EMNLP survey organizes LLM routing into predictive, adaptive and generative paradigms
kalyan_kpl · x · 2026-11-11
A systematic survey published in Findings of EMNLP 2026, A Systematic Survey of Large Language Model Routing, organizes the fragmented LLM routing literature (which model should handle which query?) along a single axis: the algorithmic operation a router performs.
- Predictive routers score candidate models before answers are generated
- Adaptive routers try a model, observe the outcome, and adjust
- Generative routers let an LLM itself reason about the routing decision
- Infrastructure work studies routing via benchmarks and adversarial analysis rather than performing it
The paper formalizes the routing decision under this four-paradigm lens, traces methodological development, compares assumptions and tradeoffs, and identifies open research directions. The authors maintain a continuously updated reading list on GitHub (wuxingyu-ai/LLM-Routing).
More from coding & agent
- Meme: what your subagents do when the task needs 5 lines of code — neal_lathia · 2026-11-11
- Visa tested 22 models on 15 shopping tasks: ~43% failure, agents bought wrong items anyway — Thick_Mulberry_2235 · 2026-11-11
- MCP or CLI? Yes — Sentry ships a unified experience bridging both — zeeg · 2026-11-11
- Stanford Launches CS329Z: Engineering AI Agents, Covering Memory, Evals and Proactive Agents — stanfordnlp · 2026-11-11
- Solo dev took down his multi-agent site for 8 hours by draining the shared model budget — wolfdogKC · 2026-11-11
- Stanford CS329Z releases assignment on building terminal-bench-style agent evals with Harbor — Diyi_Yang · 2026-11-11