Label Dynamic Routing (LDR) Overview

• Label Dynamic Routing (LDR) is a dynamic routing framework that uses labels or label-like descriptors to determine paths, supervision targets, or expert allocations across various domains.
• It is applied in contexts such as MPLS networking, multi-business recommendation, and mixture-of-experts, leveraging dynamic updates and feedback to refine routing and supervision.
• Extensions of LDR include differentiable routing methods like Sparsegen, risk-aware utility frameworks, and convergence analyses, highlighting both its versatility and challenges in stability.

Searching arXiv for recent and foundational uses of “Label Dynamic Routing” and closely related routing formulations. Label Dynamic Routing (LDR) denotes a family of dynamic routing and supervision mechanisms in which labels, label-like descriptors, or label-conditioned targets determine how traffic, training signals, experts, or models are selected over time. The term is not used uniformly across the literature. In early MPLS work, it appears only implicitly through label-based dynamic path selection with feedback rather than as a formal protocol name (Dumka et al., 2012). In multi-business generative recommendation, LDR is an explicit module that routes supervision to the next future interaction from the same business and thereby “transforms sparse multi-business labels into dense labels” (Li et al., 3 Apr 2026). Related work in mixture-of-experts, model routing, and routing supervision extends the same underlying idea toward differentiable expert allocation, dynamic model selection, and uncertainty-aware label construction (Zhuang et al., 30 Sep 2025, Lai et al., 5 Jun 2026, Jitkrittum et al., 12 Feb 2025).

1. Terminology, scope, and disambiguation

The literature uses LDR in several domain-specific senses rather than as a single canonical protocol. In networking, the central idea is label-based forwarding with dynamically refreshed path state and congestion feedback. In recommendation, LDR is label-side target routing. In expert and model routing, closely related mechanisms make routing decisions depend on dynamic supervision, dynamic sparsity, or per-model behavioral labels.

Context	Mechanism	Distinctive point
MPLS networking	Dynamic label-switched path selection	Best path, backup path, hop-by-hop feedback
Multi-business recommendation	Business-conditioned label routing	Next same-business future target
Mixture of LoRA Experts	Learnable dynamic expert routing	Token-dependent, layer-wise sparsity
LLM routing supervision	Distribution-aware label construction	Rewrites, multiple decodes, risk-aware utility
Dynamic LLM inference	Routing to unseen models	Per-model correctness-vector or cluster representation

A further source of ambiguity is acronym reuse. In imaging, “LDR” commonly means “low dynamic range,” as in LDR-to-HDR conversion for video and image generation, which is unrelated to Label Dynamic Routing (Yu et al., 7 Apr 2026, Bemana et al., 2024). Within routing research itself, the surveyed papers therefore support a narrow reading of LDR as an explicit label-routing module only in some settings, and a broader reading as a label-conditioned or label-side dynamic routing principle in others.

2. MPLS and the emergence of label-based dynamic path selection

The networking precursor is the MPLS framework described in “Dynamic MPLS with Feedback,” where forwarding is based on labels rather than full routing-table lookups, and routing-table lookups are reduced from all routers to label edge routers (LERs) while label switch routers (LSRs) use LDP or RSVP and a label table for forwarding (Dumka et al., 2012). Ingress LERs receive normal IP packets, look up routing information, add an MPLS label, and push the packet into the MPLS cloud; egress LERs remove the label and forward the packet onward as a normal IP packet. Inside the cloud, LSRs swap labels according to label forwarding tables. The control plane uses RIB and LIB, and the data plane uses FIB and LFIB.

The dynamic component is carried by two update packets. Route Update (RU) “shows all paths of the network,” is sent from edge routers at regular intervals, and makes routing information dynamic. Best Path (BP) selects the “best path” and a “backup path” for the destination, and the label table is updated only on the basis of these selected paths. The paper states that update packets “give the details of Label Switch Paths,” so the mechanism disseminates path topology, a preferred LSP, and an alternate LSP. The system therefore behaves as a dynamic label-routing scheme even though the paper does not formally name it LDR.

The most distinctive feature is the hop-by-hop feedback mechanism. Packets are initially forwarded through the best or shortest path. If congestion appears, a feedback indication is returned through an alternate LSP using a 1-bit header or label: 1 denotes congestion and 0 denotes a free path. When the source router receives feedback 1, it stops using the congested path, diverts traffic to the backup path, and then RU and BP updates are sent again so that a new best path and backup path can be selected. The paper contrasts this with end-to-end feedback and describes hop-by-hop signaling as “more reliable and much faster and congestion free path for the packets” (Dumka et al., 2012).

This formulation is conceptually important because it already contains most elements later associated with LDR: labels are the forwarding primitive, path state is refreshed dynamically, the decision is not fixed to a single path, and supervision comes from network-state feedback. The paper compares its mechanism with “Optimal congestion with N+1 label,” “Random Races (RR),” “Shortest Widest Path (SWP),” and “Widest Shortest Path (WSP),” and claims qualitative advantages such as short table lookup, reduced forwarding time, backup-path availability, and better performance with increased number of users (Dumka et al., 2012).

3. Explicit LDR as business-conditioned supervision in generative recommendation

The clearest explicit definition of Label Dynamic Routing appears in MBGR, where LDR is the module that “searches for the label of the next interaction with the same business from the user interaction sequence” and “transforms sparse multi-business labels into dense labels” (Li et al., 3 Apr 2026). Rather than training on the immediate next event in a mixed-business sequence, MBGR routes each training position to the next future item that belongs to the same business. For user $u$, position $t$, and business $b_k$, the routed target is

$$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$

If no such future interaction exists for business $b_k$, the loss for that business is masked.

The motivation is a label sparsity and label mixing problem. Multi-business user sequences interleave events from different businesses, so ordinary next-item prediction yields noisy supervision for business-specific prediction. LDR addresses this at the label level. Each sequence position can supply multiple business-specific targets, each obtained from the nearest future same-business event; invalid cases are masked rather than forced into noisy labels. The paper argues that this makes the supervision denser and more semantically consistent, especially for small or sparse businesses (Li et al., 3 Apr 2026).

LDR operates jointly with the two other MBGR modules. BID provides business-aware semantic IDs and preserves semantic integrity in the representation space. MBP provides business-specific prediction capacity through a shared MoE-style architecture. LDR then determines what each business-specific prediction head should learn from. In the paper’s implied training flow, a user interaction sequence is encoded by BID, MBP produces business-specific predictions, LDR chooses the correct future same-business target, and loss is computed only on valid business targets. The overall objective is

$$\mathcal{L} = \mathcal{L}_{\text{infoNCE}} + \lambda \cdot \mathcal{L}_{\text{recon}}.$$

Here LDR affects the InfoNCE term because it supplies the business-conditioned ground-truth target.

The ablation study isolates LDR’s contribution and reports the largest degradation when it is removed (Li et al., 3 Apr 2026).

Model	All	A
w/o LDR	0.0268	0.0215
Full MBGR	0.0410	0.0252

Model	B	C
w/o LDR	0.0312	0.0245
Full MBGR	0.0554	0.0398

For business D, the reported values are $0.0301$ without LDR and $0.0421$ for the full model. The paper states that removing LDR hurts the most among the three ablations and that the gap is particularly large for smaller businesses B and C. This makes LDR, in this specific line of work, a supervision-routing mechanism rather than a forward-pass routing layer.

4. Differentiable and distribution-aware extensions of dynamic routing

In mixture-of-experts adaptation for LLMs, LD-MoLE replaces fixed TopK routing with a learnable dynamic routing mechanism based on Sparsegen (Zhuang et al., 30 Sep 2025). The routing logits are computed from token embeddings, and a shared MLP predicts a token-dependent sparsity parameter $\lambda$. Sparsegen then maps logits and $\lambda$ into a sparse probability distribution over experts, so the number of active experts is not fixed in advance. The routed output is

$t$0

The closed-form Sparsegen solution is

$t$1

with a threshold $t$2 determined by sorted logits and the support size. The paper highlights four properties of this routing rule: it is fully differentiable, has a well-defined subgradient, has an upper-bounded derivative, and always selects at least one expert for $t$3. It also introduces an analytical sparsity control objective and reports that on Qwen3-1.7B and Llama-3.2-3B, LD-MoLE achieves the highest average scores among tested methods, including gains of more than $t$4 on ARC-E, SWAG, and HellaSWAG relative to MoLA, about $t$5 on OpenBookQA, and more than $t$6 on CommonsenseQA (Zhuang et al., 30 Sep 2025).

A different generalization appears in DARS, which argues that routing labels themselves are noisy when they are derived from a single sampled response per query-model pair (Lai et al., 5 Jun 2026). DARS constructs routing supervision from repeated observations over semantically equivalent rewrites and multiple stochastic decoding runs. It estimates mean quality, mean cost, and quality variability, and then defines a risk-aware utility

$t$7

The preferred model is the one maximizing this utility, and DARS can also produce soft routing targets through a temperature-scaled distribution over utilities. Using 5 rewrites and 5 decoding runs per training query-model pair, the paper reports that single-shot labels are highly unstable, with winner flip rates of $t$8 on GPQA, $t$9 on DROP-800, and $b_k$0 on MATH-500, and that distribution-aware supervision improves rewrite-based utility, decoding-based utility, and routing stability across router families (Lai et al., 5 Jun 2026).

These works suggest a broader interpretation of LDR as dynamic supervision construction rather than only dynamic path selection. In MBGR, the dynamic object is the target label within a mixed sequence. In LD-MoLE, it is the set of active experts per token and layer. In DARS, it is the routing label assigned to a query-model pair after accounting for uncertainty. The common theme is that routing is conditioned on dynamic evidence rather than on a fixed label assignment.

5. Capsule routing, multi-interest extraction, and convergence behavior

Dynamic routing became prominent in capsule networks and recommendation systems before the explicit MBGR formulation. In MIND, the Multi-Interest Network with Dynamic routing represents one user with multiple vectors instead of a single user vector and uses Behavior-to-Interest (B2I) dynamic routing to cluster historical behaviors into latent interest capsules (Li et al., 2019). For behavior embedding $b_k$1, interest capsule $b_k$2, and shared bilinear map $b_k$3, the routing logit is

$b_k$4

Unlike standard capsule routing, MIND uses a shared $b_k$5 and initializes routing logits by $b_k$6 to avoid collapse. It also adapts the number of interests per user as

$b_k$7

After routing, label-aware attention uses the target item embedding to select the most relevant interest vector during training. The training-time representation is

$b_k$8

and serving uses the maximum item–interest dot product. The paper reports that MIND has been deployed for major online traffic at Mobile Tmall and that best online performance occurs around 5–7 interests (Li et al., 2019).

For text classification, capsule networks with dynamic routing provide a more explicitly label-oriented variant because the fully connected capsule layer has one capsule per category plus an orphan capsule (Zhao et al., 2018). The routing coefficient is amended as

$b_k$9

so the existence probability of a child capsule modulates how strongly it can participate in routing. The paper introduces three stabilization strategies—an orphan category, leaky-softmax, and coefficients amendment—and reports that capsule networks achieve state of the art on 4 out of 6 datasets, while 3 routing iterations work better than 1 or 5 (Zhao et al., 2018).

The convergence analysis of capsule routing was formalized later in “The Convergence of Dynamic Routing between Capsules,” which identifies the objective

$$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$0

under simplex constraints on the coupling matrix $$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$1 (Ye et al., 8 Jan 2025). The paper rewrites routing as a nonlinear gradient method and proves the monotone decrease inequality

$$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$2

It also reports a practical caveat: repeated routing often does not materially change classification results but does polarize link strengths, pushing some coefficients toward zero. This observation aligns with the empirical finding in text capsules that more iterations are not necessarily better.

6. Stability guarantees, unseen-model routing, and common misconceptions

A separate line of work studies label-like dynamic routing rules with explicit stability guarantees. In a single-origin-single-destination queuing network, the Generalized Shortest Path policy assigns each path a piecewise-linear score $$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$3 and routes an arrival to the path minimizing $$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$4, where $$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$5 biases routing toward paths whose current bottleneck has a higher service rate (Wu et al., 2024). For the three-path bridge network, the labels are

$$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$6

The stability theorem states that the bridge network is stable if it is stabilizable and

$$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$7

A policy-iteration algorithm then learns $$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$8 from sample paths while remaining within the stability-feasible set. Numerically, the paper reports a $$i_{u,t+1}^{(k)} = i_{u,t'} \quad \text{where} \quad t' = \min\{t'' > t \mid b_{u,t''} = b_k\}.$$9–$b_k$0 reduction in training time relative to a neural-network policy-iteration baseline, with an optimality gap of $b_k$1–$b_k$2 (Wu et al., 2024). This is not called LDR in the paper, but it is explicitly described as a label-like, path-scoring dynamic routing rule.

Dynamic routing among previously unseen LLMs pushes the idea in another direction. Universal model routing represents each candidate LLM by a correctness vector over a small labeled validation set, or by per-cluster average error, and then routes a prompt to the model minimizing estimated error plus a cost penalty (Jitkrittum et al., 12 Feb 2025). With prompt-cluster assignment $b_k$3 and per-cluster error vector $b_k$4, the routing rule is

$b_k$5

The paper evaluates on EmbedLLM, MixInstruct, RouterBench, and Chatbot Arena, and reports effective routing among more than 30 unseen LLMs. This extends LDR-like reasoning from path labels or sequence labels to model-behavior labels.

Several recurrent misconceptions are corrected by this body of work. First, LDR is not a universally standardized protocol name; its meaning depends on domain and paper. Second, dynamic routing is not always a forward-pass selection mechanism: in MBGR and DARS, the main intervention is on target construction and supervision rather than on the predictor architecture itself (Li et al., 3 Apr 2026, Lai et al., 5 Jun 2026). Third, more routing iterations do not necessarily imply better performance, since convergence can coincide with coefficient polarization rather than improved decisions (Ye et al., 8 Jan 2025, Zhao et al., 2018). Fourth, dynamic routing does not imply differentiability: MPLS-style LDR relies on periodic updates and hop-by-hop 1-bit congestion feedback, whereas LD-MoLE derives a differentiable closed-form router from Sparsegen (Dumka et al., 2012, Zhuang et al., 30 Sep 2025).

Taken together, these results support a concise technical characterization: Label Dynamic Routing is best understood not as one algorithm but as a design pattern in which labels, routed labels, or label-derived state are updated dynamically to choose paths, targets, experts, or models. Its principal functions are densification of supervision, business or context specificity, congestion or risk sensitivity, and cost-aware selection. Its principal difficulties are instability under noisy labels, over-polarization under repeated routing, and the absence of a single cross-domain formalism.