Reading List — More Recent

Zooter (Lu et al., 2024, NAACL) — distills a reward model's judgments into a lightweight routing policy (“routing to the expert” via reward-guided ensembling).

Taxonomy: decision = user preference (reward model distilled into the router) · workflow = single choice · offline

GraphRouter (Feng, Shen & You, ICLR 2025) — models routing as a graph problem over queries, tasks, and LLMs.

Taxonomy: decision = clustering / similarity, as a graph over queries, tasks and models · adds new models without full retraining

Cascade Routing (Dekoninck et al., 2025) — unifies routing and cascading into one framework, letting the system skip/reorder models dynamically rather than committing to one model or a fixed cascade order.

Taxonomy: workflow = cascade and single choice unified · metrics = best quality under a cost budget

RouterEval (Huang et al., 2025) — a very large-scale benchmark (8,500+ LLMs, 200M+ performance records) used to study a “model-level scaling” effect: better routers keep improving as you add more candidate models to the pool.

Taxonomy: benchmarks · gap to oracle, and how it moves as the candidate pool grows

UniRoute (Jitkrittum et al., 2026) — addresses the harder practical case of routing to LLMs that weren't available at training time, using feature vectors derived from model behavior.

Taxonomy: learning and retraining = models unseen at training time · decision = capability / difficulty features

RouterDC: Query-Based Router by Dual Contrastive Learning for Assembling Large Language Models — moves toward multi-model routing rather than simply strong-vs-weak routing. The key problem they identify is that multiple LLMs may all be good for a particular query. RouterDC uses an encoder plus LLM representations and trains with dual contrastive learning, reporting improvements over existing routing methods on both in-distribution and OOD tasks.

Taxonomy: decision = clustering / similarity (contrastive embeddings) · workflow = single choice from many peers, not strong-vs-weak · offline

IRT-Router: Effective and Interpretable Multi-LLM Routing via Item Response Theory — one of the more interesting recent methodological directions. Instead of treating the router as a generic classifier, they borrow Item Response Theory from psychometrics to explicitly model LLM capability, query characteristics and difficulty, and expected model performance. They evaluate on 20 LLMs and 12 datasets, emphasizing both performance and interpretability.

Taxonomy: decision = difficulty, modeled explicitly as model capability × query difficulty · offline · interpretability

Pandora’s AI Model Router (DeepMind) — Cheap estimators (e.g., embedding-based predictors) are fast but noisy, while accurate estimators (e.g., fine-tuned models with access to retrieval results or partial reasoning traces) are expensive. They formalize this tradeoff as an instance of Pandora’s Box, the classical problem of optimal search with costly inspection. Under a Gaussian signal model, the resulting policies have closed-form value-of-information expressions that determine, for each specialist and input, whether refining the value estimate is worth its cost — the centralized policy is “Pandora’s Router,” extended to a decentralized setting, “Pandora’s Bidder,” where specialists independently decide whether to invest in self-assessment before accepting an offered price to claim a query.

Taxonomy: metrics = value of information against the cost of estimating it · workflow = verifier / escalation with costly inspection · decision = uncertainty

Awesome-Routing-LLMs (GitHub, curated list, not a paper) — an actively maintained paper index for the Routing LLMs field, organized around pre-judgment routing (decide before generation), verification routing (adapt during generation), and memory-based routing (use historical context), plus system-level work on theory, benchmarks, and safety.

Taxonomy: meta · a map of the field rather than a single method