---
title: Consensus Graphs (ConGrs) Overview
url: https://www.emergentmind.com/topics/consensus-graphs-congrs
type: topic
---

# Consensus Graphs (ConGrs) Overview

Consensus Graphs (ConGrs) denote graph-based representations of agreement, disagreement, or shared structure across multiple objects, but the term is not used uniformly across the literature. Its most explicit formalization is a weighted directed acyclic graph \(G(R)=(V,E)\) built from multiple sampled language-model responses, with consensus nodes \(V_C\), disagreement nodes \(V_D\), and edge weights equal to the fraction of responses following each edge [2510.03527]. In adjacent research areas, closely related “consensus graph” constructions appear as latent shared graphs in multiview graph learning, sparse co-association graphs in ensemble clustering, and graph sequences whose joint action performs exact averaging [2401.13769][1809.05578][2311.01317]. This suggests that ConGrs functions both as a specific DAG data structure and as a broader graph-centric paradigm for representing collective agreement.

## 1. Terminology, scope, and formal meaning

In the strictest sense, a ConGr is defined over a set \(R=\{r_1,\dots,r_m\}\) of responses sampled from a language model for the same prompt. Each response corresponds to a path through a weighted DAG. Consensus nodes contain anchor spans present across all sampled responses, disagreement nodes contain information that varies across responses, and each edge weight lies in \((0,1]\) and records the fraction of responses that traverse that edge [2510.03527].

Several neighboring literatures use the phrase “consensus graph” in a broader but structurally related way. In multiview graph learning, the consensus graph is a latent graph \(G=(V,E,\mathbf W)\) learned jointly with view-specific Laplacians to represent the structure shared across views [2401.13769]. In federated feature fusion, the consensus graph is the graph among local owners as graph nodes, used by a GCN to combine aligned local embeddings [2306.01240]. In graph clustering, ECG constructs a sparse consensus-induced reweighted graph whose edge weights encode co-clustering frequency across an ensemble [1809.05578]. In social-choice-style graph aggregation, a consensus graph is the output of an aggregation rule
\[
F:(2^{V\times V})^n \to 2^{V\times V},
\]
mapping \(n\) input graphs on a common vertex set to a single compromise graph [1609.03765].

This terminological spread is substantive rather than merely lexical. Some works study a graph that summarizes shared information, some study a graph that mediates agreement dynamics, and some study aggregation rules that output a collective graph. Several of these papers explicitly note that they do not use the exact acronym “ConGrs,” even though their central object is a consensus graph in this broader sense [2306.01240][1809.05578].

## 2. DAG-based ConGrs for multi-sample response synthesis

The clearest formal ConGr construction appears in graph-based response synthesis for language models. Given sampled responses \(R\), the ConGr \(G(R)=(V,E)\) is a DAG because it preserves left-to-right token order while allowing divergence and reconvergence. Construction begins with a lightweight lexical sequence-alignment stage adapted from Needleman–Wunsch partial order alignment, using full-word tokenization and an affine gap scheme with \(gap\_open\_penalty=-1\), \(gap\_extend\_penalty=-1\), \(match\_penalty=1\), and \(mismatch\_penalty=-2\). Consecutive nodes that appear in all responses are merged into consensus nodes, and the variable regions between adjacent consensus nodes are semantically clustered by a secondary LM judge into disagreement nodes [2510.03527].

The resulting graph alternates between stable shared scaffold and localized semantic variation. This enables two decoding regimes. In consensus decoding, nodes are retained in topological order when their weighted degree satisfies \(d^w(v)\ge \tau\), producing a high-support synthesized response; a final cleanup stage may return “Abstain” if the text is too fragmentary. In guided self-verification, disagreement branching after a consensus node is used to localize candidate reasoning errors; partial solutions are then pairwise verified only at those disagreement regions, and surviving full responses are synthesized into a final answer [2510.03527].

Empirically, these ConGrs improve factual precision on biography generation by up to 31% over an average response, reduce reliance on LM judges by more than 80% compared to other methods, increase abstention rate by up to 56% on refusal-style tasks, and improve reasoning accuracy on MATH and AIME by up to 6 points over self-verification and majority-vote baselines [2510.03527]. The graph therefore acts simultaneously as a representation of shared testimony and as a decoding substrate for task-dependent synthesis.

## 3. Learned shared graphs in multiview and federated settings

In multiview graph learning, the consensus graph is a latent shared topology learned jointly with view-specific graphs. For views \(i=1,\dots,N\), each with observations \(\mathbf X^i\) and Laplacian \(\mathbf L^i\), the paper formulates
\[
\sum_{i=1}^N \big\{\operatorname{tr}({\mathbf{X}^i}^\top \mathbf{L}^i \mathbf{X}^i) + \alpha \|\mathbf{L}^i\|_F^2\big\}
+ \beta c(\{\mathbf{L}^i - \mathbf{L}\}_{i=1}^N) + \gamma r(\mathbf{L}),
\]
subject to Laplacian constraints on \(\mathbf L^i\) and \(\mathbf L\). Here \(\mathbf L\) is the consensus graph Laplacian, \(c(\cdot)\) enforces similarity between each view graph and the consensus, and \(r(\mathbf L)\) regularizes the shared graph itself. Two regularization families are proposed: an \(\ell_{1,1}\) consensus penalty, which acts like a consensus-based extension of fused graphical lasso and favors sparse view-specific deviations, and an \(\ell_{2,1}\) penalty with \(\|\ell\|_1\) regularization, which acts like a consensus-based extension of group graphical lasso and promotes grouped deviations across views [2401.13769].

A closely related construction appears in federated feature fusion with models pre-trained on different features. There, each client learns a local representation \(\mathbf h_k^i=\phi_i^\ast(\mathbf x_k^i)\), the server aligns latent dimensions with matrices \(\mathbf P_i\), stacks the aligned embeddings into \(\mathbf H_k\), and applies a GCN whose adjacency \(\mathbf A\) is the consensus graph among local owners as graph nodes. When no graph is given, \(\mathbf A_{ij}\sim \operatorname{Ber}(\theta_{ij})\) is learned via an ICDF reparameterization, and only the global parameters—alignment matrices, graph parameters, and GCN weights—are trained after local models are frozen [2306.01240].

Across these settings, the consensus graph is not merely a post hoc summary. It is a learned latent object that mediates information sharing among heterogeneous views or clients. A plausible implication is that ConGrs in this sense are best understood as estimators of shared relational structure under explicit disagreement models.

## 4. Aggregation, co-association, and probabilistic consensus graphs

A graph-native form of consensus graph construction appears in ensemble clustering. ECG builds an ensemble \(\mathcal P=\{P_1,\dots,P_k\}\) from \(k\) randomized level-1 Louvain runs, computes edgewise co-clustering frequencies
\[
v_{P_i}(u,v)=\sum_{j=1}^{l_i}\mathbf 1_{C_i^j}(u)\mathbf 1_{C_i^j}(v),
\]
and then reweights each original edge by
\[
W_{\mathcal P}(u,v)=
\begin{cases}
w_* + (1-w_*)\cdot \left(\frac{\sum_{i=1}^k v_{P_i}(u,v)}{k}\right), & \text{if } (u,v) \text{ is in the 2-core of } G,\\
w_*, & \text{otherwise.}
\end{cases}
\]
The resulting weighted graph is reclustered by Louvain. The construction is sparse because it reweights only existing edges rather than forming a dense all-pairs co-association matrix; the recommended defaults are \(k=16\) and \(w_*=0.05\) [1809.05578].

At a more abstract level, graph aggregation studies consensus graph construction as a social-choice problem. Given \(n\) input graphs on a common vertex set, an aggregation rule \(F:(2^{V\times V})^n\to 2^{V\times V}\) produces a single output graph. Edge-wise quota rules, successor-approval rules, representative-voter rules, oligarchies, and dictatorships are all formal instances. The central impossibility results show that if a rule is unanimous, grounded, and satisfies Independence of Irrelevant Edges, then collective rationality with respect to graph properties that are contagious and implicative forces oligarchy on nonreflexive edges, and adding disjunctiveness forces dictatorship; this generalizes Arrow-style impossibility to a broad class of graph properties and application domains [1609.03765].

In probabilistic graphical models, the same tension reappears in a belief-aggregation form. Even when all agents agree on a common Bayesian-network topology, no method of combining beliefs can generally maintain that structure under mild assumptions, and local aggregation inside CPTs is also ruled out. The positive result is that the logarithmic opinion pool preserves commonly held Markov independencies, which supports a straightforward procedure for constructing a consensus Markov network [1301.6732]. Taken together, these works show that consensus graphs obtained by aggregation are often constrained by nontrivial impossibility frontiers whenever structural validity is required.

## 5. Consensus over dynamic, motion-induced, and graphon-defined topologies

Another major use of the term concerns consensus dynamics on graphs that are not static communication networks. In the moving-neighborhood model on the Petersen graph, \(n\) agents perform independent random walks on an undirected graph with 10 vertices and 15 edges and communicate only when they meet at the same node simultaneously. The state update is
\[
X_i(t+1)=X_i(t)+\varepsilon\sum_{j\in N_i(t)}b_{ij}(t)(X_j(t)-X_i(t)),
\]
with \(\varepsilon\in(0,1/\triangle)\) and \(\triangle=\max_i\sum_j b_{ij}(t)\). The work is primarily numerical, but it demonstrates that global consensus can be achieved asymptotically on an intermittently connected, encounter-driven communication graph induced by motion [1203.1900].

In evolutionary graphs, the communication graph itself is state-dependent and strategically updated. With continuous-time dynamics
\[
\dot x_t=-L_t x_t,
\]
and edges created or maintained by Bernoulli variables derived from a Continuous Actions Iterative Prisoner’s Dilemma mechanism, the paper proves average consensus almost surely and in mean square for any initial condition and graph topology, provided the feasible union graph \(G'\) is connected. The expected convergence rate is lower-bounded by \(\lambda_2(G')\), even though each realized graph \(G_t\) may be disconnected [1803.02564].

A different direction studies graph sequences whose product performs exact averaging in finitely many steps. A sequence \(\{G^{(l)}\}_{l=0}^{\tau-1}\) has the finite-time consensus property when the doubly stochastic matrices satisfy
\[
W^{(\tau-1)}W^{(\tau-2)}\cdots W^{(0)}=\frac{1}{n}\mathbf 1\mathbf 1^\top.
\]
One-peer exponential graphs, one-peer hyper-cubes, and \(p\)-peer hyper-cuboids provide explicit constructions, and the resulting topologies can be embedded in gradient tracking via GT-FT, whose rate depends on \(\tau\) but not on the connectivity of any individual graph in the sequence [2311.01317].

Graphon methods push the same theme to large or open systems. For dense graph sequences, if the continuum voting model on a graph limit \(W\) reaches consensus, then sufficiently large finite models become arbitrarily close to consensus outside a set of pairs of arbitrarily small measure; connected twin-graphons give a concrete consensus-inducing class [1603.06385]. For open multi-agent systems with replacements or with arrivals and departures, when topologies are sampled from a graphon, the paper derives upper bounds on disagreement in expectation and ties them to the spectrum of the expected sampled graph; for SBM graphons, the relevant spectral computation reduces to a matrix whose dimension depends only on the graphon, not on the number of agents [2503.24025].

## 6. Robustness, signed interactions, and topology-sensitive performance

Consensus graph structure strongly controls convergence speed, delay tolerance, and resilience. Self-similar hierarchical graphs \(\mathcal H(n,k)\) and Sierpiński graphs \(\mathcal S(n,k)\) have the same number of vertices and edges at each iteration,
\[
N_n=k^n,\qquad E_n=\frac{k^{n+1}-k}{2},
\]
yet their consensus properties differ sharply. Hierarchical graphs are non-fractal and small-world, yielding larger algebraic connectivity and lower first- and second-order coherence, whereas Sierpiński graphs are fractal and large-world, giving better delay robustness because \(\lambda_N(\mathcal S(n,k))=k+2\) while convergence and coherence scale less favorably [1712.06496].

For directed crash-tolerant consensus, the correct graph condition is not undirected connectivity but directional reachability after failures. In synchronous systems, exact crash consensus is possible iff the graph satisfies \(f\) Crash-Tolerant Node Connectivity, meaning every reduced graph obtained after removing up to \(f\) crashed nodes contains a source that reaches every surviving node. In asynchronous systems, exact consensus remains impossible, but approximate consensus is possible iff every partition \(L,C,R\) with nonempty \(L\) and \(R\) satisfies
\[
(L\cup C \to R)\ \vee\ (C\cup R \to L),
\]
where the propagate relation requires at least \(f+1\) distinct cross-cut influencers [1412.8532].

Resilient robotic consensus gives a geometric version of the same idea. For W-MSR, the communication graph built from Voronoi adjacency, equivalently the Delaunay triangulation \(G_\Delta\), is automatically \((2,2)\)-robust for \(N>2\), which suffices to reject one globally bounded non-cooperative robot. Augmenting the graph to \(G_{\Delta 2}\) by connecting two-hop Delaunay neighbors yields \((3,3)\)-robustness for \(N>4\), supporting rejection of two globally bounded non-cooperative robots or one locally bounded adversary per neighborhood [2209.04381].

Weight asymmetry and signed couplings enlarge the design space further. For weighted directed graphs with a rooted in-branching, Nyquist-based analysis gives an admissible variation interval for a single perturbed edge weight; consensus can survive even when that edge becomes negative, with graph-theoretic interpretations for DAGs and directed cycles [1609.00283]. Pseudo-undirected path graphs retain bidirectional connectivity but allow opposite-direction weights to differ. Their Laplacian is generally non-symmetric, the consensus value is determined by the positive left null vector \(\mathbf p\in\mathfrak N(\mathcal L^\top)\), and admissible negative weights can place the final consensus outside the convex hull of the initial state set [2509.20314]. In signed social networks with exactly three clustering-balanced camps, suitable diagonal stubbornness gains in a modified DeGroot law yield either tripartite consensus,
\[
\lim_{t\to\infty}\mathbf x(t)=
\begin{bmatrix}
c_1\mathbf 1_{n_1}\\
c_2\mathbf 1_{n_2}\\
c_3\mathbf 1_{n_3}
\end{bmatrix},
\]
or sign consensus with positive, zero, and negative asymptotic classes [2103.04938].

Across these literatures, consensus graphs are not a single object class but a family of graph constructions used to encode common structure, induce agreement dynamics, or aggregate multiple graph-valued inputs. The unifying feature is that the graph is not treated as passive background: it is itself the carrier of agreement, disagreement, robustness, or shared representation.

Source: https://www.emergentmind.com/topics/consensus-graphs-congrs