Routing Collapse: Mechanisms & Mitigations

• Routing Collapse is a phenomenon where dynamic routing systems fail, leading to inefficient data transfer and potential systemic overload.
• Mathematical models, using metrics like entropy and coefficient of variation, clearly detect collapse and guide load-balancing and mitigation strategies.
• Algorithmic interventions such as EquiRouter and low-dimensional gating regularization are key to restoring routing stability and preventing expert or representational collapse.

Routing collapse refers to degenerate or catastrophic behaviors in dynamic, distributed, or learned routing systems, where the routing mechanism ceases to provide correct, efficient, or intended operation. Manifestations range from network-level throughput collapse under adversarial flooding, persistent oscillations or withdrawal in interdomain routing protocols, expert under-utilization in learned neural routers, to catastrophic representation collapse in mixture-of-experts networks. Although the mechanisms and models differ by context—classical networks, multipath optimization, neural architectures, or decision-aware model selection—the unifying thread is the breakdown of routing diversity or stability, often through a combination of overload, misdesign, or unbalanced learning.

1. Forms and Mechanisms of Routing Collapse

Routing collapse can be categorized by its setting and cascade process:

• Classical Networks: In MANETs and AODV/DSR-like protocols, malicious nodes can flood the network with control packets (e.g., rogue RREQs). This consumes neighbor forwarding quotas, results in route request losses or detours, and, as in ACRR, leads to denied legitimate connectivity and sharply degraded throughput (Kataria et al., 2010).
• Dynamical Network Flows: In the ODE-based framework for flow networks with finite-capacity links, collapse is triggered by link capacity reductions that eventually force spill-backs. Once a link's occupancy hits its cap, outflow collapses to zero, causing upstream congestion and a domino effect, resulting in either full throughput or complete collapse (outflow drops to zero) (Como et al., 2012).
• Routing in BGP: Network-destabilizing attacks such as fixed-route (persistent bogus path) announcements can create lasting oscillations (“bad gadgets”), or persistent withdrawals. This “collapse” is avoided only under strict commercial policy constraints (Gao-Rexford: customer>peer>provider, valley-free export) (Lychev et al., 2012).
• Mixture-of-Experts (MoE) and Sparse MoE Models: Expert collapse arises in neural routers when the gating mechanism concentrates input traffic on a few experts, starving others and defeating the premise of conditional computation. The same effect, termed representation collapse, appears when hidden representations cluster in the low-dimensional span of expert centroids under SMoE training (Rokah et al., 21 Jan 2026, Chi et al., 2022).
• LLM Router Collapse: In dynamic LLM selection, as budget constraints loosen, degenerate routers route all traffic to the largest model, regardless of sufficiency, wasting computation and cost—termed routing collapse in this context (Lai et al., 3 Feb 2026).
• Capsule Networks: Variance collapse is a pathology of classical EM capsule routing, where an expert’s or capsule’s covariance shrinks to zero, overfitting and causing numerical instability—a consequence of unconstrained MLE during capsule assignment (Ribeiro et al., 2019).

2. Mathematical Formulations and Detection

The signature of routing collapse in various domains is formally characterized by observable metrics and equations.

LLM Routers:

Let $\mathcal{M} = \{1, ..., K\}$ be a model pool; for query $q$, let $a_j(q)$ be model $j$'s true accuracy, $c_j(q)$ its cost, and $C$ a budget. The call rate is:

$$\text{call-rate}_j(C) = \text{fraction of queries assigned to model }j.$$

Collapse is observed if, over a wide range of $C$,

$\text{call-rate}_{j^*}(C) \to 100\%,$

where $j^*$ is the most powerful/expensive model, even when $a_{j^*}(q) \approx a_{j}(q)$ for many $j$ (Lai et al., 3 Feb 2026).

MoE/SMoE:

For $N$ experts, per-expert load (SoftMoE) is $\bar{p}_i = (1/B)\sum_b p_i(x^{(b)})$; routing collapse is indicated when $\max_i \bar{p}_i \to 1$, $\bar{p}_{j\neq i} \to 0$.

Metrics:

• Entropy: $$H_{\text{avg}} = -\sum_{i=1}^N \bar{p}_i \log \bar{p}_i$$; collapse $\Rightarrow H_{\text{avg}} \to 0$.
• Coefficient of Variation: $CoV = N\sqrt{(1/N)\sum_i(\bar{p}_i - 1/N)^2}$; collapse $\Rightarrow CoV \gg 1$ (Rokah et al., 21 Jan 2026).

Dynamical Flow Cascades:

Routing collapse is formalized via the dichotomy:

$$\lim_{T\to\infty} \frac{1}{T}\int_0^T \lambda_n(s) ds = \lambda_0 \text{ or } 0$$

where $\lambda_n$ is destination outflow; zero limit denotes total collapse (Como et al., 2012).

BGP Oscillation:

Collapse is defined by nonconvergent routing tables or persistent oscillations, linked combinatorially to “dispute wheels” or induced by attacker's fixed-path announcements (Lychev et al., 2012).

3. Root Causes and Theoretical Explanations

Collapse typically traces to a failure of regulatory or balancing constraints under dynamics or optimization.

• Objective–Decision Mismatch: LLM routers trained by MSE for performance prediction, but evaluated by $\arg\max$, permit small errors to flip top choice in tight margins, over-selecting large models (Lai et al., 3 Feb 2026).
• Gradient Coupling: In MoE, gating gradients drag hidden states toward a subspace spanned by a few experts; if unchecked, this induces representation collapse (Chi et al., 2022).
• Absence of Load-Balancing: Without KL-to-uniform or load/importance losses, gating networks minimize global loss by specializing on few experts, starving the rest (Rokah et al., 21 Jan 2026).
• Overload in Distributed Systems: In AODV/DSR, honest node quotas on RREQs are exploited by attackers. Exhaustion of local limits recursively disables genuine route formation, effectively collapsing routing functionality (Kataria et al., 2010).
• Singularities in EM: Capsule routing by MLE drives Gaussian cluster covariance to zero when “captured” by a single datapoint; no inherent mechanism prevents this singularity except regularizing priors (Ribeiro et al., 2019).

4. Algorithmic and Mathematical Mitigation Strategies

Multiple algorithmic variants and regularization strategies have been developed to counter routing collapse.

• LLM Routers: EquiRouter replaces scalar prediction with pairwise ranking loss, provably aligning training and discrete $\arg\max$ inference, thus restoring utilization of smaller models and reducing cost (by 17% at GPT-4-level performance in RouterBench) (Lai et al., 3 Feb 2026).
• SMoE/X-MoE: Routing from a low-dimensional, L2-normalized sphere decouples gating gradients from full hidden space, suppressing collapse and yielding more robust, reproducible expert assignments. X-MoE achieves improved cross-lingual and downstream performance and stable expert usage across seeds (Chi et al., 2022).
• MoE Regularization: Batch-level KL to uniform or importance/load penalties ensure equitable expert utilization. Early injected Gaussian gating noise and top-k>1 routing further smooth gradients and mitigate collapse (Rokah et al., 21 Jan 2026).
• ACRR for MANETs: Shifting RREQ_RATELIMIT to per-neighbor, dynamic quotas plus exponentially back-off blacklisting isolates malicious nodes, bounding per-hop damage and preserving overall route formation under attack (Kataria et al., 2010).

Table: Collapse Setting and Remedy

Domain	Collapse Mechanism	Mitigation Approach
MANETs (AODV/DSR)	Malicious RREQ overload	Per-neighbor quotas, blacklisting
BGP	Fixed-route attack/oscill.	Valley-free policy, Gao-Rexford model
LLM Routers	Degenerate model overuse	Pairwise ranking (EquiRouter)
MoE/SMoE	Load unbalance, rep. coll.	Load KL, low-dim routing (X-MoE)
Capsule Networks	Variance-coll. in EM	VB-routed entropy regularization

5. Empirical Characterization and Metrics

Collapse is detected and measured through direct metrics and observable outputs:

• LLM Routing Collapse Index (RCI): Frequency of dominated assignments (strongest router: RCI ≈ 0.78; EquiRouter: RCI ≈ 0.69).
• Expert Loads: Actual per-expert utilization, entropy, and variability across input batches and seeds (Rokah et al., 21 Jan 2026, Chi et al., 2022).
• Route Formation and TCP/AODV Ratio: Number of successful source→destination paths; sharp drop of TCP/AODV packet ratio toward zero signals collapse in AODV (Kataria et al., 2010).
• Hessian and Loss Sharpness: Increased loss curvature, especially along top eigenvectors, is consistent with sharp, unstable routing transitions characteristic of collapse (Rokah et al., 21 Jan 2026).
• Representation Collapse Metric: Trace-ratio RC (within/between cluster covariance); decline in RC reflects representational collapse to expert subspace (Chi et al., 2022).

6. Routing Collapse in Graph Augmentation and Failure Recovery

In next-hop routing, collapse is modeled via the Tree Augmentation problem: maximizing the number of routers with alternate next-hops, subject to acyclic constraints. This problem is NP-hard, but admits a simple $\frac12$-approximation and, in special cases, polynomial/dynamic-programming solutions (two-arm trees, bounded treewidth, BFS-planar graphs). Failure to adequately augment the routing tree leads to collapse: a single failure can disconnect swathes of the network or trigger loops. Proper augmentation directly improves resilience (Borradaile et al., 2014).

7. Broader Implications and Limitations

Routing collapse represents a generic path to systemic failure across technical domains involving distributed or learned decision-making under constraints. While algorithmic interventions (decision-aware training, explicit regularization, policy design) can remedy many such pathologies, several caveats persist:

• Parameter Sensitivity: Proper setting of regularization weights and quota parameters is nontrivial; under- or over-regularization leads to, respectively, residual collapse or underutilization (Kataria et al., 2010, Rokah et al., 21 Jan 2026).
• Attack Adaptivity: Sophisticated adversaries or distributed attacks (e.g., Sybil in ACRR) may circumvent node-local controls without cryptographic reinforcement (Kataria et al., 2010).
• Nonlocal Sharpness: Even with balanced expert utilization, MoE models can experience abrupt nonlocal loss jumps under small perturbations due to routing threshold effects (Rokah et al., 21 Jan 2026).
• Incomplete Theoretical Guarantees: While specific mitigations prevent collapse under tested regimes, formal global guarantees may be lacking, especially in large, heterogeneous, or adversarially-evolving networks (Como et al., 2012, Lai et al., 3 Feb 2026).

Routing collapse thus serves as a critical motif—and caution—for the design and analysis of both classical and modern routing systems, emphasizing the necessity for carefully matched objectives, robust regularization, and adaptive resource control.