LLMRouter Says Routing Is Worth 14.6%. Stripe Just Paid $7B to Agree.
Tao Feng, Jiaxuan You and ten others put out LLMRouter on arXiv on August 7, and it's been sitting near the top of HuggingFace's paper board since. The timing is almost comedic — a paper formalizing why model routing matters, published nine days before Stripe bought the biggest router on the market.
The core move is treating routing as a sequential decision process instead of a lookup table. Five parts: context encoders, model encoders, scoring functions, decision rules, learning signals. Once you write it that way, every routing scheme anyone has shipped becomes a specific choice in each of those five slots, and you can actually compare them. The repo carries 16-plus router implementations built against that skeleton.
The headline result is that learned routers beat the strongest single fixed model by 14.6% relative. Not "beat the average model" — beat the best one you could have picked and stuck with. That's the number that justifies the whole category, and it's the first time I've seen it measured cleanly rather than claimed in a launch post.
Two secondary findings matter more for anyone building. Cheap lightweight routers get relatively better as you tighten the cost budget, which is the regime agents actually live in — thousands of calls, most of them not worth a frontier model. And user-conditioned routing improves personalization, meaning the right model depends on who's asking, not just what's being asked.
They also shipped xRouteBench, covering generic LLM, memory-augmented, vision, time-series and personalized routing. Benchmarks are how a research direction stops being vibes, and routing has been vibes for two years. https://arxiv.org/abs/2608.06867
← Back to all articles
The core move is treating routing as a sequential decision process instead of a lookup table. Five parts: context encoders, model encoders, scoring functions, decision rules, learning signals. Once you write it that way, every routing scheme anyone has shipped becomes a specific choice in each of those five slots, and you can actually compare them. The repo carries 16-plus router implementations built against that skeleton.
The headline result is that learned routers beat the strongest single fixed model by 14.6% relative. Not "beat the average model" — beat the best one you could have picked and stuck with. That's the number that justifies the whole category, and it's the first time I've seen it measured cleanly rather than claimed in a launch post.
Two secondary findings matter more for anyone building. Cheap lightweight routers get relatively better as you tighten the cost budget, which is the regime agents actually live in — thousands of calls, most of them not worth a frontier model. And user-conditioned routing improves personalization, meaning the right model depends on who's asking, not just what's being asked.
They also shipped xRouteBench, covering generic LLM, memory-augmented, vision, time-series and personalized routing. Benchmarks are how a research direction stops being vibes, and routing has been vibes for two years. https://arxiv.org/abs/2608.06867
Comments