CoDiCon: Competitive Diversity via Constructive Conflict

• CoDiCon is a design principle that balances local cooperative alignment with global competitive pressures to sustain diversity in multi-agent settings.
• In lattice models, it drives self-organized clusters by coupling neighbor matching with a dissimilarity penalty to stabilize distinct community interfaces.
• In MARL and LLM frameworks, CoDiCon employs rank-ordered intrinsic rewards and controlled deviations to foster strategic diversity without losing overall team coherence.

Searching arXiv for the specified CoDiCon-related papers to ground the article in the cited literature. Competitive Diversity through Constructive Conflict (CoDiCon) denotes a class of formulations in which diversity is sustained not by suppressing coordination, but by placing cooperative alignment under a countervailing competitive or differentiating pressure. In 2025 work, the acronym is used for a minimal lattice-evolution model of cultural or linguistic diversification, for a constructive conflict-driven multi-agent reinforcement learning algorithm, and for an LLM-based multi-agent methodology centered on the consensus-diversity tradeoff (Noronha et al., 13 Oct 2025, Mai et al., 16 Sep 2025, Wu et al., 23 Feb 2025). Across these usages, the common mechanism is that moderate conflict is treated as constructive when it prevents premature homogenization while preserving collective performance.

1. Conceptual core

CoDiCon is organized around an antagonistic pairing of forces. In the lattice-evolution model, agents gain fitness by matching with their four nearest neighbors, while also gaining a dissimilarity-driven term with respect to the entire population; the resulting “constructive conflict” produces a self-organized mosaic of distinct clusters whose interfaces are stabilized by equalizing boundary fitnesses (Noronha et al., 13 Oct 2025). In the MARL formulation, the same phrase refers to competitive incentives injected into fully cooperative training through a centralized intrinsic reward module that ranks agents and distributes distinct intrinsic rewards, thereby encouraging policy exchange and strategic diversity without abandoning the team objective (Mai et al., 16 Sep 2025). In the LLM-based multi-agent setting, the operative idea is partial disagreement under implicit consensus: agents exchange information yet independently form decisions via in-context learning, preserving enough deviation from group norms to maintain exploration and robustness in dynamic environments (Wu et al., 23 Feb 2025).

A concise way to compare the three formulations is to track which pressure promotes coherence and which pressure prevents collapse into uniformity.

Context	Coherence mechanism	Diversity mechanism
Lattice evolution	Matching with four nearest neighbors	Global dissimilarity term
MARL	Shared team reward under CTDE	Ranking-based intrinsic rewards
LLM-based MAS	Public discussion and transcript sharing	Role prompts and deviation with probability $\beta$

This suggests that CoDiCon is less a single algorithm than a transferable design principle: coherent local structure or team-level optimization is retained, but is systematically frustrated by a mechanism that rewards differentiation, rank separation, or partial deviation.

2. Minimal evolutionary formulation on a lattice

In "A Self-Organized Tower of Babel: Diversification through Competition" (Noronha et al., 13 Oct 2025), each agent’s “language,” or more generally cultural state, is a high-dimensional bit-vector. The lattice has size $L \times L$, total population $N=L^2$, and bit-vector dimension $B$. Agent $i$ at site $x$ has state

$x_i \equiv c_i \in \{0,1\}^B.$

Two pairwise measures are defined between bit-vectors $c$ and $c'$: the understandability

$$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$

and the Hamming distance

$L \times L$0

The total fitness of agent $L \times L$1 is

$L \times L$2

with $L \times L$3 weighting local cooperation and $L \times L$4 weighting global dissimilarity; the paper sets $L \times L$5 and varies $L \times L$6 (Noronha et al., 13 Oct 2025).

The same decomposition can be expressed as

$L \times L$7

$L \times L$8

so that

$L \times L$9

To emphasize the conflict, the global term can be interpreted as an inhibitory penalty, with an effective form

$N=L^2$0

where $N=L^2$1 in that sign convention, although the paper keeps both terms positive and interprets the second as discommunication.

Evolution proceeds in discrete generations with two stages. In the replication or invasion stage, there are $N=L^2$2 trials per generation. A random lattice site $N=L^2$3 is selected; its fitness and the fitnesses of its four nearest neighbors are computed; if $N=L^2$4, then $N=L^2$5 invades the weakest neighbor $N=L^2$6 and copies its bit-vector to that site, $N=L^2$7. The newly created clone is then flagged “immune” for the remainder of the replication phase, so it cannot itself replicate or be overwritten in that generation. Mutation is applied once per generation: for each site $N=L^2$8 and each bit $N=L^2$9, the bit flips independently with probability $B$0 (Noronha et al., 13 Oct 2025).

The key structural claim is that local alignment promotes homogeneity within each community, while the global dissimilarity term drives communities to differentiate from one another. The model is therefore minimal in the sense that diversification is not produced by exogenous heterogeneity, but by the conflict between short-range alignment and population-wide differentiation.

3. Cluster organization, boundary equalization, and phase structure

After a transient, agents coalesce into clusters of identical or nearly identical bit-vectors. Within each cluster, the local $B$1-terms are high; across the global term, large clusters suffer a dissimilarity penalty, so a trade-off sets the preferred fraction each cluster can occupy (Noronha et al., 13 Oct 2025). On a square lattice, clusters tend to meet along straight horizontal or vertical boundaries. Boundary agents lose some local fitness because they have fewer matching neighbors than interior agents.

A central result is that coexistence of clusters with different bulk fitness is possible because boundary fitnesses are equalized. If one cluster had stronger boundary fitness than its neighbor, it would invade until the boundary fitnesses matched again. Boundaries therefore act as pressure-equalizing interfaces rather than merely passive separators. This is the mechanism by which distinct communities persist without requiring equal bulk properties.

With nonzero mutation $B$2, the system does not settle into a single fixed configuration; it wanders among metastable mosaics of languages. New languages nucleate almost exclusively at existing cluster boundaries. A mutant in the center of a large cluster is usually unfit because it loses many neighbor matches and dies out, whereas a mutant at a boundary can surpass its neighbors because local fitness is already low there. Once established, a new language expands by invading weaker boundary segments until a new balance of inter-cluster fitnesses is reached. The observed transitions are therefore punctuated shifts between quasi-stationary states rather than smooth relaxations.

The reported phase structure is specific. For $B$3, the system splits into approximately $B$4–$B$5 clusters on a $B$6 grid; for larger $B$7, it coarsens to one cluster. The largest-cluster fraction remains approximately $B$8 of the lattice over a wide range of $B$9, set by the global dissimilarity penalty. For $i$0 and $i$1, a transition to global homogeneity occurs above a critical $i$2. The threshold shifts slightly with $i$3, $i$4, and initial conditions, showing hysteresis akin to nucleation-driven phase transitions (Noronha et al., 13 Oct 2025).

Mutation controls the diversification regime. As $i$5 increases, the number of clusters $i$6 rises and the largest-cluster fraction falls, with approximate power-law scaling in $i$7. Above a high-$i$8 error-catastrophe line, all structure collapses and the lattice becomes a patchwork of random bit-vectors. In one dimension, clusters become nearly equal-sized; a boundary-balance calculation in the Supplement shows that if two adjacent languages $i$9 have equal bulk fitness $x$0, then their populations $x$1 must also be equal in steady state (Noronha et al., 13 Oct 2025).

4. Constructive conflict-driven MARL

In "Constructive Conflict-Driven Multi-Agent Reinforcement Learning for Strategic Diversity" (Mai et al., 16 Sep 2025), CoDiCon is formulated in the standard fully-cooperative Dec-POMDP or Markov game setting

$x$2

Here $x$3 is the set of global states, $x$4 is the discrete action set of agent $x$5, $x$6 is the joint action space, $x$7 is the transition probability, $x$8 is the shared team reward, and $x$9 is the discount factor. Each agent observes only its own partial observation and follows a decentralized stochastic policy $x_i \equiv c_i \in \{0,1\}^B.$0, while a global critic can access $x_i \equiv c_i \in \{0,1\}^B.$1 and $x_i \equiv c_i \in \{0,1\}^B.$2 under centralized training and decentralized execution. The classical cooperative objective is

$x_i \equiv c_i \in \{0,1\}^B.$3

CoDiCon augments each agent’s learning signal with a competition-driven intrinsic reward. At time $x_i \equiv c_i \in \{0,1\}^B.$4, agent $x_i \equiv c_i \in \{0,1\}^B.$5 receives

$x_i \equiv c_i \in \{0,1\}^B.$6

with $x_i \equiv c_i \in \{0,1\}^B.$7 balancing extrinsic and intrinsic signals, and discounted return

$x_i \equiv c_i \in \{0,1\}^B.$8

The intrinsic reward is generated by a centralized ranking module. A small neural network $x_i \equiv c_i \in \{0,1\}^B.$9, parameterized by $c$0, takes as input the global state $c$1 and optionally joint actions $c$2, and emits $c$3 scalar scores $c$4. These are sorted in ascending order,

$c$5

guaranteeing that each agent receives a distinct intrinsic reward (Mai et al., 16 Sep 2025).

This induces deliberate competition inside a cooperative task. When an agent’s action increases its rank, its intrinsic gradient may oppose the team’s extrinsic gradient; the paper identifies conflict whenever

$c$6

The conflict is constructive because the outer optimization is still defined by the global extrinsic return. Architecturally, the ranking module is a small fully-connected network with one or two MLP layers, outputting $c$7 real-valued scores followed by a non-differentiable sort layer. To regularize the pre-sort outputs toward a user-provided target ordering $c$8, two auxiliary losses are used:

$c$9

and

$c'$0

combined as

$c'$1

Minimizing $c'$2 spreads the unsorted outputs along the prescribed ordering and helps maintain distinct intrinsic signals.

Training is posed as a bilevel problem. The outer problem maximizes the true extrinsic objective over $c'$3,

$c'$4

subject to the inner problem

$c'$5

with the ranking constraint $c'$6 absorbed into the auxiliary loss. The coupling is approximated with a one-step meta-gradient. If

$c'$7

then

$c'$8

In implementation, policies are updated by a clipped-PPO step on the hybrid advantage, and the ranking module is updated both by the auxiliary loss and by the meta-gradient on $c'$9 (Mai et al., 16 Sep 2025).

5. Empirical behavior in MARL and adaptive multi-agent systems

The MARL evaluation in (Mai et al., 16 Sep 2025) uses SMAC scenarios $$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$03s_vs_5z, 8m_vs_9m, MMM2, 2c_vs_64zg, 5m_vs_6m, 3S5Z$$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$1 and GRF tasks $$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$2academy_3_vs_1_with_keeper, academy_counterattack_easy, academy_counterattack_hard$$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$3. Baselines are MAPPO, LIIR, EOI, and CDS. Key hyperparameters are PPO clip $$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$4, learning rates $$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$5 and $$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$6 for $$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$7, intrinsic weight $$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$8, ranking targets $$U(c,c') \equiv c \cdot c' = \sum_{b=1}^B [c_b \text{ AND } c'_b],$$9 fixed at initialization with $L \times L$00 positive and $L \times L$01 negative, and ranking-loss weights $L \times L$02, $L \times L$03. The reported outcome is that CoDiCon converges faster and to higher win-rates on most maps, particularly 5m_vs_6m and MMM2, where it reaches about $L \times L$04–$L \times L$05 win versus about $L \times L$06–$L \times L$07 for MAPPO and about $L \times L$08–$L \times L$09 for LIIR or CDS. On GRF it matches or slightly exceeds MAPPO on the easiest task and significantly outperforms all baselines as difficulty increases, with final win-rates above $L \times L$10 on the hard scenario versus $L \times L$11–$L \times L$12 for MAPPO. In a synthetic Pac-Men testbed, CoDiCon agents gravitate in pairs to the rich room and achieve about $L \times L$13 mean episode return versus about $L \times L$14 for baselines. Removing either the MSE or variance ranking loss degrades performance significantly, indicating that both are necessary to maintain a well-spread, distinct intrinsic reward signal (Mai et al., 16 Sep 2025).

The LLM-based multi-agent formulation in (Wu et al., 23 Feb 2025) makes the consensus-diversity tradeoff explicit. Over $L \times L$15 agents with actions $L \times L$16, consensus is

$L \times L$17

and diversity is

$L \times L$18

System performance $L \times L$19 is described as following an inverted-U in diversity,

$L \times L$20

peaking at an intermediate $L \times L$21. The method assigns each agent a distinct role prompt, provides access to the public transcript and new observation, and then separates a discussion phase from a decision phase in which each agent independently forms its chain-of-thought and selects an action. A prompt clause can specify that the agent may deviate from group signals with probability $L \times L$22, thereby tuning the average diversity.

The theoretical rationale is presented as an exploration-exploitation tradeoff under environment volatility. If the true optimal action shifts unpredictably with shock rate $L \times L$23, then full consensus with $L \times L$24 yields regret on the order of the shock rate because the team never explores, while full fragmentation yields regret near one because coherent learning disappears. If a fraction $L \times L$25 of agents continually test alternatives, shift detection occurs within $L \times L$26 rounds, total regret scales like $L \times L$27, and minimizing with respect to $L \times L$28 gives $L \times L$29 and regret $L \times L$30; by contrast, forced consensus incurs $L \times L$31 (Wu et al., 23 Feb 2025).

Empirically, the paper reports five seeds per setting and four LLMs—GPT-4, Claude, Qwen, and Llama-2—at temperature $L \times L$32 and 256-token context. Three scenarios are studied: Dynamic Disaster Response, Information Spread and Manipulation, and Dynamic Public-Goods Provision. The comparison is between explicit consensus and implicit consensus.

Scenario	Explicit	Implicit
Disaster Response	CR $L \times L$33, MP $L \times L$34, RD $L \times L$35	CR $L \times L$36, MP $L \times L$37, RD $L \times L$38
Misinformation	MS $L \times L$39, CT $L \times L$40, CD $L \times L$41	MS $L \times L$42, CT $L \times L$43, CD $L \times L$44
Public Goods	PR $L \times L$45, TW $L \times L$46, FD $L \times L$47	PR $L \times L$48, TW $L \times L$49, FD $L \times L$50

Under high volatility in disaster response, CoDiCon maintains coverage rate at or above $L \times L$51 versus below $L \times L$52 for explicit consensus. Across all three tasks and LLMs, implicit consensus outperforms forced alignment by $L \times L$53–$L \times L$54 on key metrics, with a clear correlation to intermediate deviation $L \times L$55 (Wu et al., 23 Feb 2025).

6. Scope, interpretation, and recurring issues

The three CoDiCon formulations share a common claim: diversity is maintained not by removing cooperation, but by constraining it with an opposing pressure. In the lattice model, constructive conflict is the interplay between short-range alignment benefit and long-range differentiation penalty; in MARL, it is the tension between the shared team reward and rank-ordered intrinsic rewards; in the LLM-based setting, it is the balance between consensus formation and partial deviation (Noronha et al., 13 Oct 2025, Mai et al., 16 Sep 2025, Wu et al., 23 Feb 2025).

A recurrent misconception is that conflict here denotes breakdown of cooperation. The papers use the term more narrowly. In the MARL formulation, conflict occurs when intrinsic and extrinsic gradients oppose one another, but the outer optimization still maximizes the environmental return $L \times L$56 (Mai et al., 16 Sep 2025). In the LLM-based setting, the claim is not that maximal disagreement is desirable; rather, both full consensus and full fragmentation are described as suboptimal, with performance peaking at intermediate diversity $L \times L$57 (Wu et al., 23 Feb 2025). In the lattice-evolution model, coexistence does not require equal bulk fitness; it relies on equalized boundary fitnesses, so diversity is stabilized by interface dynamics rather than by uniformity across communities (Noronha et al., 13 Oct 2025).

The domain generalizations proposed for the minimal model are broad but specific: cultural evolution beyond language, including institutions, norms, and fashions; ecological or microbial communities, described as “cryptographical arms races” in quorum sensing signals; technological standards, characterized by local compatibility versus diverse global niches; and network-formation models with a local-global frustration term in spin or Potts-like systems (Noronha et al., 13 Oct 2025). A plausible implication is that CoDiCon is most relevant where systems simultaneously require within-group coherence and between-group differentiation, or team-level coordination and agent-level exploration.

Another recurring issue concerns whether diversity is a transient by-product or a stable regime. The lattice model supports quasi-stationary mosaics with punctuated shifts triggered by boundary nucleation (Noronha et al., 13 Oct 2025). The MARL algorithm maintains strategic diversity through distinct intrinsic rewards while reporting faster convergence and robust performance in sparse, hard-exploration domains such as GRF counterattacks and SMAC micromanagement (Mai et al., 16 Sep 2025). The LLM-based methodology ties performance gains to preserved partial diversity under dynamic shocks, rather than to a single consensus policy (Wu et al., 23 Feb 2025). Taken together, these results indicate a coherent research program in which constructive conflict functions as a mechanism for preventing premature homogenization while preserving effective collective behavior.