In-Collective Adaptive Topologies

• In-Collective Adaptive Topologies are dynamic structures that reconfigure communication links in real time based on the collective’s state, workload, and task progress.
• They integrate feedback mechanisms and schedule-level adaptations to optimize throughput and minimize delays in diverse systems ranging from photonic networks to multi-agent setups.
• Applications span high-performance computing, decentralized agent systems, and biologically inspired networks, highlighting trade-offs between reconfiguration overhead and performance gains.

In-collective adaptive topologies are topological adaptation mechanisms in which the communication, interaction, or dependency graph is not fixed for an entire system run, but is changed from within the collective process itself in response to the structure of the collective, the current collective state, or the ongoing task. In the most explicit formulation, a collective is decomposed into a sequence of steps, and the interconnect is reconfigured synchronously between steps only when the reduction in congestion and path length exceeds the reconfiguration delay (Addanki, 9 Oct 2025). Related literatures use different terminology—adaptive networks in steady state, collective adaptivity, topology-aware collective synthesis, co-evolutionary rewiring, or self-adaptive networks—but converge on the same structural idea: topology is an internal degree of freedom of the collective rather than an external constant (Wieland et al., 2015, Wang et al., 29 May 2025, Won et al., 2023, 0811.0980).

1. Conceptual definition and scope

The central distinction is between a topology chosen once and a topology adapted during collective execution. In programmable photonic scale-up domains, the topology is adapted within a collective and synchronized to the collective’s step structure, rather than being chosen once for the whole system (Addanki, 9 Oct 2025). In embodied collective systems, the analogous claim is that a collective should be able to dynamically adjust its size, topology, and functionality during inference, rather than merely executing a fixed interaction graph learned at design time (Wang et al., 29 May 2025). In adaptive-network theory, the same motif appears as co-evolution: node states and interaction structure change together, and topology is rewired using local information generated by the collective’s own dynamics (0811.0980).

Across these domains, the term denotes a family of mechanisms rather than one algorithm. In collective communication, the collective is a communication primitive such as AllReduce or All-to-All, and the adaptive object is the communication fabric or logical routing schedule (Addanki, 9 Oct 2025, Won et al., 2023). In multi-agent systems, the collective is a set of interacting agents, and the adaptive object is the communication or collaboration graph induced by task-conditioned relations or learned interaction modes (Zhang et al., 2024, Sun et al., 15 Jan 2026). In coevolutionary and biological networks, the collective is the dynamical population itself, and the adaptive object is the network of contacts, couplings, or leadership relations through which dynamics propagate (Wieland et al., 2015, Eom et al., 2015, Pais et al., 2013).

A useful common characterization is that in-collective adaptive topology is a feedback architecture. The collective generates state, demand, or performance signals; those signals alter the topology; and the altered topology in turn changes subsequent collective behavior. In photonic collectives this feedback is step-synchronous and algorithmically scheduled (Addanki, 9 Oct 2025). In adaptive oscillator and epidemic networks it is state-dependent rewiring (Eom et al., 2015, Wieland et al., 2015). In LLM-based or embodied collectives it is task- or state-conditioned graph construction (Zhang et al., 2024, Wang et al., 29 May 2025). This suggests that the concept is best understood as a systems-level design principle rather than a domain-specific mechanism.

2. Formalization in collective communication systems

The most explicit mathematical treatment appears in adaptive photonic scale-up domains. A scale-up domain with $n$ GPUs is modeled with one photonic port per GPU, transceiver bandwidth $b$, and reconfiguration delay $\alpha_r$. The collective is represented as a sequence of $s$ pairwise matching steps, where step $i$ exchanges $m_i$ units of data according to a permutation or matching matrix $\mathcal{M}_i$, yielding the aggregate demand

$$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$

This is interpreted as a Birkhoff–von Neumann decomposition of the aggregate traffic matrix, with the explicit caveat that a BvN decomposition of aggregate demand does not necessarily define a valid collective schedule because collective steps may have temporal dependencies (Addanki, 9 Oct 2025).

The topological performance of a given step pattern on a base graph $G=(V,E)$ is expressed through the maximum concurrent flow $\theta(G,\mathcal{M}_i)$. The demand completion time for step $b$0 is

$b$1

and, with fixed per-step latency $b$2, propagation delay $b$3, and $b$4, becomes

$b$5

Summing over steps produces a collective completion time that is explicitly topology-dependent through both path length and concurrent-flow throughput (Addanki, 9 Oct 2025).

The reconfiguration decision is then cast as a $b$6-$b$7 optimization. Binary variables $b$8 select between the base topology $b$9 and the step-matched topology $\alpha_r$0, while $\alpha_r$1 encode whether adjacent steps both use the base topology. The objective minimizes propagation delay, reconfiguration delay, fixed latency, and congestion penalties: $\alpha_r$2 subject to the stated adjacency constraints on $\alpha_r$3, with $\alpha_r$4 and $\alpha_r$5 (Addanki, 9 Oct 2025).

The main conclusion is a regime statement rather than an unconditional endorsement of reconfiguration. Adaptive reconfiguration is beneficial only when message sizes are large enough, reconfiguration delay is small enough, and the throughput gain from matching topology to step outweighs switching overhead. Preliminary simulations with $\alpha_r$6 GPUs, $\alpha_r$7 Gbps links, $\alpha_r$8 ns, a ring base topology, and collectives including recursive doubling, Swing, and All-to-All show three regimes: high-$\alpha_r$9/small-message settings where optimized schedules outperform naive BvN schedules; low-$s$0/large-message settings where optimized schedules beat static rings; and an intermediate regime where selective reconfiguration outperforms both always-static and always-reconfigure baselines (Addanki, 9 Oct 2025). Because the decision at each step depends only on the previous step, the paper notes that the problem admits an efficient dynamic programming solution despite the ILP form (Addanki, 9 Oct 2025).

3. Topology-aware synthesis inside collectives

A second systems lineage treats in-collective adaptation not as runtime physical reconfiguration but as synthesis of a collective schedule whose logical topology changes chunk by chunk and time step by time step. TACOS formalizes this with a Time-Expanded Network (TEN), a directed acyclic graph that replicates the physical network across time so that chunk motion, link occupancy, and dependency evolution are represented in a single structure (Won et al., 2023). Rather than solving a global ILP, TACOS frames synthesis as a link-chunk matching problem: each chunk is matched to a TEN link, each link carries at most one chunk at a time, and the procedure greedily maximizes the number of successful matches at each time span (Won et al., 2023).

This yields a schedule that is static once produced but adaptive to the topology and to the collective’s own progress during synthesis. TACOS supports All-Gather, Broadcast, Reduce, Reduce-Scatter, and All-Reduce via composition, models heterogeneous links through the $s$1-$s$2 cost $s$3, and claims polynomial synthesis time $s$4 (Won et al., 2023). Reported results include synthesis of an All-Reduce algorithm for a heterogeneous 128-NPU system in 1.08 seconds, up to 4.27× performance improvement over state-of-the-art synthesizers, and synthesis for 40K NPUs in 2.52 hours (Won et al., 2023). These results are used to argue that collective schedules should be matched to heterogeneous and asymmetric topologies rather than selected from a small fixed library.

PCCL extends the same synthesis logic to process groups. Its key claim is that a collective typically involves only a subset of devices, but the optimal schedule may still use links and intermediate devices outside that subset; hence synthesis should be process group-aware while reasoning over the full physical topology (Won et al., 5 Jun 2026). PCCL retains the TEN representation, defines operators such as $s$5, $s$6, and $s$7, and uses BFS-based path construction over the TEN to find paths from each chunk source to its destinations (Won et al., 5 Jun 2026). Conditions are chunk-centric, which permits arbitrary collective patterns including point-to-point, multicast, $s$8, and $s$9 (Won et al., 5 Jun 2026).

Its congestion-avoidance heuristic sorts conditions by

$i$0

schedules longer paths first, and removes used TEN links before synthesizing the next condition (Won et al., 5 Jun 2026). For heterogeneous networks PCCL uses $i$1, explicitly models switches as devices with state and optional multicast restrictions, and reuses the standard reductions-from-nonreductions transformations for Reduce, Reduce-Scatter, and AllReduce (Won et al., 5 Jun 2026). Reported results include 512-NPU All-to-All synthesis in 11.68 minutes, a 1,000-NPU case in 2.01 hours, 4,404× faster synthesis than TE-CCL on a 36-NPU 2D mesh, and process-group-aware speedups of 2.33–3.03× over pairwise direct on a mesh, averaging 2.68× (Won et al., 5 Jun 2026).

Taken together, these works define an important subclass of in-collective adaptive topologies: topology-aware but schedule-level adaptation. The physical network may remain unchanged, yet the collective’s effective communication graph is re-authored internally across time, chunks, and phases so as to exploit topology diversity, bandwidth asymmetry, and otherwise idle links (Won et al., 2023, Won et al., 5 Jun 2026). A plausible implication is that, in practice, “adaptation” spans a continuum from physical rewiring to logical schedule synthesis rather than a single implementation choice.

4. Co-evolutionary adaptive networks and steady-state topology

Outside high-performance communication, the concept is rooted in adaptive-network theory, where network structure and node states coevolve. In the adaptive contact process, nodes are susceptible $i$2 or infected $i$3, the mean degree $i$4 is fixed, infection proceeds along $i$5 links with rate $i$6, recovery occurs with rate $i$7, and susceptible nodes rewire away from infected neighbors with rate $i$8. The key dimensionless control parameter is

$i$9

which sets the balance between topological adaptation and state change (Wieland et al., 2015). The model admits an active steady state when $m_i$0, and the stationary fractions satisfy $m_i$1 (Wieland et al., 2015).

Using degree-resolved master equations, the stationary ensemble degree distributions $m_i$2 and $m_i$3 can be solved in closed form. The analysis yields coupled recurrence relations and explicit product-form solutions for $m_i$4, together with a special point at $m_i$5 where $m_i$6 and the degree distribution is exactly Poisson: $m_i$7 The mean-degree ordering obeys

$m_i$8

and the variance relations imply that rewiring reduces heterogeneity relative to Erdős–Rényi when $m_i$9 and increases heterogeneity when $\mathcal{M}_i$0 (Wieland et al., 2015). The paper emphasizes that agreement with simulation is strong except when rewiring is much faster than state update, because the mean-field assumption that neighbor-state fractions are independent of degree breaks down (Wieland et al., 2015).

In oscillator networks, a related state-dependent rewiring rule produces concurrent enhancement of percolation and synchronization. Kuramoto oscillators evolve according to

$\mathcal{M}_i$1

while edges are redrawn with probability

$\mathcal{M}_i$2

Because phase-similar oscillators are preferentially connected, synchronization and connectivity reinforce one another through a positive feedback loop (Eom et al., 2015). Steady-state synchronization is measured by the Kuramoto order parameter $\mathcal{M}_i$3, percolation by the largest-component fraction $\mathcal{M}_i$4, and even partial adaptation with update probability $\mathcal{M}_i$5 is reported to noticeably boost the collective transitions (Eom et al., 2015). The adaptive case yields sharper, first-order-like transitions, while a blinking-network control shows that fast switching alone does not reproduce the percolation enhancement (Eom et al., 2015).

Earlier work on Boolean and threshold networks formulates two canonical local rewiring schemes: active nodes lose links while frozen nodes gain links, and correlated nodes connect while decorrelated nodes disconnect. In random threshold networks the evolved mean connectivity obeys

$\mathcal{M}_i$6

with $\mathcal{M}_i$7 and $\mathcal{M}_i$8, approaching the critical connectivity $\mathcal{M}_i$9 in the thermodynamic limit; analogous finite-size scaling is reported for random Boolean networks and correlation-driven neural models (0811.0980). The evolved systems display inhomogeneous topologies, broad homeostatic plateaus, $$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$0 noise with exponent $$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$1 in RTN simulations, and scale-free attractor periods in the Boolean-network setting (0811.0980). In this literature, in-collective adaptive topology denotes self-organization toward criticality by purely local rewiring rules.

5. Collective intelligence, embodied systems, and task-conditioned graphs

In embodied multi-agent systems, the corresponding notion appears as collective adaptivity. A Collective Adaptive Intelligence framework argues that a collective should adapt not only its task policy but also its organization: which agents interact, which participate, and how responsibilities are distributed under partial observability and decentralized communication constraints (Wang et al., 29 May 2025). The formal setting is

$$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$2

with discounted return

$$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$3

For a collective $$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$4, agent $$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$5 updates according to

$$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$6

and its time-varying neighborhood is

$$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$7

Here $$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$8 functions as a position, role, or organization descriptor, and topology is induced dynamically from learned or updated latent states (Wang et al., 29 May 2025). The framework links this mechanism to decentralization, self-adaptation, self-assembly, collective resilience, and collective scalability (Wang et al., 29 May 2025).

LLM-based multi-agent systems make the same principle more explicit. G-Designer represents agents as nodes $$\mathcal{M} = m_1 \cdot \mathcal{M}_1 + m_2 \cdot \mathcal{M}_2 + \ldots + m_s \cdot \mathcal{M}_s.$$9 in a directed graph $G=(V,E)$0, augments the graph with a task-specific virtual node $G=(V,E)$1, and uses a variational graph auto-encoder to decode a task-adaptive sparse communication topology (Zhang et al., 2024). The task node is encoded from the query, agents are encoded with all-MiniLM-L6-v2 using embedding dimension $G=(V,E)$2, and the learned graph is optimized for utility, anchor consistency, and sparsity (Zhang et al., 2024). Reported outcomes include MMLU accuracy of 84.50%, HumanEval pass@1 of 89.90%, token reduction by up to 95.33% on HumanEval, and adversarial robustness with merely 0.3% accuracy drop (Zhang et al., 2024). The stated interpretation is that simple tasks should induce smaller graphs, while harder tasks should induce richer communication.

TopoDIM replaces iterative multi-round topology refinement with one-shot generation of a heterogeneous directed graph $G=(V,E)$3 whose edges are typed as conditioned, feedback, or debate (Sun et al., 15 Jan 2026). The framework samples a topology via a stochastic policy

$G=(V,E)$4

selects relation types $G=(V,E)$5 over $G=(V,E)$6, and later distills a centralized policy into decentralized local policies $G=(V,E)$7 (Sun et al., 15 Jan 2026). The reward combines task success with Shannon-entropy-based diversity of edge types, and the final graph is sparsified with Top-K selection and breadth-first execution (Sun et al., 15 Jan 2026). Reported results include 46.41% token reduction and 1.50% average performance improvement over state-of-the-art methods, together with a decentralized 1,000-node run requiring 6.36 GB of memory (Sun et al., 15 Jan 2026).

These works shift the meaning of topology adaptation from network routing to organization design. The graph becomes a task-conditioned computation structure. This suggests that in-collective adaptive topologies can be interpreted as a general mechanism for allocating communicative bandwidth, epistemic roles, and decision paths inside a collective, irrespective of whether the substrate is a photonic fabric or a language-agent network.

6. Biological, robotic, and thermodynamic perspectives

Biological and robotic studies further broaden the concept by making topology adaptation part of collective learning or collective function. In cloud-based robotics, cooperative adaptive control treats parameter identification as a collective problem. For multiple manipulators sharing unknown load parameters, a centralized update law

$G=(V,E)$8

leads to tracking convergence, while parameter convergence is guaranteed under collective persistency of excitation: $G=(V,E)$9 The same collective richness result is extended to decentralized, time-varying, and delayed network topologies (Wensing et al., 2017). Here topology matters because it mediates the aggregation of excitation information, but identifiability remains a collective property of the group rather than any one robot (Wensing et al., 2017).

In collective migration, leadership investment and social coupling co-evolve. Each agent has strategy $\theta(G,\mathcal{M}_i)$0, and motion combines direct environmental tracking with graph-mediated social influence: $\theta(G,\mathcal{M}_i)$1 With the simplification $\theta(G,\mathcal{M}_i)$2 and $\theta(G,\mathcal{M}_i)$3, the fast-timescale stochastic dynamics depend on both $\theta(G,\mathcal{M}_i)$4 and the graph Laplacian $\theta(G,\mathcal{M}_i)$5, while social noise is reduced when neighbors are highly invested (Pais et al., 2013). The all-to-all analysis yields cost-dependent bifurcation structure, hysteresis in migration recovery, and branching into leaders and followers when connectivity is sufficiently high (Pais et al., 2013). For the examples shown, approximate branching thresholds are about $\theta(G,\mathcal{M}_i)$6 nearest neighbors for a ring lattice, about 9 neighbors on average for random undirected graphs, and about 8 for random directed graphs (Pais et al., 2013). In small networks, topology also influences where leaders emerge; star graphs and ring lattices produce different leadership locations (Pais et al., 2013).

A more abstract thermodynamic treatment constructs a self-adaptive network that uses only macroscopic observables such as modularity $\theta(G,\mathcal{M}_i)$7, average shortest path length $\theta(G,\mathcal{M}_i)$8, and clustering coefficient $\theta(G,\mathcal{M}_i)$9. Environmental edge additions or removals are accepted with probability

$b$00

where the environment landscape $b$01 is estimated online through a Wang–Landau-like update

$b$02

The reported signature is power-law decay of the relative entropy $b$03, with exponents near $b$04 in several settings, contrasting with the $b$05 behavior of memoryless systems (Bai et al., 2024). Applications in the paper include confinement on phase space, geographic constraints, brain functional networks, and wireless communication networks (Bai et al., 2024).

These examples show that in-collective adaptive topology need not be limited to explicit message passing. It can also denote adaptive redistribution of sensing, influence, excitation, or structural regularity across a collective. A plausible implication is that the concept unifies communication optimization, collective control, and self-organization through a common question: how should a collective rewire itself while it is acting?

7. Trade-offs, misconceptions, and open problems

A recurrent misconception is that any ability to rewire implies benefit. The photonic collective framework rejects that view directly: naive per-step reconfiguration can be counterproductive when $b$06 is large or messages are small (Addanki, 9 Oct 2025). Likewise, oscillator networks show that the relevant mechanism is not arbitrary topology variation or “blinking” but state-dependent co-evolutionary feedback; random switching does not reproduce the percolation enhancement (Eom et al., 2015). In distributed collective synthesis, topology-aware schedules are not equivalent to runtime packet-level adaptivity: TACOS explicitly synthesizes a static collective algorithm offline, albeit one that is adaptive to the topology and to collective progress during synthesis (Won et al., 2023).

A second misconception is that aggregate demand decompositions or generic graph optimization automatically yield valid collective executions. The Birkhoff–von Neumann connection in photonic collectives is explicitly one-way: collective algorithms that proceed via matchings induce a BvN decomposition, but arbitrary BvN decompositions need not satisfy temporal dependencies of real collective schedules (Addanki, 9 Oct 2025). This distinction separates traffic engineering abstractions from collective-algorithm semantics.

The open problems are similarly recurrent across literatures. In adaptive photonic domains, the agenda includes fast heuristics for scheduling, simplified proxies for $b$07, routing under dynamic topologies, variable reconfiguration delays, overlap of reconfiguration with computation, and generalization beyond a single base topology to multiple rings or unions of rings (Addanki, 9 Oct 2025). In embodied collective intelligence, open problems include how to formalize and evaluate topology adaptation, how to pretrain agents that later self-organize into larger collectives, and whether adaptive organization produces genuinely emergent behavior beyond the capabilities of isolated agents (Wang et al., 29 May 2025). In LLM multi-agent systems, sparsity, anchor structure, decentralization, and semantically typed edges emerge as central design variables, but the cost of multi-agent execution remains significant relative to single-model systems (Zhang et al., 2024, Sun et al., 15 Jan 2026).

Across the broader field, the common trade-off is between adaptation benefit and adaptation cost. In communication systems the cost is reconfiguration delay or synthesis overhead (Addanki, 9 Oct 2025, Won et al., 2023). In adaptive networks the cost may appear as approximation breakdown when rewiring is too fast relative to state dynamics (Wieland et al., 2015). In migration and embodied systems the cost is leadership investment, communication overhead, or reduced parallelization efficiency (Pais et al., 2013, Wang et al., 29 May 2025). The shared conclusion is not that topology should always adapt, but that topology should be treated as a control variable whose utility depends on timing, state, workload, and structural constraints.

In that sense, in-collective adaptive topologies designate a general research program: collective systems should be analyzed not only in terms of algorithms or node dynamics on a graph, but also in terms of when the graph itself is a lawful, optimizable, and dynamically coupled component of the collective process.