Recent interest in Multi-Agent Systems of Large Language Models (MAS LLMs) has led to an increase in frameworks leveraging multiple LLMs to tackle complex tasks. However, much of this literature appropriates the terminology of MAS without engaging with its foundational principles. In this position paper, we highlight critical discrepancies between MAS theory and current MAS LLMs implementations, focusing on four key areas: the social aspect of agency, environment design, coordination and communication protocols, and measuring emergent behaviours. Our position is that many MAS LLMs lack multi-agent characteristics such as autonomy, social interaction, and structured environments, and often rely on oversimplified, LLM-centric architectures. The field may slow down and lose traction by revisiting problems the MAS literature has already addressed. Therefore, we systematically analyse this issue and outline associated research opportunities; we advocate for better integrating established MAS concepts and more precise terminology to avoid mischaracterisation and missed opportunities.

This is shown by a recent collaborative study between Johannes Gutenberg University Mainz (JGU) and the universities of Hamburg and Washington. The results, presented at this year's Conference on Empirical Methods in Natural Language Processing (EMNLP) - one of the world's leading conferences in computational linguistics - show that all tested AI systems reproduce social stereotypes. "Dialects are an essential part of cultural identity," emphasized Minh Duc Bui, a doctoral researcher in von der Wense's Natural Language Processing (NLP) group at JGU's Institute of Computer Science. "Our analyses suggest that language models associate dialects with negative traits - thereby perpetuating problematic social biases." Using linguistic databases containing orthographic and phonetic variants of German dialects, the team first translated seven regional varieties into Standard German.

The Price of Anarchy (PoA) is a standard metric for quantifying inefficiency in socio-technical systems, widely used to guide policies like traffic tolling. Conventional PoA analysis relies on exact numerical costs. However, in many settings, costs represent agents' preferences and may be defined only up to possibly arbitrary scaling and shifting, representing informational and modeling ambiguities. We observe that while such transformations preserve equilibrium and optimal outcomes, they change the PoA value. To resolve this issue, we rely on results from Social Choice Theory and define the Invariant PoA. By connecting admissible transformations to degrees of comparability of agents' costs, we derive the specific social welfare functions which ensure that efficiency evaluations do not depend on arbitrary rescalings or translations of individual costs. Case studies on a toy example and the Zurich network demonstrate that identical tolling strategies can lead to substantially different efficiency estimates depending on the assumed comparability. Our framework thus demonstrates that explicit axiomatic foundations are necessary in order to define efficiency metrics and to appropriately guide policy in large-scale infrastructure design robustly and effectively.

Abstract--This article presents a novel coordination and task-planning framework to enable the simultaneous conflict-free collaboration of multiple unmanned aerial vehicles (UA Vs) for aerial 3D printing. The proposed framework formulates an optimization problem that takes a construction mission divided into sub-tasks and a team of autonomous UA Vs, along with limited volume and battery. It generates an optimal mission plan comprising task assignments and scheduling, while accounting for task dependencies arising from the geometric and structural requirements of the 3D design, inter-UA V safety constraints, material usage and total flight time of each UA V. The potential conflicts occurring during the simultaneous operation of the UA Vs are addressed at a segment-level by dynamically selecting the starting time and location of each task to guarantee collision-free parallel execution. An importance prioritization is proposed to accelerate the computation by guiding the solution towards more important tasks. Additionally, a utility maximization formulation is proposed to dynamically determine the optimal number of UA Vs required for a given mission, balancing the trade-off between minimizing makespan and the deployment of excess agents. The proposed framework's effectiveness is evaluated through a Gazebo-based simulation setup, where agents are coordinated by a mission control module allocating the printing tasks based on the generated optimal scheduling plan while remaining within the material and battery constraints of each UA V. A video of the whole mission is available in the following link: https://youtu.be/b4jwhkNPT Note to Practitioners--This framework addresses the critical need for efficiency and safety in planning and scheduling multiple aerial robots for parallel aerial 3D printing. Existing approaches lack safety guarantees for UA Vs during parallel construction. This work tackles these challenges by ensuring safety during parallel operations and effectively managing task dependencies.

Influential critiques argue that Large Language Models (LLMs) are a dead end for AGI: "mere pattern matchers" structurally incapable of reasoning or planning. We argue this conclusion misidentifies the bottleneck: it confuses the ocean with the net. Pattern repositories are the necessary System-1 substrate; the missing component is a System-2 coordination layer that selects, constrains, and binds these patterns. We formalize this layer via UCCT, a theory of semantic anchoring that models reasoning as a phase transition governed by effective support (rho_d), representational mismatch (d_r), and an adaptive anchoring budget (gamma log k). Under this lens, ungrounded generation is simply an unbaited retrieval of the substrate's maximum likelihood prior, while "reasoning" emerges when anchors shift the posterior toward goal-directed constraints. We translate UCCT into architecture with MACI, a coordination stack that implements baiting (behavior-modulated debate), filtering (Socratic judging), and persistence (transactional memory). By reframing common objections as testable coordination failures, we argue that the path to AGI runs through LLMs, not around them.

The significant gap between rising demands for clinical training and the scarcity of expert instruction poses a major challenge to medical education. With powerful capabilities in personalized guidance, Large Language Models (LLMs) offer a promising solution to bridge this gap. However, current research focuses mainly on one-on-one knowledge instruction, overlooking collaborative reasoning, a key skill for students developed in teamwork like ward rounds. To this end, we develop ClinEdu, a multi-agent pedagogical simulator with personality-driven patients and diverse student cohorts, enabling controlled testing of complex pedagogical processes and scalable generation of teaching data. Based on ClinEdu, we construct ClinTeach, a large Socratic teaching dialogue dataset that captures the complexities of group instruction. We then train MedTutor-R1, the first multimodal Socratic tutor designed for one-to-many instruction in clinical medical education. MedTutor-R1 is first instruction-tuned on our ClinTeach dataset and then optimized with reinforcement learning, using rewards derived from a three-axis rubric, covering structural fidelity, analytical quality, and clinical safety, to refine its adaptive Socratic strategies. For authentic in-situ assessment, we use simulation-based interactive evaluation that redeploys the tutor back into ClinEdu. Experimental results demonstrate that our MedTutor-R1 outperforms the base model by over 20% in average pedagogical score and is comparable to o3, while also exhibiting high adaptability in handling a varying number of students. This promising performance underscores the effectiveness of our pedagogical simulator, ClinEdu.

Current clinical agent built on small LLMs, such as TxAgent suffer from a \textit{Context Utilization Failure}, where models successfully retrieve biomedical evidence due to supervised finetuning but fail to ground their diagnosis in that information. In this work, we propose the Executor-Analyst Framework, a modular architecture that decouples the syntactic precision of tool execution from the semantic robustness of clinical reasoning. By orchestrating specialized TxAgents (Executors) with long-context foundation models (Analysts), we mitigate the reasoning deficits observed in monolithic models. Beyond simple modularity, we demonstrate that a Stratified Ensemble strategy significantly outperforms global pooling by preserving evidentiary diversity, effectively addressing the information bottleneck. Furthermore, our stress tests reveal critical scaling insights: (1) a \textit{Context-Performance Paradox}, where extending reasoning contexts beyond 12k tokens introduces noise that degrades accuracy; and (2) the \textit{Curse of Dimensionality} in action spaces, where expanding toolsets necessitates hierarchical retrieval strategies. Crucially, our approach underscores the potential of training-free architectural engineering, achieving state-of-the-art performance on CURE-Bench without the need for expensive end-to-end finetuning. This provides a scalable, agile foundation for the next generation of trustworthy AI-driven therapeutics. Code has been released on https://github.com/June01/CureAgent.

Quantitative Systems Pharmacology (QSP) modeling is essential for drug development but it requires significant time investment that limits the throughput of domain experts. We present \textbf{GRASP} -- a multi-agent, graph-reasoning framework with a human-in-the-loop conversational interface -- that encodes QSP models as typed biological knowledge graphs and compiles them to executable MATLAB/SimBiology code while preserving units, mass balance, and physiological constraints. A two-phase workflow -- \textsc{Understanding} (graph reconstruction of legacy code) and \textsc{Action} (constraint-checked, language-driven modification) -- is orchestrated by a state machine with iterative validation. GRASP performs breadth-first parameter-alignment around new entities to surface dependent quantities and propose biologically plausible defaults, and it runs automatic execution/diagnostics until convergence. In head-to-head evaluations using LLM-as-judge, GRASP outperforms SME-guided CoT and ToT baselines across biological plausibility, mathematical correctness, structural fidelity, and code quality (\(\approx\)9--10/10 vs.\ 5--7/10). BFS alignment achieves F1 = 0.95 for dependency discovery, units, and range. These results demonstrate that graph-structured, agentic workflows can make QSP model development both accessible and rigorous, enabling domain experts to specify mechanisms in natural language without sacrificing biomedical fidelity.

Large language models display a peculiar form of inconsistency: they "know" the correct answer but fail to act on it. In human philosophy, this tension between global judgment and local impulse is called akrasia, or weakness of will. We propose akrasia as a foundational concept for analyzing inconsistency and goal drift in agentic AI systems. To operationalize it, we introduce a preliminary version of the Akrasia Benchmark, currently a structured set of prompting conditions (Baseline [B], Synonym [S], Temporal [T], and Temptation [X]) that measures when a model's local response contradicts its own prior commitments. The benchmark enables quantitative comparison of "self-control" across model families, decoding strategies, and temptation types. Beyond single-model evaluation, we outline how micro-level akrasia may compound into macro-level instability in multi-agent systems that may be interpreted as "scheming" or deliberate misalignment. By reframing inconsistency as weakness of will, this work connects agentic behavior to classical theories of agency and provides an empirical bridge between philosophy, psychology, and the emerging science of agentic AI.

Recent progress in leveraging large language models (LLMs) has enabled Neural Architecture Design (NAD) systems to generate new architecture not limited from manually predefined search space. Nevertheless, LLM-driven generation remains challenging: the token-level design loop is discrete and non-differentiable, preventing feedback from smoothly guiding architectural improvement. These methods, in turn, commonly suffer from mode collapse into redundant structures or drift toward infeasible designs when constructive reasoning is not well grounded. We introduce RevoNAD, a reflective evolutionary orchestrator that effectively bridges LLM-based reasoning with feedback-aligned architectural search. First, RevoNAD presents a Multi-round Multi-expert Consensus to transfer isolated design rules into meaningful architectural clues. Then, Adaptive Reflective Exploration adjusts the degree of exploration leveraging reward variance; it explores when feedback is uncertain and refines when stability is reached. Finally, Pareto-guided Evolutionary Selection effectively promotes architectures that jointly optimize accuracy, efficiency, latency, confidence, and structural diversity. Across CIFAR10, CIFAR100, ImageNet16-120, COCO-5K, and Cityscape, RevoNAD achieves state-of-the-art performance. Ablation and transfer studies further validate the effectiveness of RevoNAD in allowing practically reliable, and deployable neural architecture design.