The paper investigates the macroscopic dynamics of large language model (LLM)-driven agents by applying a least‑action principle to estimate the generative directionality of LLMs embedded within agents. Through experimental measurement of transition probabilities between LLM‑generated states, the authors find evidence of a tendency toward detailed balance, suggesting that LLM generation may be governed by underlying potential functions rather than explicit learning rules. This approach offers a model‑agnostic, effective description of LLM generative dynamics that could serve as a foundational theory for complex AI systems.
arXiv:2608. 11215v1 Announce Type: new Abstract: Simulating societies of many large language model (LLM) agents is expensive, yet the questions asked of such simulations are usually macroscopic: phase behaviour, stylised facts, and scaling with the number of agents $N$, not the cognition of any single agent.
arXiv:2611.07065v1 Announce Type: new Abstract: This paper develops a stochastic dynamical-systems perspective on three classes of generative AI models. First, we formulate large language models (LLM...
arXiv:2610.03033v1 Announce Type: new Abstract: Large language model (LLM)-based agents increasingly operate in multi-agent systems (MAS) characterised by strategic interaction. However, little is kn...
The paper introduces Markovian generation chains, an iterative inference process in large language models where each step uses a prompt template and the previous output as input, excluding prior data. Experiments in iterative rephrasing and round‑trip translation show that outputs either converge to a small recurrent set or continue generating novel sentences within a finite horizon. Sentence‑level Markov chain modeling reveals that the process can either increase or reduce diversity, depending on factors such as temperature and the initial input sentence.
The paper shows that classical consensus theory, which relies on network topology and convex update rules, fails for large language models (LLMs). By measuring token‑level updates across eight instruction‑tuned models and 73,080 exchanges, the authors find that LLMs violate convexity, exhibit negative weights, and are driven by internal update bias rather than network structure. Simulations with 10,000 agents confirm that algebraic connectivity explains almost none of the outcome variance, leading populations to converge on confidently false claims regardless of their network topology.
arXiv:2510.15047v2 Announce Type: replace-cross Abstract: Large Language Models (LLMs) as agents often fail to improve in new environments. We identify and characterize a failure mode we call explora...
arXiv:2602. 21556v2 Announce Type: replace Abstract: When designing compound AI systems, a common approach is to query multiple copies of the same model and aggregate the responses to produce a synthesized output.
The paper introduces VHD-Play, a pipeline that first samples and solves a mathematical model before generating agentic reinforcement learning environments, ensuring that dynamics and evaluation are aligned from the outset. This approach yields 3,300 diverse environments at a low cost and significantly improves the performance of a large language‑model agent (Qwen3.6‑35B‑A3B) across multiple diagnostic families and external benchmarks. The study demonstrates that stateful interaction is a key factor in learning gains and that scaling the training substrate can further enhance performance.
The paper introduces the Universe of Universes (UoU) framework, treating the ecosystem of major large language models as a structured retrieval corpus and proposing a compositional Automated Reasoning and Machine Learning architecture for cross-model retrieval‑augmented generation. It formally defines the Benefit Yield Function (BYF), measuring marginal performance gain per added model, and identifies an implosion threshold θ* where BYF becomes zero and ensemble performance degrades. The work highlights gaps in current LLM ensemble research, such as lack of performance analysis across full model universes, and connects these findings to implications for DoD AI acquisition policy and testing of AI‑enabled systems.
The paper investigates how transformer models learn latent structure by training a small decoder-only transformer on three variants of the Alchemy benchmark. It finds that the model acquires different components of latent structure in discrete stages, with a notable asymmetry: it robustly composes fundamental transitions but struggles to decompose complex examples into atomic transitions. Layer‑specific causal interventions reveal plasticity windows where freezing layers delays or prevents stage completion, offering a detailed view of capability evolution during training.
arXiv:2607. 28667v1 Announce Type: cross Abstract: Recent work has shown that classifying large language models (LLMs)' responses can be distinguished by modeling token embeddings as trajectories of a black-box dynamical system (DS) and comparing prediction residuals of two DSs.
Language-model agents increasingly face long-horizon tasks with evolving state, interdependent decisions, and delayed outcomes. Scaling their training requires diverse agentic environments, dependable...