: A Comprehensive Review

Mohamed Amine Ferrag, Norbert Tihanyi, Merouane Debbah

arXiv:2504.19678v2cs.AIcs.LG

TL;DR

Research on LLM reasoning and autonomous agents remains fragmented across benchmarks, frameworks, applications, and protocols. This survey unifies these areas through benchmark comparison and taxonomy, framework and application reviews, and protocol analysis, reporting a broad synthesis that includes a 90% human-judgment alignment result for Agent-as-a-Judge.

  • Problem

    The field lacks a unified taxonomy and comprehensive survey of rapidly expanding benchmarks, agent frameworks, applications, and collaboration protocols.

  • Method

    The paper compares benchmarks from 2019–2025, proposes a taxonomy of approximately 60 benchmarks, reviews frameworks and applications, and surveys ACP, MCP, and A2A protocols.

  • Results

    90% alignment with human judgments was achieved by Agent-as-a-Judge, compared with the conventional LLM-as-a-Judge approach’s typical 60–70% alignment.

  • Takeaways & Limitations

    The survey provides an integrated reference spanning evaluation, autonomous-agent design, applications, collaboration protocols, and future research directions.

Abstract

from arXiv · show

Large language models and autonomous AI agents have evolved rapidly, resulting in a diverse array of evaluation benchmarks, frameworks, and collaboration protocols. Driven by the growing need for standardized evaluation and integration, we systematically consolidate these fragmented efforts into a unified framework. However, the landscape remains fragmented and lacks a unified taxonomy or comprehensive survey. Therefore, we present a side-by-side comparison of benchmarks developed between 2019 and 2025 that evaluate these models and agents across multiple domains. In addition, we propose a taxonomy of approximately 60 benchmarks that cover general and academic knowledge reasoning, mathematical problem-solving, code generation and software engineering, factual grounding and retrieval, domain-specific evaluations, multimodal and embodied tasks, task orchestration, and interactive assessments. Furthermore, we review AI-agent frameworks introduced between 2023 and 2025 that integrate large language models with modular toolkits to enable autonomous decision-making and multi-step reasoning. Moreover, we present real-world applications of autonomous AI agents in materials science, biomedical research, academic ideation, software engineering, synthetic data generation, chemical reasoning, mathematical problem-solving, geographic information systems, multimedia, healthcare, and finance. We then survey key agent-to-agent collaboration protocols, namely the Agent Communication Protocol (ACP), the Model Context Protocol (MCP), and the Agent-to-Agent Protocol (A2A). Finally, we discuss recommendations for future research, focusing on advanced reasoning strategies, failure modes in multi-agent LLM systems, automated scientific discovery, dynamic tool integration via reinforcement learning, integrated search capabilities, and security vulnerabilities in agent protocols.

I. INTRODUCTION

The survey addresses fragmentation in evaluations, agent frameworks, applications, and collaboration protocols by organizing these developments into a unified treatment. It compares benchmarks, proposes a taxonomy, reviews frameworks and applications, and identifies future research directions.

  • Benchmarks and taxonomy: The survey compares state-of-the-art LLM and autonomous-agent benchmarks developed between 2019 and 2025 across multiple domains.
  • Benchmarks and taxonomy: Its taxonomy covers approximately 60 benchmarks spanning knowledge reasoning, mathematics, software engineering, retrieval, domain-specific, multimodal, orchestration, and interactive tasks.
  • Agent frameworks: The paper reviews agent frameworks that combine LLMs with modular toolkits for autonomous decision-making and multi-step reasoning.
  • Applications and protocols: It surveys applications across science, engineering, healthcare, finance, multimedia, synthetic data generation, and geographic information systems.
  • Applications and protocols: The survey covers ACP, MCP, and A2A as protocols for agent-to-agent collaboration.
  • Future directions: It recommends research on advanced reasoning, multi-agent failures, scientific discovery, reinforcement-learning-based tool integration, search, and protocol security.

C. Multi-Agent Systems

Multi-agent systems extend LLM capabilities through communication, cooperation, competition, and specialized roles across digital, social, graphical, personal, scientific, and healthcare settings. Existing surveys emphasize both their applications and challenges in scalability, security, multimodal integration, planning, and reasoning.

  • Multi-agent systems: Multi-agent systems use communication among specialized agents to support cooperative and competitive interactions on tasks difficult for individual agents.
  • Multi-agent systems: Surveyed challenges include scalability, security, and multimodal integration in multi-agent architectures.
  • Applications: Healthcare-agent research covers clinical decision support, medical documentation, training simulations, and healthcare-service optimization.
  • Applications: Social-agent surveys examine preferences, beliefs, reasoning, game frameworks, communication scenarios, and game-specific or game-agnostic metrics.
  • Applications: GUI-agent research describes multimodal systems that interpret graphical interfaces and execute multi-step tasks from conversational commands.
  • Applications: Personal LLM agents integrate personal data and devices to improve assistance while addressing traditional assistants’ limitations in intent understanding, planning, and tool use.
  • Scientific applications: Agentic AI research applications include literature review, hypothesis generation, experimental design, and data analysis across chemistry, biology, and materials science.

E. Comparison with Our Survey

The survey unifies fragmented coverage of benchmarks, frameworks, applications, protocols, and open challenges, while comparing demanding evaluations across multimodal reasoning, broad knowledge, function calling, academic expertise, grounding, process verification, and document extraction.

  • Survey scope: The survey integrates benchmarks, agent frameworks, applications, protocols, and open problems that prior studies typically treat separately.It presents this integration as a unified treatment and roadmap for LLM-based autonomous AI agents.
  • Multimodal reasoning: 1,184 puzzles make ENIGMAEVAL a multimodal, long-context test of visual-semantic integration and multi-step deductive reasoning.State-of-the-art systems achieve only about 7% accuracy on standard puzzles and fail on the hardest ones.
  • General knowledge reasoning: 57 tasks make MMLU a broad evaluation of world knowledge and problem-solving across subjects including mathematics and professional law.The benchmark emphasizes zero-shot and few-shot generalization without task-specific fine-tuning.
  • Task orchestration: ComplexFuncBench evaluates multi-step function calling under constraints, implicit parameters, long inputs, and contexts reaching 128k tokens.Its ComplexEval framework assesses more than 1,000 scenarios across five function-calling aspects; observed failures include value errors and premature termination.
  • Academic reasoning: 3,000 questions across more than 100 subjects make Humanity’s Last Exam substantially harder than traditional benchmarks, with leading models achieving less than 10% accuracy.The questions are contributed by nearly 1,000 experts from over 500 institutions and are multimodal and resistant to quick internet retrieval.
  • Grounding and structured evaluation: The benchmark landscape also covers factual grounding, reasoning-error detection, document extraction, and tabulated comparisons of LLM evaluations.FACTS Grounding uses 1,719 long-form grounding examples, ProcessBench uses 3,400 annotated mathematical cases, and OmniDocBench spans nine document types with 19 layout categories and 14 attribute labels.

H. Agent-as-a-Judge

Agent-as-a-Judge evaluates agentic systems with granular intermediate feedback rather than relying only on final outcomes or extensive manual assessment. On realistic software-development tasks, it achieves high human alignment while substantially reducing evaluation cost and time.

  • Agent-as-a-Judge uses agentic systems to evaluate other agentic systems and provide granular feedback throughout task solving.
  • DevAI contains 55 realistic automated AI development tasks annotated with 365 hierarchical user requirements for evaluating the framework.
  • 90% human-judgment alignment exceeds the 60–70% typically achieved by conventional LLM-as-a-Judge methods.
  • $30.58 versus $1,297.50 reduces evaluation cost to approximately 2.29%, while evaluation time falls to 118.43 minutes from 86.5 hours.

Q. OCCULT Benchmark

The review surveys benchmarks that extend evaluation from core reasoning and knowledge toward dynamic, multimodal, interactive, and agentic capabilities. These benchmarks expose persistent gaps in general-purpose reasoning, strategic planning, multi-agent coordination, and real-world tool use.

  • Dynamic reasoning: DIA evaluates problem-solving across mathematics, cryptography, cybersecurity, and computer science using dynamic questions rather than static question-answer pairs.Its metrics assess reliability and confidence across multiple attempts.
  • Multi-agent evaluation: MultiAgentBench evaluates LLM-powered multi-agent systems across six interactive domains, including research writing, coding, gameplay, database analysis, and bargaining.The benchmark targets capabilities that single-agent or narrow-domain evaluations do not capture.
  • General-purpose evaluation: 15% versus 92%: GPT-4 with plugins achieved 15% on GAIA’s 466 real-world questions, while humans achieved 92%.GAIA evaluates reasoning, multimodality, web browsing, and tool use on conceptually simple questions.
  • Recent expansion: Recent benchmarks broaden coverage to complex puzzles, function calling, medical reasoning, expert academic tasks, cybersecurity, and freelance software engineering.The review identifies this expansion as a substantial increase in evaluation depth and breadth.
  • Benchmark landscape: The benchmark taxonomy spans academic and general knowledge, mathematics, code, factual grounding, domain-specific, multimodal, task-selection, and agentic evaluations.Fig. 2 organizes the evaluation landscape across these categories for AI-agent settings.

IV. AI AGENTS

AI-agent frameworks combine large language models with modular tools, reusable components, and iterative workflows to support autonomous decision-making. Agentic workflows and Agentic RAG add planning, execution, evaluation, and dynamic retrieval to conventional model responses.

  • Frameworks: AI-agent frameworks integrate LLMs with modular tools for context management, tool integration, iterative refinement, autonomous decision-making, and multi-step reasoning.The reviewed frameworks include LangChain, LlamaIndex, CrewAI, and Swarm.
  • Agentic workflows: Agentic workflows replace fixed rules with a cycle in which agents formulate strategies, execute tasks using tools, and evaluate outcomes.This makes the process dynamic and adaptive rather than deterministic.
  • RAG applications: Retrieval-augmented generation is described as effective for up-to-date, accurate responses in domains such as healthcare and law.These domains require precise, domain-specific information.
  • Agentic RAG: Agentic RAG combines fact-based grounding from retrieval-augmented generation with agents’ dynamic adaptability and autonomous decision-making.The combination is described as minimizing errors while keeping responses current.
  • LangChain: LangChain agents select appropriate tools for subtasks and iteratively refine responses within workflows built from language models, search integrations, and utility functions.The framework supports custom roles, tasks, and tool sets.

2) LlamaIndex:

LlamaIndex provides a flexible platform for building autonomous agents that reason over tasks and iteratively call external tools. The surrounding frameworks and applications extend this tool-using paradigm to multi-agent coordination, GUI control, reasoning assistance, and domain-specific systems.

  • LlamaIndex: LlamaIndex agents receive a task and tools, select tools iteratively, and loop through reasoning and execution until generating a final answer.Tools can range from simple Python functions to complete query engines.
  • Multi-agent frameworks: CrewAI orchestrates autonomous teams whose specialized agents collaborate through roles, tools, objectives, processes, and assigned tasks.Its components coordinate individual assignments toward a larger goal.
  • Multi-agent frameworks: Swarm uses a stateless loop of agent interactions, function calls, and dynamic handoffs to manage multi-agent systems with fine-grained control.Context variables maintain continuity and information sharing across the conversation.
  • GUI agents: Claude 3.5 Computer Use completed 16 of 20 test cases translating instructions and screenshots into desktop actions.The evaluation covered web search, productivity, gaming, and file-management tasks.
  • Tool-using reasoning: OctoTools improved average accuracy by 9.3% over GPT-4o and by up to 10.6% over other frameworks using the same toolset.Its planner, standardized tool cards, and executor support tool integration and multi-step problem solving.

7) Agents SDK:

The reviewed agent systems apply specialized, collaborative architectures to diagnosis, therapy, patient simulation, and medical machine-learning workflows. Across these applications, agents emphasize interpretability, iterative reasoning, multimodal analysis, and domain-specific coordination.

  • Clinical diagnosis: DiagnosisGPT uses Chain-of-Diagnosis to provide transparent step-by-step reasoning and disease-confidence distributions while diagnosing 9,604 diseases.The approach is reported to improve diagnostic rigor and outperform existing LLMs on diagnostic benchmarks.
  • Medical ML workflows: M3Builder coordinates four agents for medical-imaging data processing, environment configuration, auto-debugging, and model training.It achieved a 94.29% success rate with Claude-3.7-Sonnet across 14 training datasets.
  • Clinical diagnosis: PathFinder uses Triage, Navigation, Description, and Diagnosis Agents to collaboratively interpret gigapixel whole-slide images.Risky slides trigger iterative navigation and characterization of significant regions.
  • Clinical training: EvoPatient generates simulated patients for multi-turn diagnostic dialogues in which patient and doctor agents engage in iterative consultations.The framework emphasizes unsupervised simulation of standardized patient-presentation patterns.
  • Mental health: PsyDraw uses specialized agents to extract drawing features and generate professional reports for mental-health screening from 290 primary-school submissions.It targets early screening of left-behind children in rural China.
  • Mental health: HamRaz is a Persian-language Person-Centered Therapy dataset whose trained models produced more empathetic, contextually nuanced, and realistic counseling interactions.The comparisons included Script Mode and Two-Agent Mode approaches.

2) Materials Science:

LLM-based agents are being applied to automate astronomical observation workflows and address specialized materials-science research challenges. These systems combine domain knowledge, curated resources, and tools to improve research efficiency and reliability.

  • Materials-science applications aim to streamline complex scientific workflows, data collection, analysis, and specialized research tasks.
  • StarWhisper automates observation-list generation, telescope operations, real-time image analysis, and follow-up proposal formulation for astronomical surveys.
  • HoneyComb targets materials-science research by addressing complex computational tasks and outdated implicit knowledge that can cause inaccuracies and hallucinations.

3) Biomedical Science:

The reviewed applications span biomedical knowledge discovery, collaborative research, scientific ideation, data-science workflows, scientific discovery, and software engineering. They use specialized agents, structured literature, multi-agent coordination, and executable environments to support complex research and development tasks.

  • Biomedical Science: GeneAgent combines self-verification, biological databases, and specialized domain knowledge for gene-set knowledge discovery.
  • Biomedical Science: BioKGBench evaluates biomedical AI scientist agents through scientific-claim verification and interaction with research literature.
  • Academic Ideation: AgentRxiv enables autonomous agent laboratories to share research reports through a common preprint server and build iteratively on prior findings.
  • Academic Ideation: The CoI agent structures literature into progressive chains and achieves idea quality comparable to human experts at approximately $0.50 per candidate idea and experimental design.
  • Scientific Discovery: AI co-scientist uses seven specialized agents to generate, refine, rank, and organize scientific hypotheses.
  • Software Engineering: 86.0% success on 420 Python repositories made Repo2Run outperform existing baselines by 63.9% in automated environment configuration.

6) Synthetic data generation:

The reviewed systems apply multi-agent methods to synthetic-data generation and finance, combining automated instruction creation, financial-data integration, market simulation, investment analysis, and trustworthy attribution. Reported applications range from model post-training to risk-sensitive financial decision support.

  • Synthetic Data Generation: AgentInstruct automates high-quality instructional-data creation through a multi-agent Generative Teaching workflow.
  • Synthetic Data Generation: 25 million prompt-response pairs from AgentInstruct improved Orca-3 performance by 19% to 54% across several benchmarks and reduced summarization hallucinations.
  • Finance Applications: FinCon uses a manager–analyst communication hierarchy to coordinate sequential investment decisions in volatile environments.
  • Finance Applications: LLM agents in multi-commodity Cournot markets can dynamically adjust pricing and resource allocation to monopolize commodities and maximize profitability.
  • Finance Applications: MarketSenseAI integrates news, prices, fundamentals, macroeconomic indicators, filings, earnings calls, and institutional reports for stock analysis.
  • Finance Applications: 125.9% cumulative returns versus 73.5% for the index were reported for S&P 100 stocks during 2023–2024.

8) Chemical Reasoning:

The surveyed work addresses chemical and mathematical reasoning through domain-specific agents, structured memories, efficient architectures, collaborative inference, and formal proof support. Applications include chemistry, materials discovery, tutoring, numerical reasoning, and theorem proving.

  • Chemical Reasoning: Tox-chat uses hierarchical section search and distilled tool-use capabilities to support efficient Korean chemistry agents under resource constraints.
  • Chemical Reasoning: ChemAgent decomposes chemical tasks into sub-tasks and stores reusable knowledge in a dynamic, self-updating structured memory library.
  • Chemical Reasoning: Materials-discovery agents use expert-curated goals, constraints, and methodologies alongside a scalable metric for evaluating generated hypotheses.
  • Mathematical Problem Solving: MACM increased GPT-4 Turbo accuracy on level-five MATH problems from 54.68% to 76.73%.
  • Mathematical Problem Solving: Inductive-reasoning agents improved global mathematical accuracy by 20.96% and solved 17.54% of problems missed by the baseline.
  • Mathematical Problem Solving: Combining varied prompting methods achieved maximal MATH-hard search-space utilization with approximately 43% fewer inference runs.

10) Geography Applications:

The survey presents autonomous-agent applications across geography and multimedia, spanning spatial analysis, film, video, drama, fashion, music, poetry, and lyrics. These systems use modular or collaborative agents to address domain-specific reasoning and generation challenges.

  • Geography Applications: MineAgent applies a modular multimodal-agent framework to remote-sensing mineral exploration, addressing geological knowledge, multi-image reasoning, and long-context challenges.Experiments demonstrate its effectiveness and potential in remote-sensing mineral exploration.
  • Geography Applications: An autonomous GIS-agent framework generates, executes, and debugs programs to discover and retrieve geospatial data from predefined sources.Source-specific handbooks guide the selection of appropriate data sources.
  • Multimedia Applications: Multimedia agents coordinate text, audio, image, and video capabilities across film production, story-to-video generation, drama scripting, fashion assistance, and music-related tasks.The survey organizes these applications as an emerging frontier for creative and interpretive agent systems.
  • Multimedia Applications: IBSEN uses a director agent and actor agents to maintain plot progression and improve control over drama-script development.The director writes plot outlines, directs role-play, and adjusts the plot toward the intended objective.
  • Multimedia Applications: ComposerX generates coherent polyphonic compositions with engaging melodies and follows user instructions, improving LLM-based creative music composition.The framework targets long-dependency difficulties in music composition.
  • Multimedia Applications: Multi-agent social learning improves poetry diversity by 3.0–3.7 percentage points and novelty by 5.6–11.3 percentage points in experiments on 96k poems.Non-cooperative interaction particularly enhances diversity for prompting-based agents.

C. AI Agents Protocols

The survey compares ACP, MCP, and A2A as protocols for standardizing interactions among agents, tools, data sources, and services. Their designs target interoperability, context sharing, and collaborative workflows across heterogeneous systems.

  • ACP: ACP standardizes communication among agents and supports framework-independent interaction across diverse agents and client systems.ACP is central to IBM Research’s BeeAI platform and uses session-based stateful communication.
  • MCP: MCP standardizes how AI systems connect to external tools and data sources, enabling agents to identify, select, and manage services contextually.Its ecosystem includes integrations for files, databases, development tools, productivity platforms, and communication services.
  • MCP: MCP uses a client-server architecture to unify secure access to local data and remote web services through structured interfaces.Host applications connect to multiple lightweight servers that expose resources and tools.
  • A2A: A2A enables dynamic collaboration among autonomous agents across isolated data systems, applications, frameworks, and vendors.It builds on HTTP, SSE, and JSON-RPC to simplify interoperability.
  • A2A: A2A groups interactions with shared task identifiers, streams context updates through SSE, and supports resubscription for recovering ongoing tasks.The protocol also defines standard JSON-RPC error handling.
  • Protocol Comparison: The survey’s comparison distinguishes MCP for tool and data integration, ACP for local-first multi-agent communication, and A2A for agent collaboration.These protocols differ in primary purpose, setup, core features, and intended use cases.

D. Training datasets

The survey reviews datasets designed to expand reasoning, multilingual coverage, synthetic instruction generation, and collaborative-agent training. It also highlights human oversight and structured communication as complementary supports for reliable agent development.

  • Training Datasets: NaturalReasoning contains 2.8 million cross-domain questions with reference answers for expanding reasoning beyond traditional mathematics and coding.Its domains include STEM, economics, and social sciences.
  • Training Datasets: FineWeb2 provides 8TB of cleaned multilingual text containing over 3 trillion non-English words across 1,893 languages.The corpus is drawn from 96 CommonCrawl snapshots spanning 2013 to 2024.
  • Training Datasets: MagPie-Ultra is an open synthetic dataset of 50,000 instruction pairs generated with Meta Llama 3.1 405 B-Instruct FP8.Its challenging tasks include coding, mathematics, data analysis, creative writing, advice seeking, and brainstorming.
  • Training Datasets: MagPie-Ultra adds metadata for quality, difficulty, embeddings, topics, and safety assessments to support training and evaluation.Safety assessments use tools including ArmorRM and LlamaGuard.
  • Human Oversight: Human-in-the-loop mechanisms insert confirmation steps between agent decisions and sensitive tool execution in business-critical settings.They address risks from incorrect tool selection or parameter interpretation.
  • Collaborative Agent Methods: TalkHier combines structured messaging with hierarchical refinement, using raw input, intermediate output, and background context to organize inter-agent information flow.Experiments report improved performance on subdomain problem solving, open question answering, and Japanese text generation.
  • Collaborative Agent Methods: CoMM assigns distinct expert roles and reasoning trajectories to multiple agents, enabling richer problem-space exploration than single-chain prompts.Few-shot exemplars are distributed across role-play agents.

V. CHALLENGES AND OPEN PROBLEMS

The survey identifies open problems in reasoning, multi-agent reliability, scientific discovery, tool integration, search, and protocol security. It pairs these challenges with approaches including Meta-CoT, failure-mode analysis, Chain-of-Tools, and reinforcement-learning-based search.

  • Advanced Reasoning: Meta-CoT addresses Chain-of-Thought limitations by modeling latent reasoning and the in-context search underlying visible reasoning steps.The approach aims to formalize cognitive processes not explicitly represented by traditional CoT.
  • Multi-Agent Failure Modes: A study of five open-source frameworks across 150 tasks identified fourteen multi-agent failure modes involving design, inter-agent alignment, verification, and termination.Examples include ignored role specifications, unnecessary repetition, memory lapses, and flawed verification.
  • Automated Scientific Discovery: Scientific-discovery agents are evaluated on inspiration retrieval, hypothesis composition, and hypothesis ranking across research domains.The benchmark extracts research questions, backgrounds, inspirations, and hypotheses from scientific papers.
  • Automated Scientific Discovery: Scientific-discovery systems remain constrained by risks of biased or spurious associations, domain complexity, data contamination, and demands for transparency and ethical standards.These constraints complicate generating novel, scientifically valid hypotheses without sufficient human oversight.
  • Dynamic Tool Integration: Chain-of-Tools enables dynamic selection and use of external tools, and experiments report performance above conventional baselines across several datasets.Evaluations include SimpleToolQuestions, GSM8K-XL, FuncQA, and KAMEL.
  • Autonomous Search: ReSearch integrates search into reasoning through reinforcement learning, treating search as an actionable tokenized operation guided by answer accuracy and structured behavior.The method trains models such as Qwen2.5 to decide when and how to consult external knowledge.
  • Protocol Security: MCP security risks include inconsistent defenses and missing standardized authentication, while distributed workflows can produce state inconsistencies and operational glitches.These weaknesses can increase unauthorized-access and data-breach risks.
  • Conclusion: The survey concludes that multi-step intermediate processing improves accuracy and reliability across complex mathematics, code-generation, and logical-reasoning tasks.It also synthesizes reasoning strategies and compares approximately 60 benchmarks from 2019 to 2025.
Loading 2504.19678v2…