Collective human activities like using an Ouija board (or Kokkuri-san) often produce emergent, coherent linguistic outputs unintended by any single participant. While psychological explanations such as the ideomotor effect exist, a computational understanding of how decentralized, implicit linguistic knowledge fuses through shared physical interaction remains elusive. We introduce CoCre-Sam (Collective-Creature Sampling), a framework modeling this phenomenon as collective Langevin dynamics sampling from implicitly fused language models. Each participant is represented as an agent associated with an energy landscape derived from an internal language model reflecting linguistic priors, and agents exert stochastic forces based on local energy gradients. We theoretically prove that the collective motion of the shared pointer (planchette) corresponds to Langevin MCMC sampling from the sum of individual energy landscapes, representing fused collective knowledge. Simulations validate that CoCre-Sam dynamics effectively fuse different models and generate meaningful character sequences, while ablation studies confirm the essential roles of collective interaction and stochasticity. Altogether, CoCre-Sam provides a novel computational mechanism linking individual implicit knowledge, embodied collective action, and emergent linguistic phenomena, grounding these complex interactions in the principles of probabilistic sampling.
Figures & tables
Figure 1: Conceptual illustration of the CoCre-Sam framework. Multiple agents interact with a shared planchette on a letter board. Each agent i possesses an internal language model Pi(cτ+1∣c1:τ) defining an energy landscape Ei . They exert micro-forces ai(t) based on their local energy gradient −∇Ei and stochastic noise (Langevin dynamics). The collective force a(t)=∑iai(t) guides the planchette x(t) , effectively performing sampling from the fused energy landscape Efused=∑iEi . As a result, a character sequence is generated not from either agent’s distribution alone, but from a fused probabilistic model that approximates a product-of-experts form. This leads to the attribution of agency to an imaginary creature named “Kokkuri-san,” distinct from either agent, as the perceived author of the outcome.
Figure 2: (Left) The CoCre-Sam simulation environment, representing the Ouija board as a 2D embedding space. Each letter, plus ’BOS’ (Begin) and ’EOS’ (End) tokens, is assigned a specific target location gc . (Right) An example trajectory of the planchette x(t) from a simulation with N=2 agents. The collective Langevin dynamics, driven by the agents’ fused energy landscapes (Eq. (4)), stochastically sample the coherent sequence ’P-E-O-N-Y-EOS’. Each color represents the trajectory segment as the dynamics converge on the goal location for each successive character, demonstrating the model’s generative process.
Agent 1 ( P1 )
Agent 2 ( P2 )
CoCre-Sam ( ∝P1P2 )
#
String
Freq (mean ± SD)
String
Freq (mean ± SD)
String
Freq (mean ± SD)
1
petunia
21.6±3.9
quassia
27.6±5.2
quaker
23.6±1.6
2
jacobinia
16.0±2.6
quaker
17.6±3.4
jacobinia
17.6±2.9
3
quince
12.0±2.4
primrose
09.8±4.3
petunia
15.8±3.5
4
iris
11.2±2.3
rue
09.6±2.4
rockrose
12.4±3.0
5
kalmia
08.8±1.6
jacobinia
08.2±1.6
quassia
08.4±4.5
Table 1: Top 5 generated flower names and their frequencies under various conditions. CoCre-Sam ( ∝P1P2 ) is the proposed model. ’No Noise’ ( Ti=0.0 ) is a deterministic gradient descent, ’Middle Noise’ ( Ti=0.5 ) an intermediate setting, and ’High Noise’ ( Ti=1.0 ) approaches a random walk. Frequencies are reported as mean ± SD across 5 random seeds , with 100 trials per (seed, condition).
Evaluated by LM:
Gen. by
P1
P2
Fused ( P1P2 )
Agent 1
1.840±0.027
2.128±0.076
1.920±0.038
Agent 2
2.509±0.133
1.935±0.046
2.139±0.074
CoCre-Sam
1.951±0.037
1.855±0.020
1.862±0.017
Table 2: Perplexity of generated valid words (mean ± SD across 5 random seeds, 100 trials each; word-level, then averaged with denominator 100 to match the generative process). Each row indicates the generating model; each column indicates the evaluation LM. Lower perplexity means higher predictability. The lowest entry in each row is bold ; for CoCre-Sam, P2 and Fused are both bolded because they are statistically tied (paired t -test, ∣t∣=0.64 , p>0.40 , n=5 seeds).
Figure 5
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A.1 : Flowchart of the hierarchical control in CoCre-Sam. The outer loop for character sequence generation (see Algorithm A.1 ) determines the energy function by weighting elemental energy functions based on the context c1:τ , while the inner loop simulates the probabilistic convergence of the planchette dynamics driven by the resulting energy landscape.
Figure D.2 : Step-by-step visualization of the word generation process for the word PEONY , followed by the [EOS] token. The six panels, arranged left-to-right, then top-to-bottom in a 3 × 2 layout, show the evolution of the fused energy landscape Efused and the planchette’s trajectory as each new character is selected.
Complexity and interpretability rarely coincide: systems rich enough for complex behaviours to emerge are usually too opaque to question, while transparent ones are too simple for anything complex to emerge. A single large language model (LLM) is a static artefact, hardly exhibiting any of the emergent properties we associate with life. This changes through interaction: populations of LLMs display emergent dynamics absent from isolated models. Furthermore, LLMs can be endowed with persistent memory, tools and shared skills, and the capacity to initiate actions unprompted, i.e., turning LLMs agentic. In this paper, we argue that such collectives of agents can serve as a computational substrate for Artificial Life (ALife) research. Critically, since the agents communicate in natural language, their collective behaviour can be directly interrogated by examining textual traces and asking the agents themselves. We outline the notion of interpretability in language-model research and extend it for collectives of agents. Lastly, we survey recent examples of agentic LLM collectives that already instantiate the idea of agentic substrates, from controlled experiments to deployments in the wild.
Elias Najarro, Ane Espeseth, Eleni Nisioti +2
IT University of Copenhagen, Denmark · University of Oslo, Norway · Sakana AI, Japan +1
Language-model agents act on state encodings of their environment, yet these are treated as interchangeable interfaces. Using pretrained language models, we designed a circular-synchronization experiment applying a state-encoding intervention while holding the physical system fixed: each agent sees only a summary of its neighbours' relative phases and chooses to advance, stay or retard. Encoding that state as low-order circular moments rather than as a histogram selected different collective outcomes. In GPT the moment encoding synchronized the population in 6/6 seeds and the histogram encodings in 0/6; the effect replicated in Claude but reversed direction. Replaying identical fields shifted each agent's advance/stay/retard probabilities far beyond within-encoding repeat variation, in GPT, Claude and Gemini; in GPT, presentation alone shifted the operator with the moment values fixed. State encodings therefore form part of a model-dependent effective interaction law, not a neutral interface.
Takahiro Ezaki, Naoto Imura, Katsuhiro Nishinari
Research Center for Advanced Science and Technology, The University of Tokyo, Tokyo, Japan · Department of Aeronautics and Astronautics, School of Engineering, The University of Tokyo, Tokyo, Japan
We introduce CRAFT, a multi-agent benchmark for evaluating pragmatic communication in large language models under strict partial information. In this setting, multiple agents with complementary but incomplete views must coordinate through natural language to construct a shared 3D structure that no single agent can fully observe. We formalize this problem as a multi-sender Bounded Pragmatic Speaker problem and provide a diagnostic framework that decomposes failures into spatial grounding, belief modeling and pragmatic communication errors, including a taxonomy of behavioral failure profiles in both frontier and open-weight models. Across a diverse set of models, including 8 open-weight and 7 frontier including reasoning models, we find that stronger reasoning ability does not reliably translate to better coordination: smaller open-weight models often match or outperform frontier systems, and improved individual communication does not guarantee successful collaboration. These results suggest that multi-agent coordination remains a fundamentally unsolved challenge for current language models. Our code can be found at https://github.com/csu-signal/CRAFT