SelfOrg: Decentralized LLM Collaboration

Updated 17 July 2026

SelfOrg is a self-organizing framework for multi-agent LLM collaboration that constructs communication structures from agents’ real-time responses.
It employs dense embeddings and a cosine similarity-based contribution score to form a directed acyclic graph, ensuring that correct answers amplify through the network.
Empirical results highlight significant performance boosts in weak LLM regimes and demonstrate robust behavior even with heterogeneous agent qualities.

SelfOrg is a response-conditioned framework for multi-agent systems based on LLMs that adapts communication on-the-fly by constructing a communication graph from the agents’ current responses. It was introduced as an alternative to fixed topologies, pretrained graph generators, optimization over edges, and external LLM judges, and is described as enabling the self-organization of agents without additional supervision or training. In the wider literature, self-organization denotes the emergence of global patterns or stable organized structure from local interactions without central control; SelfOrg instantiates that general idea at the level of inter-agent communication in LLM-based collaboration (Tastan et al., 1 Oct 2025, Gershenson et al., 2018).

1. Concept and intellectual context

Self-organization has been defined across several research traditions as the emergence of order from local interactions among system components. In Artificial Life, it is characterized as the ability of a system to display ordered spatio-temporal patterns solely as the result of the interactions among the system components, without the imposition of patterns by a central authority (Gershenson et al., 2019). Closely related formulations describe self-organization as the spontaneous creation of statistical interdependencies among parts of a system, rather than merely the convergence of trajectories toward attractors (Rosas et al., 2018).

A design-oriented strand of the literature frames self-organizing systems in terms of agents, interaction, and performance. In that setting, a central claim is that reducing the “friction” or “interference” of interactions between elements of a system will result in a higher “satisfaction” of the system, i.e. better performance [0505009]. Another line distinguishes self-organization from hierarchical organization: hierarchical organization is top-down, with goals and division of labor assigned by leaders, whereas self-organization relies on local rules and emergent coordination, even though self-organization can occur in different network structures (Busseniers, 2014).

SelfOrg inherits this broader vocabulary but applies it to a specific engineering problem: how to organize communication among LLM agents so that coordination is induced by the responses generated for the current query rather than by a predetermined orchestration scheme. This places it within the family of decentralized coordination methods while also making it a concrete protocol for response-driven collaboration (Tastan et al., 1 Oct 2025).

2. Core architecture

SelfOrg begins by broadcasting a query $\mathcal{Q}$ to $N$ LLM-based agents. Each agent independently generates a response $\mathcal{R}_n^{(0)}$ in the first round, with no prior coordination. These responses are transformed into dense embeddings $\mathbf{r}_n^{(0)}$ using a lightweight embedding model, and the resulting semantic representations are used to estimate each agent’s contribution and to form the communication structure for the next round (Tastan et al., 1 Oct 2025).

The communication structure is a directed acyclic graph (DAG). Information is propagated from higher-contributing agents to lower-contributing agents, and the graph is dynamically updated based on the agent responses from the previous collaboration round. The process proceeds for $T$ rounds: agents update their answers based on incoming neighbors, embeddings and contribution scores are recomputed, and the DAG is rebuilt. At the end, the system computes a contribution-weighted centroid in embedding space, and the final response is selected as the agent whose embedding best aligns with this centroid (Tastan et al., 1 Oct 2025).

Stage	Mechanism	Role
Initial round	Independent responses to $\mathcal{Q}$	Uncoordinated candidate generation
Intermediate rounds	Response embeddings, contribution scores, DAG updates	Response-conditioned propagation
Final selection	Contribution-weighted centroid	Response choice

This workflow is explicitly positioned against orchestration methods that depend on static graphs or external adjudication. SelfOrg is therefore both a collaboration policy and a mechanism for online topology formation (Tastan et al., 1 Oct 2025).

3. Contribution scoring and graph construction

The technical center of SelfOrg is its approximation to contribution assessment. The paper takes the Shapley value as the ideal notion of marginal contribution for agent $n$ ,

$\phi_n = \sum_{\mathcal{S} \subseteq [N] \setminus \{n\}} \frac{|\mathcal{S}|! (N-|\mathcal{S}|-1)!}{N!} \left[v(\mathcal{S} \cup \{n\}) - v(\mathcal{S})\right],$

but treats exact computation as intractable because enumerating all coalitions is exponential in $N$ (Tastan et al., 1 Oct 2025).

The operational proxy is

$\psi_n := \cos(\mathbf{r}_n, \mathbf{r}_{\mathrm{avg}}), \qquad \mathbf{r}_{\mathrm{avg}} = \frac{1}{N}\sum_{n=1}^N \mathbf{r}_n.$

This approximation is described as linear in $N$ 0, efficient, model-agnostic, and grounded in the cosine similarity of response embeddings. The paper also states an approximation bound and a ranking-stability corollary, arguing that relative orderings are preserved when contributions are well-separated (Tastan et al., 1 Oct 2025).

Graph construction follows this scoring step. Pairwise similarities $N$ 1 are computed, and an incoming edge $N$ 2 is retained when three conditions are satisfied: $N$ 3, $N$ 4, and, optionally, only the $N$ 5 most similar neighbors are kept. Cycles are then removed by breaking the edge from the lower to higher $N$ 6 agent, ensuring a DAG. The resulting topological ordering places higher-contribution agents upstream and lower-contribution agents downstream (Tastan et al., 1 Oct 2025).

This architecture means that communication is neither fully symmetric nor externally imposed. It is selectively routed through a topology determined by semantic agreement and estimated marginal utility, which is the precise sense in which the framework describes its own organization as self-organized (Tastan et al., 1 Oct 2025).

4. Correctness amplification under stochasticity

A defining claim of SelfOrg is that it goes beyond task- and query-level optimization and takes into account the stochastic nature of agent responses. Its theory is built around the observation that, when individual agent outputs are stochastic, multiple agents can improve the probability that correct information appears in the pool of responses and then dominates the communication process (Tastan et al., 1 Oct 2025).

If each agent has correctness probability $N$ 7, the probability that at least two agents are correct is

$N$ 8

The paper states that this quantity rapidly approaches $N$ 9 as $\mathcal{R}_n^{(0)}$ 0 grows. It further argues that correct answers cluster in embedding space, whereas mistakes are scattered across diverse wrong alternatives, so agreement on the same answer is disproportionately likely to reflect correctness (Tastan et al., 1 Oct 2025).

Under mild conditions, correct responders obtain larger $\mathcal{R}_n^{(0)}$ 1 values than incorrect responders. The paper’s corollary on correctness amplification states that if at least two agents generate the correct answer, their scores strictly exceed those of incorrect peers, ensuring that the DAG channels information from correct responders downstream. In effect, the communication graph is not merely a routing device: it is the mechanism through which correct signals dominate the information flow and random errors are suppressed (Tastan et al., 1 Oct 2025).

A broader interpretation is that SelfOrg treats agent interaction as a selection process over stochastic candidates. That interpretation aligns with wider self-organization research in which coordination under noise depends on the persistence or reinforcement of a small subset of more stable or more informative states, although the formal objects differ across traditions (Blumenfeld, 3 Aug 2025, Blumenthal et al., 26 Jun 2026).

5. Benchmarks and operating regimes

The empirical case for SelfOrg is strongest in the weak-LLM regime. For Qwen-2.5-1.5B-Instruct, the reported average accuracy and average rank are as follows (Tastan et al., 1 Oct 2025):

Method	AVG Accuracy	AVG Rank
Single	41.24	2.57
CoT	41.18	2.71
DyLAN	37.82	4.00
MacNet	36.60	4.57
G-Designer	33.47	5.86
AutoGen	18.93	6.06
SelfOrg	45.05	1.00

These results are described as showing robust performance, with significant gains in the weak regime where prior methods collapse. In the strong regime, SelfOrg is reported to retain the highest average performance as well, with 70.19% on LLaMA-70B and AVG-Rank 1.25 (Tastan et al., 1 Oct 2025).

The reported scaling behavior is asymmetric. SelfOrg shows its largest benefit on small and medium LLMs; the summary gives Qwen-3B as 65.35% versus 73.62% on AQUA-RAT and MMLU-Pro as 42.60% versus 46.20%. As model size increases, the marginal gain drops, and for large LLMs all approaches converge, indicating diminishing returns when the base LLM is already highly reliable (Tastan et al., 1 Oct 2025).

The framework is also evaluated with heterogeneous agent pools. There, it nearly matches the performance of the strongest agent, because weak agents are reliably assigned lower contribution and thus do not degrade the group output. This supports the claim that response-conditioned communication can be robust not only to stochasticity within a model family but also to quality heterogeneity across agents (Tastan et al., 1 Oct 2025).

6. Relation to adjacent research and interpretations

The broader literature distinguishes self-organization from several neighboring concepts. Self-configuration is characterized as a system’s capability to autonomously alter its structural or functional settings to adapt to changing environments, goals, or internal states, whereas self-organization is often rooted in emergence of order from local interactions (Lodwich, 2016). Organic computing, by contrast, treats self-organization as one self-* property among others, alongside self-configuration, self-integration, self-management, and robustness (Gill, 2018).

Quantification also varies substantially across fields. Some work measures self-organization through statistical complexity and proposes that self-organization occurs when complexity grows (Milovanovic et al., 2012). Other work treats binding information as a metric of global structural strength and identifies self-organization when that quantity increases over time (Rosas et al., 2018). A different tradition measures organization as the inverse of the average action per element per unit motion, multiplied by Planck’s constant, and interprets self-organization as minimization of average action (Georgiev, 2012). Still another proposes the Entropic Principle of Self-Organization, according to which equilibrium states maximize evolutionary entropy, contingent on the production rate of the external energy source (Demetrius, 2023).

These divergences matter for interpreting SelfOrg. SelfOrg is not a universal definition of self-organization, nor does it attempt to unify thermodynamic, biological, and information-theoretic accounts. It is a designed, response-conditioned mechanism for decentralized LLM collaboration. A plausible implication is that its significance lies less in providing a general ontology of organization than in showing how a concrete multi-agent system can induce communication structure from local response statistics without additional supervision or training (Tastan et al., 1 Oct 2025).

The wider self-organization literature repeatedly notes that the term is often confusing or misinterpreted, and that similar system-level functions can arise in different structures and with different formal criteria (Gershenson et al., 2018, Gershenson et al., 2019). In that context, SelfOrg is most precisely understood as a specific orchestration framework whose self-organization consists in dynamic, unsupervised topology formation and correctness-weighted information flow among LLM agents, rather than as a synonym for self-organization in general (Tastan et al., 1 Oct 2025).