EquiRouter: Equitable Routing Framework

• EquiRouter is a framework that integrates explicit fairness metrics (e.g., Gini coefficients, ranking losses) into routing decisions across machine learning and transportation domains.
• It reframes model selection as a supervised ranking task in multi-LLM routing, reducing routing collapse and cutting operational costs by up to 17% while maintaining high quality.
• In traffic and vehicle routing, EquiRouter employs bilevel optimization and fairness indices to improve travel equity and system performance through scalable, real-time algorithms.

EquiRouter refers to a set of theoretically grounded and computational frameworks designed to achieve equitable decision-making in routing scenarios across disparate problem domains. These include machine learning model selection (LLM/MLM routing), multi-agent or vehicle routing for transportation and logistics, and multi-modal user equilibrium in traffic assignment. While methodologies diverge sharply according to application, the hallmark of each EquiRouter instantiation is the explicit modeling and optimization of fairness or equity—typically via ranking-based learning, Gini-style dispersion metrics, or explicit fairness constraints—rather than optimizing only raw efficiency or accuracy.

1. Motivation and Core Principles

The core motivation for EquiRouter frameworks is to address intrinsic trade-offs between efficiency/quality and equity/fairness. In multi-agent or multi-model routing problems, naïve optimization focusing solely on global cost minimization, best-possible predictive performance, or shortest paths often leads to degenerate allocations. These degenerate allocations can manifest as "collapse" onto a single high-performing model (in multi-model AI inference) (Lai et al., 3 Feb 2026), imbalance of workloads among agents or vehicles (Bhadoriya et al., 2024), or routings that exacerbate existing mobility inequities in transportation systems (Bang et al., 2024, Bang et al., 2023, Bai et al., 16 Mar 2025).

Key organizational principles include:

• Explicit measurement and modeling of equity, e.g., via Gini coefficients, ranking losses, or Pareto fronts.
• Algorithmic frameworks that adaptively balance efficiency and fairness, often parameterized to allow for policy trade-off selection.
• Architecture and pipeline design to maintain tractability and guarantee convergence in large-scale or real-time domains.

2. EquiRouter for Model Selection and Routing in ML Inference

The term EquiRouter was introduced in the context of LLM router design to formally address the problem of "routing collapse." In multi-LLM architecture, a router aims to allocate queries among a pool $\mathcal{M}=\{1,2,\ldots,K\}$ of models with varying cost-quality profiles, such that a user-defined budget constraint $C$ is satisfied while maximizing global query performance (Lai et al., 3 Feb 2026).

Routing Collapse and Objective–Decision Mismatch: In practice, as $C$ increases, classical routers (e.g., MLPRouter, GraphRouter) consistently default to the most expensive and accurate model (e.g., GPT-4), leading to underutilization of smaller, more efficient models—a phenomenon termed "routing collapse." This problem is traced to the mismatch between the training objective (predicting scalar scores) and the deployment decision (ranking and selection). Near-ties in model scores make the routing decision sensitive to prediction noise. Empirical evidence on RouterBench shows that Oracle routers only require the strongest model for less than 20% of queries, but standard routers saturate its usage to near 100%.

EquiRouter Solution: EquiRouter reframes multi-model selection as a supervised ranking task. Rather than training to regress per-model accuracy estimates, the router trains on pairwise comparisons:

• Query and model embeddings are combined via a FiLM-modulated architecture.
• The scoring head produces per-model scores, and a pairwise logistic loss supervises direct ranking that aligns with the desired routing decisions.
• Inference restricts feasible models to those meeting the budget and selects the highest-scoring.

Quantitative Impact: On the RouterBench benchmark, EquiRouter achieves normalized AUC (nAUC) = 0.7712 (best), QNC = 0.7731 (i.e., 77% of GPT-4's cost for same quality, versus ≥93% in baselines), and RCI (routing collapse index) = 0.6911 (lowest). At GPT-4's quality, EquiRouter saves approximately 17% of the computational/monetary cost (Lai et al., 3 Feb 2026). Ablative experiments confirm that the combination of joint model-query features and pairwise ranking supervision is essential to avoid collapse.

Implications: EquiRouter restores the role of smaller models and supports robust, cost-efficient multi-model orchestration in production LLM systems, with potential extension to dynamic or ensemble-based routing.

3. EquiRouter in Multimodal and Traffic Network Equity

In multi-agent transportation and mobility contexts, EquiRouter frameworks formalize and optimize equity among traveler populations or vehicle agents. Several scholarly contributions operationalize this principle via developed fairness metrics and bilevel/bi-objective optimization (Bang et al., 2024, Bang et al., 2023, Bai et al., 16 Mar 2025).

Equity Metrics:

• Mobility Equity Metric (MEM): Aggregates mode-specific accessibility and penalizes by cost sensitivity; implemented as a population-weighted, Gini-based complement. MEM = 1 indicates perfect equity (Bang et al., 2024, Bang et al., 2023).
• Dynamic Trip Index (DTX) and Dynamic Trip Equity (DTE): In real-time dynamic systems, the DTX captures relative trip quality per traveler, and DTE is a Gini-style equity score among simultaneously competing road users (Bai et al., 16 Mar 2025).
• Tour-length Gini/Jain indices: In multi-vehicle routing, tour-length distributions are analyzed via Gini coefficient and Jain’s fairness index (Bhadoriya et al., 2024).

Algorithmic Structure:

• Bilevel/Fixed-point Routing: System-centric routing for compliant vehicles solves a convex program minimizing weighted travel-time costs, while non-compliant drivers are modeled via cognitive hierarchy and shortest-path games. An upper-level (outer loop) adjusts mode weights or routing parameters to maximize MEM (or minimize Gini) subject to efficiency constraints (e.g., limiting delay for compliant drivers) (Bang et al., 2024, Bang et al., 2023).
• Real-time Optimization: In dynamic routing, EquiRouter computes, at each decision point, the optimal route for each vehicle to maximize its local equity score, considering both monitored and anticipated congestion and maintaining consistency among competing vehicles’ plans (Bai et al., 16 Mar 2025).

4. Mathematical Formulations and Algorithmic Techniques

Multi-Model LLM Routing:

• Feasible model set $F(q;C) = \{j \mid c_j(q)\le C\}$.
• Routing decision $j^*(q;C) = \arg\max_{j\in F(q;C)} a_j(q)$.
• Pairwise ranking loss used during training: $$L_{\mathrm{rank}}(q) = (1/|P(q)|)\sum_{(i,j)\in P(q)} \log(1+\exp(-(s_i-s_j)))$$ (Lai et al., 3 Feb 2026).

Multi-modal Traffic/Equity-Aware Transportation:

• Mobility equity ($\mathrm{MEM}$) defined as MEM = $$1 - \frac{\sum_{i,j}p_i p_j|\epsilon_i-\epsilon_j|}{2PM}$$.
• System planner solves:

$$\min_{x} J_{\mathrm{sys}}(x) = \sum_{m\in\mathcal{M}} w_m \sum_n \sum_{(i,j)} t^{ij}(x^{ij}+q^{ij})x_{m,n}^{ij}.$$

• Equity maximization solved via line/grid search:

$$\max_{w} {\rm MEM}(w) \quad \text{s.t.} \quad \delta^{pv}(w)\le \Gamma.$$

Multi-Vehicle Routing (Fair-MTSP):

• $\varepsilon$-F-MTSP (MISOCP): $$(1-\varepsilon+\varepsilon\sqrt{m})\,||\mathbf{l}||_2 \le ||\mathbf{l}||_1$$.
• $\Delta$-F-MTSP (MILP): $\mathrm{Gini}(\mathbf{l}) \leq \Delta$.
• Pareto-front generation by varying $\varepsilon$ or $\Delta$ (Bhadoriya et al., 2024).

Solution Methods:

• Branch-and-cut with dynamic subtour elimination, outer-approximation of SOC/p-norm cones, and BPR-type latency models in traffic.
• Interior-point solvers for convex programs; dynamic programming for real-time DTE maximization.
• Convergence proven or empirically demonstrated in synthetic and real-world datasets.

5. Quantitative Results and Comparative Analysis

Multi-Model Routing:

• On RouterBench, EquiRouter reduces cost by ~17% at GPT-4-level quality compared to the strongest prior router and dramatically lowers strongest-model call rates (Lai et al., 3 Feb 2026).

Traffic/Mobility Equity:

• MEM improvements of 10–20% over baseline feasible by redistributing compliant vehicle flows and increasing public transit share, even amid non-compliance (Bang et al., 2024).
• Real-time DTE: In a 1,000-vehicle Boston urban grid, EquiRouter achieves a relative DTE improvement of 13.3%, or ~11.4% when aggregated overall, as well as pronounced trip time reductions for riders of lower-cost modes (Bai et al., 16 Mar 2025).

Multi-Vehicle Routing:

• Fair-MTSP models (EquiRouter) balance workload nearly as well as min-max MTSP but at 5–10× less solve time; Gini can be tuned via parameters to approach exact equality at the expense of modest total cost increase (Bhadoriya et al., 2024).

Key trade-offs are clearly elucidated via generated Pareto fronts, enabling policy selection according to the desired cost-equity balance.

6. Limitations, Open Problems, and Future Directions

Identified limitations and open avenues include:

• Pairwise ranking supervision in LLM routing scales as $\mathcal{O}(K^2)$; listwise or sampled losses are not yet implemented for large $K$ (Lai et al., 3 Feb 2026).
• Incorporation of side-information into model embeddings (architecture, data provenance) may further improve discrimination in multi-LLM settings.
• Out-of-domain robustness under extreme distribution shift is not fully characterized (Lai et al., 3 Feb 2026).
• Dynamic and context-dependent ensemble routing remains largely open in both model selection and transportation.
• In transportation, extensions to time-dependent flows, rolling-horizon optimization, and mechanisms that guarantee or incentivize compliance are active research directions (Bang et al., 2024, Bang et al., 2023).
• For vehicle routing, scalable heuristics and real-world deployment with complex constraints (capacity, time windows, heterogeneous vehicles) are being actively pursued (Bhadoriya et al., 2024).

7. Broader Implications and Impact

EquiRouter frameworks realign learning or optimization objectives with deployment-time equity constraints across settings where efficiency-centric policies may result in undesirable societal or economic effects. In production LLM services, principled routing yields better quality-cost trade-offs and reduces resource waste. In transportation and logistics, Gini-bounded or MEM-maximizing strategies enable materially more fair access, lower impedance for systemically disadvantaged populations, and transparent management of equity-efficiency trade-offs.

As a result, EquiRouter is increasingly recognized as a flexible, theoretically grounded paradigm for architecting equitable algorithms in multiagent, multi-model, and networked decision systems (Lai et al., 3 Feb 2026, Bang et al., 2024, Bang et al., 2023, Bhadoriya et al., 2024, Bai et al., 16 Mar 2025).