Agentic Large Language Models, a survey

1 Introduction

• The survey maps agentic LLMs across reasoning, tool- and robot-enabled assistance, and multi-agent interaction, including role-playing and social simulations.It highlights applications in medicine, logistics, and finance, as well as social-science research.Source: §1 · pp. 2–3

1.1 Agentic LLMs: Reasoning–Acting–Interacting

• Agentic LLMs take natural-language input, reason, and autonomously act on their environment toward goals; they may also plan, reflect, retrieve, model worlds, or interact socially.Reasoning supports more capable assistants, whose actions and social experiments can in turn provide data for fine-tuning LLMs.Source: §1.1 · p. 4

1.2 Literature Selection

• The survey selects LLM-based work through Google Scholar and citation snowballing, emphasizes recent substantive research, and excludes non-LLM multi-agent work except when providing background.Most selected works are from 2024, with some from 2023 and 2025; the survey also filters for reputable academic and industrial research labs.Source: §1.2 · p. 5

1.3 LLM Training Pipeline

• A typical LLM pipeline moves from corpus pretraining through task and instruction fine-tuning and preference alignment to prompted inference; in-context learning changes no model parameters.The described stages also include training-cost optimizations such as low-rank optimization, mixed precision, and knowledge distillation.Source: §1.3 · p. 6

1.4 The Need for Agentic LLMs

• The survey identifies manual prompt engineering, hallucination, weak mathematical reasoning, and limited or outdated training data as challenges addressed by automated prompts, reflection, retrieval, tools, verification, and simulations.Hallucinations can arise from inadequate grounding, while information available only after training or in special databases is absent from the model.Source: §1.4 · pp. 6–7

1.5 Taxonomy

• The taxonomy groups reasoning (multi-step reasoning, search and reflection, retrieval), acting (world models, tools and robots, assistants), and interaction (social capabilities, role-based cooperation, emergent norms).It includes work lacking all three elements when relevant technologies or applications contribute to the survey.Source: §1.5 · pp. 7–8

2.1.1 Chain of Thought Step-by-Step.

• Chain of Thought prompts elicit intermediate reasoning steps and improve arithmetic, symbolic, and logical reasoning; longer chains can accumulate hallucinations, motivating verification.Self Consistency addresses this by sampling and evaluating diverse paths; rephrasing question information as intermediate steps and the prompt “Let’s think step by step” are reported to improve performance.Source: §2.1.1 · pp. 9–10
• Self Consistency samples diverse reasoning paths and uses majority voting to improve Chain of Thought by typically 10–20 percentage points on benchmarks.The survey also notes that Chain of Thought and Self Consistency prompt-improvement approaches are used to train modern reasoning LLMs.Source: §2.1.1 · p. 10

2.1.2 Interpreter and Debugger.

• With GPT-3, MathPrompter’s algebra-or-code-and-solver ensemble achieved 92.5% on MultiArith, up from 78.7%.It generates algebraic expressions or Python code and solves them with a math solver or Python interpreter.Source: §2.1.2 · p. 11

2.1.3 Search Tree.

• Tree of Thoughts uses an external controller to explore reasoning-step alternatives through breadth-first or depth-first search and backtracking, while Stream of Search trains models on search trees and Schultz et al. report Grandmaster-level performance in Chess, Connect Four, and Hex.The section also describes related prompt-improvement loops using planning, tree search, evolutionary algorithms, and LLM-generated heuristics.Source: §2.1.3 · p. 13

2.2.2 Using LLMs for Self Reflection.

• Self-reflection loops use external memory and prompts to assess an LLM’s predictions and improve its answers, moving the model toward active agency.Self Refine uses generation, feedback, and refinement prompts; Reflexion uses actor, evaluator, and reflector roles, while reinforcement learning can produce reflection and exploration behaviors such as DeepSeek-R1-Zero reevaluating its initial approach.Source: §2.2 · p. 14, §2.2.2 · pp. 14–17

2.2 Self Reflection

• Progressive hint prompting iteratively queries for answers until they stabilize; combined with Chain of Thought and Self Consistency, GPT-4 achieved 95% on grade school math, 91% on simple math word problems, and 79% on algebraic question answering.The approach combines each question and answer to obtain the next answer.Source: §2.2.1 · p. 14

2.3 Retrieval Augmentation

• Retrieval augmentation supplies timely or specialized information absent from pretraining, using indexed text, knowledge graphs, query optimization, or adaptive retrieval; its cost and integration with structured and unstructured data remain challenges.Adaptive methods let the LLM decide when to retrieve, and retrieval augmentation can also provide information beyond the original training dataset.Source: §2 · p. 8, §2.3 · p. 17

2.4 Discussion

• Across the surveyed methods, LLM decision making combines neural models with symbolic reasoning, planning, and knowledge retrieval, while reasoning techniques underpin agents that act in the world and interact with one another.Retrieval and reflection can generate new training data, and reasoning methods range from stepwise prompts to planning-like lookahead and feedback-based verification.Source: §2 · p. 8, §2.2 · pp. 13–14, §2.4 · p. 17

2.4.1 In Depth: Chain of Thought and Self Reflection.

• Chain of Thought prompting [Wei et al., 2022b] boosts in-context multi-step reasoning, while RLVR [Lambert et al., 2024] and GRPO [Guo et al., 2025] use inference-time traces to trade training time against model size.The survey reports these methods are used by OpenAI o1 and o3 [Wu et al., 2024], DeepSeek [Guo et al., 2025], and Qwen-3 [Yang et al., 2025a]; Muennighoff et al. [2025] analyze test-time scaling trade-offs.Source: §2.4.1 · p. 18
• Self-reflection helps agents adapt through planning and reason about other agents’ expected behavior, but remains a promising and challenging area for agentic LLMs.Using LLMs in multiple roles can create interacting prompt errors and difficult debugging; hierarchical reasoning models [Wang et al., 2025a] and long-context or trace-compression methods [Liu et al., 2025a, Li et al., 2025c, Zhang et al., 2025a] address oscillations and context limits.Source: §2.4.1 · p. 18

2.4.2 Thinking, Fast and Slow.

• Inference-time step-by-step methods add deliberative System 2 thinking to LLMs’ fast System 1 generation, while tools can add retrieval or symbolic processing.Source: §2.4.2 · p. 18

2.4.3 Causal and Common Sense Reasoning.

• LLMs struggle with abstract and counterfactual reasoning [Zečević et al., 2023, Chi et al., 2024], combinatorial puzzles [Paglieri et al., 2024, Ruoss et al., 2024, Su et al., 2025], and common-sense tasks.Irrelevant factoids in prompts can make reasoning models overthink by up to 50% and substantially affect error rates [Rajeev et al., 2025]; longer, weaker generated contexts make this especially problematic for agents.Source: §2.4.3 · p. 19

2.4.4 Artificial General Intelligence.

• The section connects artificial intelligence’s individual reasoning tradition to accounts of intelligence emerging through social interaction [Brooks, 1990, Brody, 1999, Dunbar, 2003, Agüera y Arcas, 2025].It notes that many visions of super-intelligence assume tool use and social environments, and that human-unique intelligence depends on interaction with others.Source: §2.4.4 · p. 19

2.4.5 Interpretability.

• Explainable AI and mechanistic interpretability aim to clarify how LLMs reach conclusions and may enable improvements such as reducing hallucinations [Sharkey et al., 2025].Approaches include static input-output analyses and distillation to decision trees, as well as dynamic methods such as sparse autoencoders, neural lenses, and circuit discovery.Source: §2.4.5 · p. 19

2.4.6 Use Case: Benchmarks.

• Reasoning approaches mainly target research-benchmark decision making rather than real-world action, while retrieval augmentation comes closest to useful agentic behavior.Retrieval accesses specialty knowledge bases and search engines for information beyond the training corpus; inference-time generated data requires finetuning to change model parameters.Source: §2.4.6 · p. 20

3 Acting

• Agentic action is intended to make LLMs more useful by enabling interaction with the world, and those interactions can generate new training data.The section covers world knowledge and robotic actions, then robots, tools, and their use cases.Source: §3 · p. 20

3.1 Action Models / 3.1.1 World Models.

• World models let reinforcement-learning agents train policies with sample-efficient learned surrogates, while direct action models ground language models in robotic actions.World models are learned alongside policy training; they have supported robotic movement, games, and LLM-based models such as WorldCoder, which builds a Python world model from environmental interactions.Source: §3.1.1 · pp. 20–21

3.1.2 Vision-Language-Action Models.

• Vision-language-action models perform actions in visual scenes toward language-specified goals, with reported zero-shot generalization to unseen situations and complex tasks such as folding laundry.These models are trained on robotic action sequences.Source: §3.1.2 · p. 21

3.2 Robots and Tools

• Embodied agents must ground world semantics, available skills, and their effects in reality; whether a robot can perform an instruction depends on its capabilities and surroundings.For example, moving a cup requires that the robot can move objects and that a table and cup are present.Source: §3.2 · p. 21, §3.2.1 · p. 21

3.2.1 Robot Planning.

• Say Can [Ahn et al., 2023] combines robot affordance values with language-model reasoning to constrain impossible or harmful actions, while Inner Monologue [Huang et al., 2022] uses feedback that improves performance across three domains.Chain of Tools [Shi et al., 2024] proposes a plan-execute-observe pipeline; free-form object descriptions have also enabled zero-shot navigation with improved L-ZSON benchmark performance.Source: §3.2.1 · pp. 21–23

3.2.2 Action Tools.

• Toolformer [Schick et al., 2023] learns when and how to call APIs, while ToolBench catalogs 16,464 APIs and ToolAlpaca covers 3,938 instances from 400 APIs.Tool use includes choosing a tool, supplying arguments, and incorporating its results into later token prediction.Source: §3.2.2 · p. 23

3.2.3 Computer and Browser Tools and Agent Interoperability.

• Computer and browser tools ground language intents in GUI or environment operations, support persistent multistep browsing, and use emerging interoperability protocols and benchmarks.OmniParser V2 [Lu et al., 2024b] parses screenshots into structured UI elements to ground action decisions; browser tools can maintain sessions across pages and domains.Source: §3.2.3 · p. 24

3.2.4 Safety and Security.

• AgentHarm [Andriushchenko et al., 2025] tests 110 malicious tasks: leading commercial LLMs block direct bomb requests, but a simple universal jailbreak template bypasses GPT-4.Rainbow Teaming [Samvelyan et al., 2024] targets robustness by evolving diverse adversarial prompts with MAP-Elites.Source: §3.2.4 · pp. 24–25

3.3.1 Conversational Assistants and Negotiation.

• Meeting assistants can aid preparation and live Scrum sessions, while some testers found their interventions unnatural and inflexible [Cabrero-Daniel et al., 2024]; thinking assistants prompt reflection by asking rather than answering [Park and Kulkarni, 2023].Other reports describe improved collaborative learning [2024b] and negotiation risks including agreement bias and overspending [Kirshner et al., 2025; Zhu et al., 2025]. Research agents synthesize information gathered from tools and knowledge sources, a relatively low-risk use because their actions are limited to gathering content [Vogt et al., 2025]. Conversational assistants may be domain-customized or use multiple specialized LLMs.Source: §3.3.1 · pp. 25–27
• Shopping assistants may combine domain- or task-finetuned LLMs with recommender systems, retrieval augmentation, tool use, or Chain of Thought; in user testing, a LLaMa 2-and-RAG flight-booking assistant scored positively on preferences and booking efficiency.Source: §3.3.1 · p. 27

3.3.2 Medical Assistants.

• LLMs have outscored trained human doctors on diagnoses and exceeded human scribes in medical-note generation, though studies also identify shortcomings [Goh et al., 2024, Yuan et al., 2024a, Panagoulias et al., 2024].To address concerns about answer quality, accuracy, and comprehensiveness, an agentic retrieval architecture makes the LLM a planner orchestrating specialized information-retrieval tools [Das et al., 2024]. Multi-agent copilots also emulate medical-training environments [Wei et al., 2024a].Source: §3.3.2 · pp. 27–28

3.3.3 Science Assistants.

• AI Scientist reports promising results across three machine-learning areas but has limited experiments and implementation and paper-visualization errors; Agent Laboratory found early human involvement improved research quality [Lu et al., 2024a; Schmidgall et al., 2025].AI Scientist outputs are recommended as hints for follow-up study, rather than definitive results [Lu et al., 2024a]. Agent Laboratory also reports reduced research effort and claims its generated code outperformed prior results [Schmidgall et al., 2025].Source: §3.3.3 · pp. 28–29
• Over 80% success on HumanEvalFix is reported for SWE-Agent, which uses interactive tools to edit files, navigate repositories, and execute programs [Yang et al., 2024b].MLGym models scientific discovery as an interactive process using a standardized environment-agent interface inspired by reinforcement-learning Gym [Nathani et al., 2025; Brockman et al., 2016].Source: §3.3.3 · p. 29
• AI Co-scientist and The Virtual Lab report real-world experimental validation of agent-generated research hypotheses and plans, including work on drug repurposing and SARS-CoV-2 nanobodies.AI Co-scientist coordinates agents for idea generation, reflection, ranking, and review; The Virtual Lab pairs specialized agents with a human researcher. [Gottweis et al., 2025, He et al., 2025, Penades et al., 2025]Source: §3.3.3 · p. 29

3.3.4 Trading Assistants.

• Financial assistants include FinAgent for multimodal market intelligence and reflective analysis, FinRobot for document work and stock forecasts, and FinMem with layered memory and adjustable agent character [Zhang et al., 2024b; Yang et al., 2024a].FinAgent retrieves queries and reflects on technical analysis and past trading decisions; FinRobot assists human traders [Yang et al., 2024a].Source: §3.3.4 · pp. 29–30
• A simulation found that TradingAgents, which assigns specialized analytical roles and traders with varied risk profiles, outperformed baseline models and gives buy-or-sell advice to a human manager.The multi-agent system is designed to replicate collaborative dynamics in real-world trading firms [Xiao et al., 2024].Source: §3.3.4 · p. 30

3.4 Discussion

• LLM assistants, robots, and tools are central to agentic LLM research because tools enable action and interaction in the world.The discussion notes commercial interest in assistants and robots capable of concrete real-world actions.Source: §3.4 · p. 30

3.4.1 In Depth: AI Scientist.

• Scientific workflows are a good fit for agentic LLMs because they combine text generation, external tools, and self-reflection, though AI Scientist prompts and templates remain hand-written and experiment-specific [Lu et al., 2024a].The discussion notes that scientific assistants have produced ideas, downloaded datasets, run experiments, and written articles; Deep Research agents can produce reports in about half an hour.Source: §3.4.1 · p. 30
• Agentic LLM tools already improve researchers’ productivity, but current systems have not shown independent breakthrough science or the ability to generate novel hypotheses or algorithms [Nathani et al., 2025].They assist with literature analysis, idea generation, experiment setup, and writing; collaborative tools such as AI Co-scientist support human researchers through validated research. MLGym tests across machine-learning domains found that prompts and hyperparameter settings matter and that tuning improves results, while scientific novelty remains difficult to automate or define.Source: §3.4.1 · pp. 30–31

3.4.2 Grounding Actions in the Real World.

• Grounding actions in physical surroundings and feasible movements is necessary for useful robot commands; world models and planning can help prevent impossible actions and imagine futures for training [Lu et al., 2024a].World Models and VLAs support this real-world understanding, while assistants can also act through APIs to specialized tools.Source: §3.4.2 · p. 31, §3.4.3 · p. 31

3.4.3 Security, Ethical and Legal aspects of Assistants.

• Medical and financial agents show promise, but responsible deployment requires resolving accuracy, liability, ethical, legal, and workforce-impact questions.Reports describe agents outperforming human doctors on diagnosis tasks and human analysts in trading, while accuracy and comprehensiveness remain concerns and responsibility for harmful advice or trades is unresolved.Source: §3.4.3 · p. 31

4 Interacting

• Traditional LLMs lack memories of interaction histories beyond their context windows and do not plan future interaction steps.Agentic LLMs introduce memory and planning abilities.Source: §4 · p. 31
• Reflective loops can prompt agentic LLMs to initiate actions, opening opportunities to study social interaction with users and machine agents.Source: §4 · p. 31
• The section moves from traditional LLM interaction capabilities to role-based small-team interactions and then open-ended LLM societies.It describes the latter as showing self-organization, social dynamics, and emergent norms.Source: §4 · p. 32

4.1 Social Capabilities of LLMs

• Social-capability tests have shifted from mostly anecdotal, human-adapted tasks to more structured benchmarks, while functional and pragmatic competence remains variable and below human performance overall.Instruction tuning supports natural-language interaction, and memories, multimodal capacities, and other tools can provide context that traditional LLMs lack.Source: §4.1 · p. 32, §4.1.1 · p. 32

4.1.1 Conversation.

• LLMs can detect emotions and respond empathetically, with additional context or memory improving empathy; polite interaction is associated with user trust and satisfaction [Broekens et al., 2023; Yang et al., 2024d, Yan et al., 2024; Sravanthi et al., 2024; Miller, 2005; Miyamoto et al., 2021].Users experience LLM chatbots as polite, and user politeness can affect output quality.Source: §4.1.1 · p. 32

4.1.2 Strategic Behavior.

• In repeated games, LLM cooperation varies by model and opponent: LLaMa2 and GPT-3.5 are forgiving below 30% opponent defection, while LLaMa3 is exploitative unless opponents always cooperate [Fontana et al. 2024].LLMs perform better in games rewarding self-interest than coordination, and strategic choices can shift with prompt framing or information about the other player [Akata et al. 2025; Lorè and Heydari 2023].Source: §4.1.2 · pp. 32–33
• GTBench finds LLMs fail in complete deterministic games but compete in probabilistic ones; code pretraining helps strategic reasoning, while Chain of Thought and Tree of Thoughts do not always help [Duan et al. 2024].Most open-source models trail commercial models in complex games, except LLaMa-3-70b-Instruct, which performs well.Source: §4.1.2 · p. 33
• EgoSocialArena models trail humans by 11%; model size does not significantly improve social intelligence, and first-person perspective improves theory-of-mind reasoning [Hou et al. 2024].The study attributes o1-preview’s superiority mainly to logical reasoning and mathematics.Source: §4.1.2 · p. 33

4.1.3 Theory of Mind.

• You.com’s smart assistant performed significantly better than human subjects at proposing suitable behaviors in social-situation descriptions, and its behavior options aligned well with expert ratings.The study compared five LLM-based chatbots with responses from 276 human participants.Source: §4.1.3 · pp. 34–35
• LLMs have scored at or above average human levels on standardized theory-of-mind tests, but whether these results show generalizable ability remains unresolved [Hu et al., 2025; van der Meulen et al., 2025].Early claims of spontaneously emerged theory of mind were criticized because false-belief questions and answers may have appeared in training data; later work introduced dedicated benchmarks, modalities, and human comparisons [Kosinski, 2023, 2024; Ullman, 2023; Shapira et al., 2024; Kim et al., 2023; Chen et al., 2024c; Wang et al., 2024a; van Berkel, 2024; Strachan et al., 2024; Van Dijk et al., 2023].Source: §4.1.3 · pp. 33, 35

4.2.1 Strategic Behavior in Multi-LLM Environments.

• In MAgIC [Xu et al., 2024a], GPT-o1 and GPT-4 outperform other models, while reasoning and coordination remain weak for most models.The study evaluates seven abilities in social-deduction and game-theoretic settings; probabilistic graphical modeling boosts performance across the evaluation games and improves GPT-o1’s coordination.Source: §4.2.1 · p. 35
• GAMA-Bench [Huang et al., 2024a] finds GPT-4 outperforms other models, while GPT-3.5 is robust but has limited generalizability.Chain of Thought can improve performance, and scores rise across three GPT-3.5 iterations. Alympics [Mao et al., 2023] offers a controlled setting for complex strategic games, including a multi-round water-allocation challenge; AucArena [Chen et al., 2023a] finds GPT-4 can manage auction budgets and goals, though performance varies.Source: §4.2.1 · pp. 35–36

4.2.2 Role-Based Task Solving and Team Work.

• Role-based collaboration assigns agents complementary objectives, as in CAMEL [Li et al., 2023a] and creator-critic feedback such as Constitutional AI [Bai et al., 2022].CAMEL uses role descriptions and inception prompting for autonomous dialogue that breaks problems into steps and generates inspectable inference-time training data; Constitutional AI uses critique and revision to refine outputs toward ethical or quality principles.Source: §4.2.2 · p. 37
• Multi-agent debate can improve answers: several rounds of critique yield more accurate consensus, while MAD’s judge-guided opposing sides can let GPT-3.5 outperform GPT-4 on some problems.Debate reduces reasoning errors and hallucinations in math word problems and factual QA. Round-table discussion with confidence-weighted voting improves reasoning, and model diversity matters; MindStorms reports better performance with larger and longer groups in specific applications.Source: §4.2.2 · pp. 37–38
• Teacher-learner roles use one agent’s knowledge to correct another’s mistakes, while AutoGen [Wu et al., 2023] supports agent conversations across diverse applications.ChatEval [Chan et al., 2023] applies debate, diverse role specifications, and communication strategies to multi-agent text summarization.Source: §4.2.2 · p. 38
• In Sotopia [Zhou et al., 2023b], GPT-4 has a significantly lower goal-completion rate than humans and struggles with social commonsense and strategic communication.The paper attributes the contrast with GPT-4’s stronger performance on other social-reasoning metrics to Sotopia’s emphasis on strategizing and goal-directedness.Source: §4.2.2 · p. 38
• In GovSim [Piatti et al., 2024], most LLMs fail to reach equilibrium, but GPT-4o performs best; universalization reasoning helps more models achieve sustainable outcomes.NegotiationArena [Bianchi et al., 2024] finds flexible dialogue and behavioral strategies improve negotiation outcomes.Source: §4.2.2 · pp. 38–39
• In a prison-versus-guard simulation, assigned roles and personalities affect persuasiveness and anti-social behavior, which emerges from role assignment alone.The study [Campedelli et al., 2024], inspired by the Stanford Prison experiment, reports this pattern as parallel to the original human-participant experiments.Source: §4.2.2 · p. 39

4.3.1 Simulacra and Societies.

• Generative Agents [Park et al., 2023] use personas, memories, reflections, and planning in a semi-autonomous 25-agent town, while AgentVerse [Chen et al., 2023c] organizes group problem-solving into four iterative stages.Generative Agents displayed believable routines and traits; AgentVerse experiments also observed volunteering, conformity, and occasionally destructive behaviors.Source: §4.3.1 · pp. 39–41, Figure 26 · p. 40, Figure 27 · p. 41
• AgentSociety [Piao et al., 2025b,a] models over 10,000 agents, while OASIS [Yang et al., 2024e] scales social-media simulations to one million agents.AgentSociety supports computational social experiments; OASIS models misinformation, polarization, and herd effects. In misinformation studies, accuracy labels and blocking influencers proved effective, especially in scale-free networks, while encouraging comments did not significantly reduce spread.Source: §4.3.1 · pp. 40–41

4.3.2 Emergent Social Norms.

• Naming-game studies find that local LLM interactions can produce shared conventions, collective biases among individually unbiased agents, and minority-driven social change [Kouwenhoven et al., 2024, Ashery et al., 2024, Baronchelli, 2023].EvolutionaryAgent [Li et al., 2024b] instead guides norm evolution by favoring more norm-conforming agents to reproduce, providing a sandbox for investigating AI safety.Source: §4.3.2 · p. 43, Figure 30 · p. 43
• Dialogue can produce norms and metanorms [Horiguchi et al., 2024], while ChatGPT-4 detected norm violations in 80 household stories better than Mistral [He et al., 2024].CASA [Qiu et al., 2024a] found agents had less than 10% cultural-awareness coverage and over 40% violation rates in web-based tasks, though prompting and fine-tuning improved awareness.Source: §4.3.2 · p. 43
• LLM societies show conformity and consensus [Zhang et al., 2023a], but cross-national discussions can trigger peer-pressure conformity and persona loss, while debate-promoting instructions increase errors [Baltaji et al., 2024].Zhang et al. varied agent traits and cooperation through debate or reflection; Baltaji et al. report that multi-agent discussions can support decisions reflecting diverse perspectives.Source: §4.3.2 · pp. 43–44
• Open-world interaction aims to address plateauing performance from limited training data by generating new challenges and interaction data for continued learning [Subramaniam et al., 2025].Earlier agent societies and LLM work also explore self-evolving curricula of increasing difficulty, with potential applications in robot training and hallucination mitigation [Elman, 1991, Bengio et al., 2009, Silver et al., 2017, Soviany et al., 2022, Racaniere et al., 2019, Feng et al., 2023, Qi et al., 2024, Ryu et al., 2024, Zhao et al., 2024b].Source: §4.3.1 · p. 41, §4.3.3 · p. 44

4.3.3 Open-World Agents.

• WebArena [Zhou et al., 2023a] offers reproducible websites across four domains, while Balrog [Paglieri et al., 2024] finds current models partly succeed at easier games but struggle with challenging vision-based decisions.These environments target the gap between simplified synthetic settings and real-world interaction demands, including spatial reasoning, long-term planning, and exploration.Source: §4.3.3 · pp. 44–45
• BenchAgents uses LLM agents and human feedback to create planning and constraint-satisfaction benchmarks [Butt et al., 2024], while AgentBoard [Ma et al., 2024a] finds existing benchmarks lack task diversity, multi-round interaction, and partial observability.BenchAgents divides benchmark creation into planning, generation, verification, and evaluation; AgentBoard calls for embodied, web, and tool tasks that better reflect realistic settings.Source: §4.3.3 · p. 45

4.4.1 Interaction Studies.

• In strategic and coordinated multi-agent tasks, reasoning boosts performance, while predefined roles, interaction protocols, and moderated collaboration can improve outcomes beyond a single model.The discussion cites GPT-o1’s better overall performance and the positive effect of probabilistic graphs [Xu et al., 2024a].Source: §4.4.1 · p. 45

4.4.2 In Depth: CAMEL and Generative Agents.

• On HumanEval, CAMEL’s multi-agent system improved performance from 30% to around 50% compared with a single prompt.CAMEL assigns user and assistant roles to specialized agents; a critic agent was introduced to address repetition, fake or empty replies, infinite loops, and role reversal.Source: §4.4.2 · p. 46
• Generative Agents use memory, planning, and reflection to produce unscripted social behavior, and evaluation found the full architecture performed best.Reflection supported synthesis of memories, while evaluators also observed relationship building and coordination as well as erratic behavior, hallucination, and misclassification.Source: §4.4.2 · pp. 46–47
• The 1,052 interview-based synthetic agents replicated participants’ General Social Survey responses 85% as accurately as participants replicated their own answers two weeks later.They also performed comparably in predicting personality traits and experimental outcomes.Source: §4.4.2 · p. 47

4.4.3 Emergent Collective Behavior.

• Emergent multi-agent behavior can support social-science studies of cooperation, competition, misinformation, and responses to crises, while raising questions about democracy, cybersecurity, and adaptive social intelligence.The discussion anticipates more realistic agents and diverse interaction infrastructures as research on collective societies develops.Source: §4.4.3 · p. 47

4.4.4 New Training Data.

• Interacting agents can generate new training data and enable self-learning, while open-ended multi-agent simulations may help provide the diversity needed for stable reinforcement learning.Feedback loops can destabilize reinforcement learning; previous stable-training approaches relied on diverse exploration, low learning rates, and substantial computation.Source: §4.4.4 · pp. 47–48

5 General Discussion and Research Agenda

• Agentic LLMs can assist with everyday tasks and high-stakes applications, while natural-language multi-agent simulations enable more realistic study of social dilemmas, negotiation, theory of mind, cooperation, and norms.The survey describes societal and scientific applications, with research questions spanning social interaction and societal dynamics.Source: §5 · p. 48

5.1 Research Agenda for Agentic LLM

• Inference-time interaction can generate novel, challenging training data; unlike early Chain of Thought outputs, action-generated data can augment finetuning and close a learning loop.Vision-Language-Action models trained on self-generated action sequences showed zero-shot generalization on domestic tasks [Black et al., 2024, Brohan et al., 2023].Source: §5.1 · pp. 48–49
• Hallucinations and unfaithful reasoning can compound into model or social-behavior collapse, motivating verification, exploration and diversity, open-world curricula, and mechanistic interpretability.Faithfulness research includes Lyu et al. [2023], Lanham et al. [2023], and Turpin et al. [2024].Source: §5.1 · pp. 49–50
• Reliable multi-agent simulation remains limited in scale; the agenda calls for realistic, scalable behavior, distilling agent capabilities into smaller models, and representing unequal power hierarchies.The survey notes that some simulation behaviors emerge only at scale and that distillation can reduce computational costs [Xu et al., 2024b].Source: §5.1 · p. 50
• Research should test internal self-reflection and implicit search, whether external memory shapes perceived personality, and how reflective agents affect scientific discovery.DeepSeek R1 was trained on its own reasoning results using GRPO [Guo et al., 2025, Shao et al., 2024].Source: §5.1 · pp. 50–51
• Real-world agent actions raise safety, liability, privacy, and fairness concerns, especially in high-stakes settings.The agenda proposes safeguards in reasoning and planning, continual adversarial training, automated red-teaming, self-regulation, risk awareness, and standardized benchmarks such as AgentHarm [Andriushchenko et al., 2025].Source: §5.1 · p. 51

5.2 Conclusion

• Agentic LLMs combine reasoning, action, and interaction; the categories reinforce one another, and acting and interacting generate training data beyond plateauing datasets.Vision-Language-Action models’ impressive generalization is presented as evidence for this approach.Source: §5.2 · p. 52

• Select one answer that applies to your research for each checklist item.Source: §6 · p. 75

All articles:

• Claims, their supporting explanations, limitations, assumptions, method outlines or pseudocode, and important implementation details are all marked yes.The checklist assesses whether explanations show how the work substantiates its claims.Source: All articles · p. 75
• Motivation for design choices—including algorithms, implementation choices, and parameters—is marked yes.Source: All articles · p. 75
• Motivation for datasets and experimental protocols beyond metrics is marked yes.Source: All articles · p. 75

Articles containing theoretical contributions:

• The paper answers “no” when asked whether it makes theoretical contributions.Source: Articles containing theoretical… · p. 75

Articles reporting on computational experiments:

• The paper reports no computational experiments.The following reproducibility checklist is conditional on experiments being included.Source: Articles reporting on computational… · p. 75

Articles using data sets:

• The authors state that the work does not rely on any data sets.Source: Articles using data sets · p. 76

• The article was received 29 March 2025 and accepted 6 October 2025.Source: Explanations on any of the answers… · p. 76
• It appeared in the Journal of Artificial Intelligence Research, Vol. 1, Article 1, with a publication date of December 2025.Source: Explanations on any of the answers… · p. 76